Experiments

Data Science

Testing AI features across different user segments

May 28, 2026

min read

A graphic of a bar chart with an arrow pointing upward.

An AI feature that looks successful in your aggregate metrics can be quietly failing a specific group of users at the same time.

This happens because AI models produce outputs based on statistical patterns learned from training data — and those patterns don't match every user equally. A summarization model trained mostly on English text will produce worse results for Spanish speakers. A coding assistant tuned for experts will confuse beginners. Your overall numbers can look fine while a real cohort has a genuinely bad experience.

This article is for engineers, PMs, and data teams who are building or shipping AI features and want to test them more honestly. If you've ever wondered why your AI feature seems to work well overall but gets complaints from specific users, AI feature segmentation is the practice that helps you find and fix that gap. Here's what you'll learn:

Why AI performance varies by user type, language, plan tier, and intent — and why aggregate metrics hide those differences
How to choose the right user segments to test before your experiment launches
How to structure A/B tests that actually surface segment-level performance gaps
How to pick the right success metrics for each segment, including guardrail metrics that catch harm early
How to build a repeatable workflow using feature flags and targeted rollouts so every AI release is segment-aware by default

The article moves in order from the "why" to the "how." It starts with the mechanics of why AI outputs differ across user groups, then walks through segment selection, experiment design, metric choice, and finally the operational workflow that makes this sustainable over time.

Why AI features don't perform the same way for every user

If you've shipped an AI feature and declared it successful based on aggregate metrics, there's a reasonable chance you've missed something important. Not because your measurement was sloppy, but because the nature of AI systems makes aggregate measurement structurally insufficient for detecting certain classes of failure.

Understanding why requires a short detour into how these models actually work — and why that's fundamentally different from how traditional software fails.

AI outputs are probabilistic, not deterministic

A conventional software function is deterministic: given the same input, it returns the same output. If it breaks, it breaks consistently and visibly. AI models don't work this way. They generate outputs shaped by statistical patterns learned from training data, and the quality of those outputs is a direct function of how closely a given user's input resembles the distribution the model was trained on.

The same prompt, submitted by two different users with different linguistic backgrounds or domain expertise, can produce outputs of meaningfully different quality — not because of a bug, but because of how the model learned.

This is the core operating characteristic of probabilistic systems, and it has a direct consequence for product teams: you cannot test an AI feature once, observe a positive result, and ship with confidence. The result you observed reflects the aggregate of your test population, which may or may not represent the full range of users who will encounter the feature in production.

The dimensions along which AI performance diverges

The axes of variance are predictable once you understand the training distribution mechanism. Language and locale are among the most significant: a summarization model trained predominantly on English-language text will produce degraded output for Spanish-speaking users, not because the model is broken, but because Spanish-language text was underrepresented in its training data.

MIT researchers documented this pattern directly in a medical context, finding that a model trained mostly on data from male patients made incorrect predictions for female patients when deployed in a hospital — a subgroup failure that was invisible in aggregate accuracy metrics.

User expertise level introduces a different kind of divergence. A recommendation engine that surfaces advanced configuration options may delight power users who know what they're looking for, while overwhelming new users who lack the context to evaluate those recommendations. The model is behaving consistently; the user populations are interpreting and benefiting from its outputs differently.

Plan tier and behavioral intent introduce similar dynamics: enterprise users tend to bring more complex, domain-specific tasks, while free-tier users may be exploring capabilities with less defined goals. A single model configuration rarely serves all of these populations equally well.

Why aggregate metrics hide segment-level failures

This is where the problem becomes operationally dangerous. A global A/B test computes an average treatment effect across all users. If an AI feature meaningfully improves outcomes for 70% of your user base while degrading them for 30%, the aggregate result can still register as neutral or positive lift — and you'll ship a feature that's actively harming a cohort you never examined.

This is qualitatively different from traditional software testing. A broken UI component fails for everyone. An AI feature that underperforms for a specific segment fails silently, masked by the majority's positive response.

The technical term for this is heterogeneous treatment effects — the same feature variant produces meaningfully different outcomes across user populations — and it's the expected default behavior of any model deployed against a non-uniform user base, not an edge case.

The risk is highest for high-value or structurally underserved cohorts: non-English speakers, new users still forming habits, enterprise accounts with specialized workflows. These are often the users whose failures are most costly and least visible in aggregate dashboards.

What "working" actually means depends on who you ask

Consider a text summarization feature built on a model trained predominantly on English-language business documents. For English-speaking enterprise users, it performs well — outputs are coherent, accurate, and useful. For Spanish-speaking users, quality degrades because the model's training distribution underrepresents Spanish-language text.

Aggregate metrics show the feature is "working" because the English-speaking majority dominates the average. The Spanish-speaking cohort's degraded experience is statistically diluted into invisibility.

The same mechanism plays out with expertise divergence. A coding assistant that generates advanced, idiomatic solutions may accelerate experienced engineers while producing outputs that junior developers can't evaluate or safely use. Both cohorts are using the same feature. The aggregate engagement metric looks fine. The junior developer cohort is quietly accumulating technical debt or abandoning the feature entirely.

As Landon Smith, Head of Post-Training at Character.AI, put it: the only way to determine which model behavior actually serves users is to compare modeling techniques "from the perspective of our users" — not from the perspective of offline evals or aggregate product metrics.

That framing captures the problem precisely. AI performance is not a property of a model in isolation. It's a property of a model in contact with a specific user population, and that population is never uniform.

Segment selection happens before the experiment, not after the results disappoint you

Most teams reach for the easiest segments first — plan tier, geography, device type — because those attributes are already in the database and require no additional instrumentation. The problem is that "easy to query" and "likely to correlate with model performance" are not the same thing.

Choosing segmentation dimensions that don't actually map to how your AI feature behaves produces one of two bad outcomes: underpowered tests that return no signal, or neutral aggregate results that give you false confidence while a specific user cohort quietly has a terrible experience. The segment selection decision has to happen before the experiment launches, not as a post-hoc analysis after you're already puzzling over flat results.

Which dimensions actually correlate with AI model performance

The right way to evaluate a potential segmentation dimension is to ask whether it plausibly changes either the input distribution the model receives or the evaluation criteria the user applies to the output. If neither is true, the dimension probably won't reveal anything meaningful about AI performance differences.

Five dimensions tend to pass this test for most AI features: user expertise level, language and locale, plan tier, device type, and use-case intent. Expertise level changes input distribution directly — a power user submitting a detailed, well-structured prompt to an AI writing assistant will receive systematically different output quality than a novice submitting a vague one-sentence request.

Language and locale change both input distribution and the model's ability to process it, particularly for models trained predominantly on English-language data. Plan tier often correlates with use-case maturity and data volume, especially in B2B contexts. Device type affects how outputs are rendered and consumed, which matters for multimodal or long-form AI features.

Use-case intent — what the user is actually trying to accomplish — is frequently the highest-signal dimension of all, and the one most often overlooked.

Research from marketing analytics contexts suggests that behavior-based segments are substantially more predictive of outcomes than demographic segments. While that finding comes from a different domain, the underlying logic applies directly to AI feature testing: what users do with a feature, and why they're using it, tells you more about how a model will serve them than where they live or what they pay.

User expertise and use-case intent as high-signal dimensions

Expertise level and behavioral intent deserve particular attention because they're the dimensions most likely to be skipped in favor of attributes that are easier to pull from a user table. A concrete illustration: a SaaS product team that clustered users by feature usage patterns discovered three distinct activation paths.

Users who engaged with reporting features first retained at roughly twice the rate of users who started with setup flows. The segment that mattered wasn't demographic — it was behavioral intent at the moment of onboarding.

For AI features, this pattern is even more pronounced. Two users on identical plan tiers, in the same country, using the same device, can have radically different experiences with the same AI feature if one is an experienced practitioner who knows how to structure inputs and evaluate outputs, and the other is encountering the feature for the first time.

Aggregate metrics blend these experiences into a result that looks acceptable while neither cohort is being served well.

Language, locale, and plan tier — operationalizing them in practice

Language and locale are worth separating from generic geography because the performance gap is usually at the model level, not the UI level. A summarization model that performs well on English-language content may degrade significantly for Spanish or Mandarin inputs, and that degradation won't appear in aggregate satisfaction scores if English speakers are the majority of your test population.

Plan tier matters most in B2B contexts where the feature's value proposition differs meaningfully across customer segments. An AI feature that synthesizes large volumes of historical data may be genuinely useful for enterprise accounts with years of accumulated records and actively harmful — or simply irrelevant — for a startup account with three months of data.

Experiment targeting rules can handle these multi-dimensional definitions through AND/OR attribute logic, which lets you construct conditions like locale = es-MX AND plan_tier = free AND feature_usage_count < 5 without custom engineering work. Reusable saved group definitions mean that when the same segment — say, non-English enterprise users — needs to be tested across multiple AI feature releases over time, you're not rebuilding the audience logic from scratch each time.

For B2B products specifically, organization-level targeting is relevant when AI performance differences exist at the account level rather than the individual user level. Teams can also pass custom user attributes — including model-specific signals like prompt complexity scores or session depth — as targeting dimensions, which opens up AI feature segmentation approaches that go well beyond standard demographic fields.

For data teams who need to derive segment membership from historical usage patterns rather than live user properties, warehouse-native experiment platforms typically support two segment types: SQL-defined segments built from arbitrary warehouse queries, and fact-table-defined segments built from structured event data. This matters when expertise level or behavioral intent has to be inferred from past behavior rather than read from a user profile field.

The cost of getting segmentation wrong

When teams choose dimensions that don't correlate with model performance, the experiment doesn't just fail to find a signal — it actively misleads. A test that shows a neutral aggregate result on an AI feature may be hiding a significant negative effect on a specific cohort that simply isn't large enough to move the overall metric. That's not a statistical problem you can solve after the fact; it's a design problem that compounds with every experiment cycle you run on the wrong dimensions.

Two practitioner failure modes are worth naming explicitly. The first is running segmentation without a clear hypothesis about why that dimension should matter for model performance — treating segmentation as a mechanical step rather than a reasoning exercise.

The second is accepting segments that are statistically valid but not actionable: a segment that exists in your data but can't be targeted in your feature flag system, or that can't be connected to a metric your team actually controls. Both failures waste experiment cycles and erode confidence in the testing process itself. Defining the right segments before you test is how you avoid spending those cycles on questions that were never going to produce useful answers.

A single global A/B test is structurally incomplete for AI features

Running a single global A/B test on an AI feature is a structurally incomplete experiment design. It will frequently return a neutral or mildly positive aggregate result while concealing significant degradation for specific user subgroups. For traditional software, aggregate metrics are a reasonable proxy for quality — a bug either fires or it doesn't, and its effect tends to be uniform.

AI features don't work that way. Their outputs are probabilistic and context-dependent, which means a model that performs well for your majority population can simultaneously be producing poor outputs for a minority segment. The majority's positive signal averages out the harm, and the experiment reads as a ship decision when it should be a rollout decision at best.

This isn't a statistical edge case. It's the default failure mode for AI feature experiments that aren't designed to look for it.

The aggregate-masking problem

The mechanism here matters. Traditional software bugs tend to produce discrete, observable failures — a broken function throws an error, a misconfigured UI renders incorrectly for everyone. AI quality degradation works differently: it's continuous and subjective.

A recommendation engine that performs well for users with dense interaction histories may produce near-random suggestions for users with sparse ones, and that degradation won't surface as an error rate or a latency spike. It shows up as lower engagement, higher abandonment, and reduced return sessions within that cohort — signals that are easy to miss when you're looking at aggregate dashboards.

Aggregate metrics have no way to distinguish "feature works well for everyone" from "feature works well for most people and poorly for a few."

Pre-specifying segment hypotheses before the experiment runs

The corrective is not post-hoc slicing of results. Slicing after the fact inflates false positive rates through multiple comparisons, and it tends to surface spurious patterns rather than real differential effects. The discipline is pre-specification: before the experiment launches, document which segments you hypothesize will show differential treatment effects and articulate why.

That "why" is the important part. If you're testing a new language model on a summarization feature, you should be able to state mechanistically why Spanish-locale users might respond differently — because the model's training data skews English, because your evaluation set didn't include Spanish-language content, because the prompt template wasn't localized. That reasoning forces better experiment design and gives you a falsifiable hypothesis rather than a fishing expedition.

Practically, this means defining your segment filters before analysis begins, configuring dimension-level breakdowns to capture treatment effects across those dimensions within a single experiment, and treating guardrail metrics for high-risk segments as first-class experiment outputs — not afterthoughts.

In GrowthBook, guardrail metrics are a distinct schema object from primary goals and secondary metrics, which means they can be configured to surface harm signals even when aggregate primary metrics look positive.

Feature flag gating by user attribute

Feature flag gating by user attribute is the mechanism that makes segment-aware AI experiments operationally tractable. Rather than exposing all users to a new AI variant and hoping your analysis catches segment-level problems after the fact, you can gate the variant to specific segments using targeting rules on user attributes — locale, plan tier, expertise level, organization ID for B2B contexts, or any custom attribute your application tracks. GrowthBook supports attribute-based targeting rules with AND/OR logic, Saved Groups for reusable audience segments, and organization-level targeting for B2B use cases.

GrowthBook's experiment override rules support AND/OR attribute logic, saved group targeting, and B2B organization-level targeting, giving teams enough precision to isolate exposure to the exact cohort they're testing. The assignment algorithm uses deterministic hashing — the experiment seed and the configured hash attribute are hashed together to produce a stable value between 0 and 1, meaning the same user always receives the same variant assignment.

That consistency matters for AI experiments specifically: variant-switching mid-experiment would corrupt any behavioral signal you're trying to measure.

When running multiple segment-specific AI experiments simultaneously — say, one for Spanish-locale users and a separate one for free-tier users — namespaces prevent users who fall into both segments from being enrolled in conflicting experiments, which would otherwise introduce noise into both results.

Warehouse-native analysis and interpreting heterogeneous treatment effects

Segment-level analysis requires joining experiment assignment data with the full user attribute table. For sensitive segment identifiers — locale, health status, plan tier, organization ID — that join needs to happen in an environment where PII doesn't leave your control.

GrowthBook's warehouse-native architecture keeps analysis inside the customer's own data warehouse (Snowflake, BigQuery, Redshift, Postgres), which means the experiment infrastructure never requires access to the sensitive attributes that define your segments.

When interpreting results, the signal to look for is not just whether the overall treatment effect is statistically significant — it's whether the treatment effect direction or magnitude differs across segments. A positive aggregate effect paired with a negative segment-level effect is not a minor footnote. It's a fundamentally different rollout decision.

For smaller segment populations where sample sizes constrain statistical power, variance reduction techniques and continuous monitoring methods — both covered in detail in the metrics section below — are configurable at the experiment level, giving teams tools to detect real effects in cohorts that would otherwise require impractically long run times.

Character.AI's Head of Post-Training, Landon Smith, described using GrowthBook to "compare different modeling techniques from the perspective of our users — guiding our research in the direction that best serves our product." That framing captures the intent precisely: segment-aware experiment design isn't about finding statistical significance in aggregate. It's about understanding which users a model actually serves well, and making deployment decisions accordingly.

Choosing the right success metrics for each user segment in AI experiments

Running a well-structured AI feature experiment with carefully defined segments still fails if you're measuring the wrong thing for each group. Metric selection isn't a one-time decision you make at the experiment level — it's a decision you need to make for each segment independently, before you launch.

The right primary metric for a power user is often structurally irrelevant for a casual one, and applying a single measurement framework across all cohorts is how teams end up declaring a successful experiment that quietly damaged a high-value user group.

Aligning metrics to segment behavior, not segment demographics

The instinct is to define segments by who users are — their plan tier, their geography, their company size — and then apply the same success metric to all of them. The problem is that users in the same demographic cohort can have entirely different behavioral goals within the same AI feature.

A power user invoking an AI writing assistant to produce a first draft has a different success signal than a casual user who opened the same feature out of curiosity. For the power user, task completion rate and output accuracy are meaningful. For the casual user, feature abandonment rate and session engagement tell you far more about whether the experience is working.

This distinction matters because AI outputs are evaluated differently depending on what the user was trying to accomplish. Aggregate metrics flatten those differences. A neutral overall result on "time to task completion" might reflect genuine improvement for one behavioral cohort and genuine degradation for another — the two effects cancel each other out in the aggregate, and you ship a feature that harms a segment you care about.

Guardrail metrics as early warning systems for high-value cohorts

The most costly version of this failure mode involves high-value cohorts — enterprise accounts, power users, non-English speakers — where harm is invisible in aggregate results but significant within the segment. This is precisely where guardrail metrics earn their place in the experiment design.

A guardrail metric isn't a primary success metric. It's a metric you're not actively trying to move, but whose degradation would be a serious problem. GrowthBook's experimentation framework makes this concrete: guardrail results appear beneath the main goal and secondary metrics with full statistics, and the frequentist engine uses color-coded thresholds to communicate severity.

Yellow indicates the metric is moving in the wrong direction regardless of statistical significance. Red indicates it's moving in the wrong direction with a p-value below 0.05. When guardrail metrics reach significance, the documented guidance is to consider ending the experiment — it functions as a kill-switch trigger.

Applied at the segment level, this means defining a guardrail metric specifically for your highest-risk cohort before launch. If you're testing an AI summarization feature, your aggregate goal metric might be user satisfaction score. Your guardrail for enterprise accounts might be retention rate or feature usage frequency — metrics that would show harm even if the overall satisfaction numbers look fine.

Handling small per-segment sample sizes

Filtering an experiment to a specific segment reduces your sample size, which reduces statistical power and increases the risk of both false positives and false negatives. This is the practical constraint that makes segment-level metric analysis harder than it looks.

Two statistical methods address this directly. CUPED (Controlled-experiment Using Pre-Experiment Data) reduces metric variance by controlling for pre-experiment user behavior, which effectively increases the sensitivity of your analysis without requiring more users. Sequential testing allows you to monitor results continuously and make decisions as data accumulates, rather than waiting for a fixed sample size that a small segment may never reach. Both methods are available in mature experiment platforms and represent the current standard for handling underpowered segment analyses.

Minimum data thresholds per metric are a practical safeguard that prevents teams from drawing conclusions when segment sample sizes are genuinely too small to be meaningful. The instructive example: a result based on 5 versus 2 conversions should not be treated as signal, and experiment tooling can be configured to enforce that.

Pre-specifying metrics before you launch

The false positive risk compounds when you analyze many metrics across many dimensions after the fact. The more metrics and dimensions you examine, the more likely you are to encounter a spurious result. The corrective is pre-specification — deciding, before the experiment runs, which metric is primary for each segment, which metric serves as the guardrail, and what sample size that segment is realistically going to generate.

That last question constrains your options. If a segment will produce only a few hundred observations over your experiment window, a metric with low baseline conversion rate may never reach detectable effect sizes. Choosing a higher-frequency behavioral metric — one that fires more often per user — may be the only viable path to a valid result. Pre-specifying forces that conversation before launch, when you can still change the design.

Turning segment testing from a one-off project into a default release practice

Design principles don't ship products — workflows do. The preceding sections of this article establish why AI performance diverges across user segments and how to structure experiments that surface those differences. This section is the operational payoff: a repeatable cycle that turns segment testing from a one-time project into a durable practice.

The goal isn't to run one good experiment. It's to build the infrastructure that makes every AI feature release segment-aware by default.

The repeatable workflow: define, gate, monitor, decide per cohort

The fundamental shift in AI feature segmentation is that the decision unit changes. You're no longer asking "should we ship this feature?" You're asking "should we ship this feature to this cohort?" That reframe has direct operational consequences.

The cycle has four steps, each building on the last. Segment definition comes first: identify the user attributes that correlate with model performance differences — locale, expertise level, plan tier, organization, behavioral intent. From there, gate AI variants using feature flags with targeting rules that map to those attributes. As the rollout progresses, monitor segment-level metrics in your data warehouse rather than waiting for a final readout.

The last step is making per-cohort decisions — expand to 100%, modify the variant, or roll back without touching a deployment — then repeat this cycle for the next AI feature release.

The key discipline is resisting the pull toward a single global decision. A new summarization model might be a clear win for English-speaking power users and a measurable regression for Spanish-speaking casual users in the same experiment. A global ship decision harms one cohort while rewarding another. The workflow above forces the decision to happen at the cohort level, where the signal actually lives.

Feature flag architecture for AI variants

Feature flags serve as the control plane for AI variant delivery, and their architecture matters for experiment integrity. Deterministic hashing — where the same user ID always resolves to the same variant — is non-negotiable for AI segment tests. Without it, a user might see different model outputs across sessions, contaminating both the user experience and the exposure data.

GrowthBook uses MurmurHash3-based deterministic hashing, ensuring consistent bucketing across the full experiment window.

The targeting layer needs to be expressive enough to encode the segment dimensions that matter for AI. AND/OR attribute rules, saved groups for reusable cohort definitions, and B2B organization-level targeting all let you gate AI variants against the exact user populations you've pre-specified as hypotheses.

When a flag is evaluated locally from a cached JSON payload — rather than via a synchronous third-party call — it's also safe to use in AI inference paths where latency is a constraint. GrowthBook SDKs download flag rules as a locally cached JSON payload and evaluate every flag check in-process with zero network latency, so flag checks resolve in sub-millisecond time and your application continues to function correctly even if GrowthBook's servers are temporarily unavailable.

Feature evaluation diagnostics close the operational loop on the engineering side. When a flag behaves unexpectedly — a user receiving the wrong AI variant, a targeting rule firing incorrectly — developers need visibility into which rules evaluated and what outcome they produced. GrowthBook's developer tools expose exactly this: which features are active, how the rules were evaluated, and the ability to manually switch variations for debugging.

Continuous monitoring and the kill-switch mechanic

AI degradation doesn't trigger error alerts. A model producing degraded outputs for users with complex, high-volume inputs won't throw a 500 error or spike your latency dashboard — it will quietly show up as lower task completion rates, higher abandonment, or reduced return sessions within that cohort. That's why continuous behavioral monitoring at the segment level is the detection mechanism, not infrastructure alerting.

The monitoring architecture that supports this keeps segment-level metrics in your own data warehouse. GrowthBook queries experiment exposure data against your existing event tracking — no PII leaves your environment — and surfaces segment-level results as the rollout accumulates data.

Sequential testing enables teams to monitor those results continuously without inflating false positive rates, which means you can make an earlier kill decision when a specific segment is being harmed rather than waiting for a pre-determined sample size to complete.

The kill-switch capability is what makes this operationally viable. GrowthBook can instantly deactivate an underperforming AI feature for a specific segment without a code deployment. That's the "Confident AI Releases" model in practice: you're not choosing between shipping and not shipping — you're choosing which cohorts receive which variant, and you can change that decision in real time.

Who owns each step, and why that division prevents the workflow from collapsing

Sustainable segment testing requires clarity on who owns each step of the cycle. Role-based access lets PMs monitor segment-level experiment results and flag anomalies without requiring engineering support for every query. Data scientists can define the metrics and statistical methods. Engineers own flag configuration and the targeting rule architecture.

A cross-functional experiment platform — where product, engineering, and data science operate from shared infrastructure rather than siloed tooling — is what makes this division of ownership durable rather than theoretical.

Flag and experiment creation directly from the IDE, using plain English prompts, removes the context-switching overhead that causes teams to skip the flagging step under deadline pressure. Launch checklists, validation hooks, and prerequisite flags make segment review a required checkpoint in the release process — not an optional step that gets skipped under deadline pressure.

The specific review cadence will vary by team and release velocity. What matters more than the interval is that the cadence exists and is tied to the rollout stages: flag creation at build time, segment monitoring during gradual rollout, cohort-level ship/modify/kill decisions at defined percentage thresholds.

Character.AI's Landon Smith described the operational intent directly: "compare different modeling techniques from the perspective of our users — guiding our research in the direction that best serves our product." That's what the workflow above is designed to produce: model decisions guided by segment-level user evidence, not aggregate intuition.

Where to start when your last AI experiment didn't break out segment results

The through-line of this article is simple: AI performance is not a property of a model. It's a property of a model in contact with a specific user population. That reframe changes what "testing" means. You're not validating that a feature works — you're validating that it works for each cohort that will encounter it, and that the cohorts most likely to be harmed are the ones you've looked at most carefully.

The minimum viable segment testing stack: what you need to get started

You don't need a sophisticated data platform to start. You need three things: a way to target feature variants by user attribute, a way to define segment membership before the experiment runs, and a metric that fires frequently enough to detect an effect within a realistic sample size.

If you have feature flags with attribute-based targeting, a data warehouse you can query, and a behavioral metric that isn't conversion-rate-shaped, you have enough to run your first segment-aware AI experiment.

Prioritizing which AI features and segments to test first

Start with the AI feature that has the largest gap between aggregate satisfaction and qualitative complaints. That gap is usually where a segment-level failure is hiding. Then pick the one segment dimension most likely to correlate with model performance for that feature — language and locale if your model's training data skews English, expertise level if your feature's value proposition assumes domain knowledge.

One feature, one segment hypothesis, one guardrail metric for your highest-risk cohort. That's a complete first experiment.

From one-off experiment to continuous AI quality program

The operational shift that makes this sustainable is changing the decision unit from "should we ship this feature?" to "should we ship this feature to this cohort?" Once that question is the default, segment review becomes part of the release process rather than a separate project.

Saved segment definitions and warehouse-native analysis make that operationally tractable — you're reusing segment definitions across experiments and keeping sensitive attribute joins inside your own environment, not rebuilding the infrastructure each time.

The honest thing to say is that this takes a few experiment cycles to feel natural. The first time you run a segment-aware experiment and find a negative effect in a cohort that your aggregate result missed, the workflow will feel worth it. That's the moment the practice becomes a habit.

This article is meant to give you enough of the reasoning — not just the mechanics — to make that first experiment a good one, and to build from there.

What to do next: Pull the last AI feature experiment your team shipped and look at whether you broke out results by language, expertise level, or plan tier. If you didn't, that's your starting point — not a new experiment, but a reanalysis of an existing one. If you have the segment data in your warehouse, run the breakdown now. If the segment-level result differs from your aggregate, you've just identified the first hypothesis for your next experiment. If you don't have the segment data, that's the instrumentation gap to close before the next AI feature ships.

Related insights

Sign up for free

Take Growthbook for a spin, no credit card required.

Create my account

Example H2

See All Articles

Experiments

Feature Flags

CLI vs MCP: TLDR, it doesn’t matter as much as you think

Oct 21, 2026

min read

Your agent should know how to do the job whichever interface it uses.

Your coding agent can already run terminal commands. Your product manager's assistant can connect to remote tools. Both need to inspect the same experiment, understand the same metric, and prepare changes to the same feature flag.

That creates a familiar question: should you give them a CLI or an MCP server?

The usual answers mix several concerns. A CLI is easy to compose with scripts. MCP gives compatible clients a standard way to discover and invoke tools. Skills can supply the instructions that make either route useful. None of those statements tells you whether the agent knows your release process or can distinguish a draft from a live change.

For GrowthBook, the choice has become smaller. We've built the agent layer around a REST API and reusable workflow skills, then made those available through different interfaces. Your choice of CLI or MCP no longer needs to determine which product workflow your agent can understand.

It still determines how the agent connects, executes requests, and handles credentials. This article explains that distinction, shows the same disabled-flag operation through both routes, and gives you a way to evaluate the integration you actually intend to use.

The debate combines three different questions

Start with a specific job: prepare a release control for a new checkout flow. The agent needs to know which project owns checkout, which environments exist, what the default behavior should be, and whether the proposed flag key already exists.

Choosing an interface answers only part of that job.

How can the agent reach the product?

A command-line interface lets an agent invoke a program in an environment where it has shell access. That program can make authenticated API requests and return data. You can also run the same commands yourself, put them in scripts, and preserve output for inspection.

The Model Context Protocol architecture connects a host application to servers through clients. An MCP server exposes capabilities that a compatible host can discover and call. The host might be a coding environment, a desktop assistant, or another agent application.

These categories overlap. A terminal-based agent can also be an MCP client. A local MCP server can be a process launched by that client. A remote MCP connection can work without the host giving the model a general-purpose shell.

So “terminal agent” doesn't automatically mean “use the CLI,” and “MCP” doesn't automatically mean “remote service.” Ask what access the runtime provides before choosing the adapter.

What procedure should the agent follow?

An API can expose a create operation without teaching an agent when to use it. A syntactically valid feature flag might still have the wrong project, an ambiguous key, or a default value that doesn't match the application.

That's the job of a skill: reusable instructions for a recognizable task. The Agent Skills format gives those instructions a portable structure, including a skill entry point and supporting references. A workflow can describe the context to inspect, the sequence to follow, and the evidence to return.

For the checkout example, useful guidance says to resolve environments, check the permanent key, choose a value type, and create the flag disabled. Merely exposing a create command leaves those decisions to the agent's general knowledge.

Skills are also useful when the right answer is to take no action yet. “Create a flag” and “design an experiment” are related requests, but they can require different objects and sequencing. An agent needs enough product guidance to make that distinction.

What context makes the procedure relevant?

Even a good playbook needs the current product state. The agent should inspect the objects in your GrowthBook organization and, when it is changing application code, the repository's SDK initialization and tests.

Those are different sources of context. GrowthBook supplies flags, experiments, metric definitions, and other authorized product resources. Your coding runtime supplies the application files it can access. An MCP connection doesn't by itself grant repository access, and a CLI doesn't by itself explain your metric definitions.

Evaluate the combination: access to the product, a maintained procedure, and enough context to apply that procedure. A comparison that changes all three at once cannot isolate the effect of CLI versus MCP.

GrowthBook puts the durable parts below the interface

We separated the agent layer into responsibilities that can be reused. The product API exposes operations. Skills explain workflows. The interface connects the agent to those two layers.

This is the architectural shift behind GrowthBook's agent experience. Instead of making every client integration carry its own interpretation of experimentation, we can share the interpretation and change how it is delivered.

The API provides the operations

GrowthBook's REST API makes product objects addressable. An agent can inspect a project, retrieve a flag, or perform an authorized operation against those objects.

The rebuilt API described in GrowthBook 4.4 uses shared schemas to support types, validation, and API documentation. The CLI is generated from the API definition. That reduces the amount of separately maintained command logic needed to expose product operations.

The OpenAPI specification explains the underlying principle: a machine-readable description can represent an HTTP API's paths, operations, inputs, and responses. Generation is useful because it derives another interface from that contract.

Generation doesn't mean any installed binary knows every future endpoint. The generated client still has a version, and your GrowthBook server still has a version. A self-hosted instance can lag behind the public documentation. Treat API compatibility as a deployment fact you verify.

Skills provide the workflow guidance

GrowthBook's feature-flags skill routes requests to more specific workflows. Creating a flag, inspecting dependencies, changing targeting, requesting review, and publishing a revision are separate procedures.

That structure lets the agent load the guidance relevant to the current job. The skill authoring guidance on progressive disclosure supports this approach: keep the entry point focused and place detail in references that are loaded as needed.

The benefit is maintainability as well as context management. A correction to a flag workflow can be made in the maintained skill rather than repeated across a series of tool descriptions and client-specific prompts.

There are two shell concepts to keep distinct. The generated growthbook CLI is a terminal interface to the product API. The skills repository also includes a small gb-call helper for making REST requests. Installing a skill doesn't make that helper a globally available command; its location depends on the installation. Both can reach the API, but their configuration and invocation are different.

MCP provides a thin bridge

The current GrowthBook MCP implementation exposes four tools: growthbook_list_skills, growthbook_read_skill, growthbook_api_read, and growthbook_api_write.

An agent discovers a domain, reads the relevant workflow, then makes API calls. It doesn't need a dedicated MCP tool for every product endpoint. Calls described as gb-call GET in a skill map to the read tool; mutating methods map to the write tool with the same path and body.

The bridge does not run the shell helper. It forwards authenticated requests. Product-side permissions and validation remain important because a small tool surface can still reach many operations.

Shared guidance also has a delivery boundary: the MCP server bundles skills when it is built, using a pinned source revision. A newly edited skill in the repository needs to be included in a subsequent server build before that server supplies it. “Shared source” is a maintainability claim, not a promise of instant synchronization.

Explore the shared flag workflows

Open the feature-flags skill to see the procedures that guide creation, targeting, reviews, and publishing. Bring the same workflow knowledge to the agent interface your team uses.

View the Flag Skill

The same disabled flag can use either route

The practical question is whether the interface changes the job. Let's create a Boolean flag called checkout-redesign, with a false default and every environment disabled.

This is an API example for a sandbox project, not an application integration tutorial. It creates a control object; your application still needs a correctly configured GrowthBook SDK to evaluate it. The feature flag documentation covers the product concepts behind that integration.

Connect and verify before preparing the write

For the terminal route, install the current GrowthBook CLI, configure a Personal Access Token through its interactive setup, and inspect the available command syntax:

npm install -g growthbook
growthbook configure
growthbook whoami
growthbook features create --help
growthbook features get --help

Check that the active server and identity are the ones you intend to use. Prefer a tested, pinned CLI version in repeatable automation. Don't include a live token in a command you plan to share.

For an MCP route, the GrowthBook Cloud setup uses the hosted endpoint and browser OAuth. In a client using the documented mcpServers configuration shape, such as Cursor, add:

{
  "mcpServers": {
    "growthbook": {
      "url": "https://mcp.growthbook.io/mcp"
    }
  }
}

Complete the browser authorization and confirm tool discovery. Other clients can use a different configuration shape. Self-hosted GrowthBook users should follow the local stdio or self-hosted HTTP instructions rather than point a private instance at the Cloud service.

Next, let the connected agent discover the flag-creation workflow. These are tool-call parameters, shown separately in the order to invoke them:

{"name":"growthbook_list_skills","arguments":{}}

{"name":"growthbook_read_skill","arguments":{"name":"feature-flags"}}

{"name":"growthbook_read_skill","arguments":{"name":"feature-flags/references/flag-create"}}

The returned guidance should describe the prerequisites for creating the flag. If the workflow isn't present, resolve the server's bundled skill version before substituting instructions from an older integration.

Resolve real IDs and construct one payload

The flag-creation workflow checks existing keys, lists projects, and resolves environments. In the MCP client, invoke the read tool with each of these argument objects:

For the CLI route, perform that discovery with the corresponding command groups documented in the CLI command catalog:

growthbook features list-keys --output-format json
growthbook projects list --output-format json
growthbook environments list --output-format json

Inspect each command's help if your installed version differs. The returned keys, project IDs, and environment IDs are the inputs to the payload below. In MCP, the equivalent read-tool arguments are:

{"path":"/api/v2/feature-keys"}

{"path":"/api/v1/projects"}

{"path":"/api/v1/environments"}

Verify that checkout-redesign isn't taken, identify the sandbox project's ID, and enumerate every actual environment. These read responses determine the payload; don't treat example environment names as a complete inventory of your organization.

Save the following as flag.json, replacing <PROJECT_ID> and the environment keys with the discovered values. This example assumes the only environments are staging and production:

{
  "id": "checkout-redesign",
  "valueType": "boolean",
  "defaultValue": "false",
  "description": "Control exposure to the redesigned checkout",
  "project": "<PROJECT_ID>",
  "environments": {
    "staging": { "enabled": false },
    "production": { "enabled": false }
  }
}

The API represents defaultValue as a string, even for a Boolean flag. Explicitly disabling each environment avoids relying on the organization's default state for newly created flags. Use a PAT for this example so the caller has a user identity for attribution.

Review the permanent key, project, value, and disabled states before submitting. If the intended job is an experiment, route to experiment design first instead of creating an unrelated flag and hoping to connect it later.

Submit once, then read the object back

With the CLI, the create command accepts a JSON body. In a Bash-compatible shell, inspect the request with dry run before submitting:

growthbook features create --body "$(cat flag.json)" --dry-run
growthbook features create --body "$(cat flag.json)" --output-format json

The first command previews a request; it doesn't prove the server will accept it. Run the second only after checking the preview and approving this creation. In PowerShell, use that shell's file-reading syntax rather than paste the Bash substitution unchanged.

The PowerShell equivalent, using the same reviewed file, is:

$flagPayload = Get-Content -LiteralPath .\flag.json -Raw
growthbook features create --body $flagPayload --dry-run
growthbook features create --body $flagPayload --output-format json

The same distinction applies: preview first, approve the specific creation, then submit. Neither command changes the application's deployed code.

Through MCP, call growthbook_api_write with the same payload serialized into its body string. Substitute the same resolved project and environment names:

{
  "method": "POST",
  "path": "/api/v2/features",
  "body": "{\"id\":\"checkout-redesign\",\"valueType\":\"boolean\",\"defaultValue\":\"false\",\"description\":\"Control exposure to the redesigned checkout\",\"project\":\"<PROJECT_ID>\",\"environments\":{\"staging\":{\"enabled\":false},\"production\":{\"enabled\":false}}}"
}

Choose one creation route. Don't run both against the same key; they target the same product object, and the second creation should encounter an existing key.

Verify through the CLI get command:

growthbook features get --id checkout-redesign --output-format json

Or invoke growthbook_api_read with:

{"path":"/api/v2/features/checkout-redesign"}

Check the returned key, project, value type, default value, disabled state of every environment, and absence of unintended rules. Report those facts rather than “success.” No targeting or exposure change is included in this example.

Choose the route that fits the runtime

Once the procedure and API operation are shared, the remaining choice is operational. Where does the agent run? What can it execute? How does it authenticate? What does your team need to inspect afterward?

Shell access favors terminal composition

A shell-capable agent can run the CLI alongside repository tools, save structured output, and combine product inspection with local checks. That is useful when the job includes code changes, generated files, or a reproducible script.

For example, a flag audit can produce an inventory file while an engineer separately checks code references. The agent should preserve the association between the GrowthBook key and each repository reference. The final artifact can include both the product state and the proposed code diff.

The shell makes composition available; it doesn't make a workflow safe by default. A script can select the wrong profile or swallow an error. Decide which failures stop the workflow, where outputs are stored, and which steps need a reviewable artifact.

Remote tool access favors MCP clients

MCP is useful when the host already provides a tool connection but doesn't expose a terminal. A PM can inspect an experiment through a compatible assistant without installing a product CLI inside that assistant's runtime.

The MCP authorization specification defines an authorization framework for HTTP transports. GrowthBook Cloud uses browser OAuth, while local stdio configurations use credentials supplied to the process. Those are distinct deployment paths.

MCP also gives hosts a structured description of tools. The tool specification includes annotations such as read-only and destructive hints. Hosts can use those signals in their interaction design, but the signals are not an access-control boundary. The product must still enforce the caller's actual permissions.

Use a small decision table

Your immediate requirement	Route to consider	What to verify
Combine product operations with local files and scripts	CLI plus maintained skills	Shell access, profile, installed version, structured output
Connect a compatible assistant without a shell	Hosted MCP	HTTP/OAuth support, identity, tools, bundled workflow
Keep requests inside a self-hosted deployment	CLI or self-hosted MCP	Network reachability, server version, credentials, trusted app origin
Reuse a procedure across different team clients	Shared skills through either route	Workflow revision, API compatibility, expected review artifact

You can support more than one row. An engineer's CLI and a PM's MCP connection can operate on the same objects with different identities and permissions. Standardizing the workflow is often more useful than standardizing everyone's client.

Compare actual outcomes, not tool counts

Four MCP tools can address a large API. A large CLI command catalog can expose the same operations. Counting entries tells you little about whether an agent completed the intended job correctly.

Likewise, there is no universal token or speed result implied by the architecture. The client, model, instructions, response size, pagination, and number of requests all affect a real run. Measure the work you need instead of importing a result from an unrelated integration.

Keep the evaluation task constant

Use one sandbox project and one bounded task. For a read test, ask each route to find a named flag and identify its enabled environments, current revision, and relevant dependencies. For a write test, use separate temporary keys with the same reviewed disabled-flag requirements.

Give both runs the same expected output. Record whether each identified the correct objects, loaded the relevant workflow, handled missing data, and preserved the boundary between creation and exposure. Measure elapsed time and context consumption only alongside those correctness checks.

If one run gets rich workflow guidance and the other gets only a command name, you're testing the guidance as well as the interface. That can still be a useful product evaluation, but describe what changed.

Test failure behavior deliberately

Try a nonexistent object, an existing key, and a credential without the required permission. The agent should identify the failure and stop the dependent action. It should not invent an ID, silently pick another environment, or reinterpret a permission error as permission to retry with a broader credential.

The revision workflow adds another important test: multiple open drafts. When a chain of changes belongs in one revision, the agent must preserve its version rather than repeatedly ask for whichever draft is latest.

A creation receipt and a publication receipt are different artifacts. If your evaluation only checks that a request returned without error, it can miss the point where a proposed change became live or where a draft failed to include the intended edits.

Record the versions behind the result

Capture the CLI version, server version when available, client configuration, and the skill text or revision used. With MCP, inspect the guidance actually returned by the connected server. With installed skills, inspect the local instructions actually loaded.

That record makes a later regression diagnosable. A changed result may come from new guidance, a client update, a changed API, or different product state. “MCP became worse” is too broad to help you fix it.

Build workflows that survive the next client

The most useful consequence of GrowthBook's design is organizational. Engineers, PMs, and analysts can choose the agent surface that fits their work while referring to the same flags, experiments, and metrics.

An engineer can prepare a disabled flag from a terminal. A reviewer can inspect its draft in the app. An assistant connected through MCP can retrieve the same object and explain its state. The handoff should carry an ID, a revision where applicable, the proposed behavior, and the decision still required.

This is where the CLI-versus-MCP question loses much of its importance. Your release procedure doesn't need to be rewritten around each client. The product holds state; the API exposes operations; skills explain the work; the chosen interface delivers access.

Start with one inspection or sandbox creation. Require the agent to return a concrete object and its verified state. Then add the next workflow only when the current one is reliable enough for your team to review.

GrowthBook's feature flag platform gives that work a common place to live. The durable choice is how you connect release control and measurement to your team's process. CLI and MCP are ways to carry that process into the tools people already use.

Start with a common foundation

Set up GrowthBook for the flags, experiments, and metrics your agents will share. Then choose the CLI or MCP connection that fits each team member's work.

Get Started With GrowthBook

Experiments

Analytics

Standard deviation vs variance: Key differences for A/B testers

Oct 21, 2026

min read

A metric can have wide user-to-user variation and still yield a precise experiment estimate.

That is why confusing standard deviation with standard error creates bad A/B testing decisions. A dashboard may show that users differ substantially while the average effect is estimated closely enough to guide a release. Those statements can both be true.

Variance and standard deviation describe spread. Variance works on a squared scale; standard deviation takes its square root to return to the metric's original units. Neither is automatically the uncertainty around an experiment result.

For A/B testers, the useful distinction is practical: use standard deviation to communicate outcome variability, use variance in the calculations that combine uncertainty, and inspect the standard error or interval for the treatment effect. Start by making sure all three describe the correct unit of analysis.

Understand the difference in units

Variance uses squared distances

Outcomes above and below a mean have deviations with opposite signs. Simply averaging those signed deviations gives zero. Variance avoids that cancellation by squaring each deviation before combining them.

The sample variance formula is s² = Σ(xᵢ − x̄)² / (n − 1), where x̄ is the sample mean and n is the number of observations. The symbol Σ means to add the terms across observations.

If the outcome is measured in seconds, variance is measured in seconds squared. That is mathematically useful but less intuitive for explaining a product experience. “Variance is 4 seconds squared” does not mean users differ by 4 seconds.

Standard deviation returns to the original scale

Standard deviation is s = √s². A variance of 4 seconds squared corresponds to a standard deviation of 2 seconds. Both describe the same dispersion, expressed on different scales.

Do not define standard deviation as the average absolute distance from the mean. That is a different measure. Squaring deviations gives large deviations more influence, and taking the square root afterward does not undo that weighting.

Quantity	What it describes	Units for task time
Variance	Squared dispersion in outcomes	Seconds squared
Standard deviation	Outcome dispersion on the original scale	Seconds
Standard error of the mean	Sampling variability of an estimated mean	Seconds

Neither variance nor standard deviation requires a normally distributed population merely to be calculated. Normality matters for particular interpretations and inferential procedures, not for the existence of these summaries when the required moments are finite.

Calculate both from the same observations

Work through a small dataset

Use four synthetic task times: 2, 4, 4, and 6 seconds. This tiny dataset is for arithmetic, not evidence about a real product or a basis for confident inference.

The mean is (2 + 4 + 4 + 6) / 4 = 4 seconds. Deviations from the mean are −2, 0, 0, and 2 seconds. Their squares are 4, 0, 0, and 4 seconds squared, totaling 8.

Treating these observations as a sample, the usual sample variance is 8 / (4 − 1) = 8/3, or about 2.667 seconds squared. The sample standard deviation is √(8/3), or about 1.633 seconds.

This is a useful check on a report. If someone describes the same sample's variance as 1.633, they have probably substituted the standard deviation. If the units are missing, the mistake is much harder to spot.

Choose sample or population formulas deliberately

If those four observations constitute the entire finite population you want to describe, population variance divides by 4: 8/4 = 2 seconds squared. Population standard deviation is √2, or about 1.414 seconds.

The population and sample distinction depends on what you are estimating. Having every recorded event from an experiment does not necessarily make those events the full population of future users or future conditions you care about.

The n − 1 denominator makes the usual sample variance an unbiased estimator of population variance under independent, identically distributed sampling with finite variance. It does not make the square root exactly unbiased for population standard deviation. Nor does it correct biased sampling or dependence between observations.

Verify the calculation with explicit function names

Python's statistics functions distinguish sample calculations from population calculations. With Python 3 installed, this standalone example requires no external packages or data:

from statistics import variance, stdev, pvariance, pstdev

times = [2, 4, 4, 6]
print(round(variance(times), 3), round(stdev(times), 3))
print(round(pvariance(times), 3), round(pstdev(times), 3))

The first line of output is 2.667 1.633; the second is 2 1.414. The different denominators explain the difference. Before comparing statistics from two tools, check their sample/population defaults and handling of missing values.

Find the noise you can reduce

Explore practical variance reduction methods, including metric design and CUPED, and the conditions that make each approach useful.

Read the Variance Reduction Guide

Use standard errors for experiment uncertainty

Distinguish users from estimated means

Standard deviation describes how users' outcomes vary. Standard error describes how an estimator would vary across repeated samples. For an independent sample, the estimated standard error of its mean is s / √n.

Suppose a user-level task-time metric has a sample standard deviation of 20 seconds among 10,000 independent users. The estimated standard error of the mean is 20/√10,000 = 0.2 seconds. Large differences between users coexist with a much smaller uncertainty around their average.

Collecting four times as many independent users would halve this standard error if the outcome variance stayed similar. It would not mechanically halve user-level standard deviation. The individual experience remains variable even as the estimated mean becomes more precise.

Combine arm variances before taking the square root

For independent treatment and control groups, a simple unadjusted difference in means has estimated variance s²ₜ/nₜ + s²꜀/n꜀. Its standard error is the square root of that sum. You do not add the two arm standard deviations and call the result the experiment's uncertainty.

Take a separate synthetic summary with 10,000 independent users in each arm and a 20-second standard deviation in both. The variance of the estimated difference is 400/10,000 + 400/10,000 = 0.08 seconds squared. Its standard error is √0.08, about 0.283 seconds.

If the observed treatment-minus-control difference is −1 second, an illustrative large-sample, unadjusted 95% normal interval is −1 ± 1.96 × 0.283, approximately −1.554 to −0.446 seconds. Negative values mean faster completion under this metric definition.

This example assumes independent groups and an appropriate large-sample approximation. It is not the exact output of GrowthBook's relative-lift, sequential, covariate-adjusted, or ratio-metric methods. GrowthBook's statistical details describe how the platform estimates effects and their uncertainty.

Avoid interpreting intervals as outcome ranges

That interval concerns the mean treatment effect. It does not say that 95% of individual users improved by between 0.446 and 1.554 seconds. An individual user's causal effect is a different quantity, and the experiment does not observe both potential outcomes for the same person.

Likewise, mean ± one standard deviation is not a universal statement about where 68% of users fall. That familiar interpretation depends on an approximately normal distribution. Skewed task times or revenue can look very different.

The frequentist confidence level describes the repeated-sampling coverage of the interval procedure under its assumptions. It is not a posterior probability assigned to a fixed realized interval.

Check what your metric actually measures

Match the variance to the randomization unit

If users are randomized and each contributes many requests, request rows are not automatically independent experimental units. Treating 100 requests from one person as 100 unrelated users can make uncertainty look smaller than it should.

Define the outcome at the intended unit or use a method that accounts for clustering. GrowthBook's aggregate data requirements are relevant when supplying precomputed summaries: the sample size, sum, and dispersion must correspond to the statistic and analysis method.

Two teams can report different standard deviations without either making an arithmetic mistake. One may summarize per-request latency while the other summarizes each user's average latency. They are describing different outcome distributions. Clarify the grain before comparing their numbers.

Paired measurements need another distinction. If the same users provide before-and-after outcomes, the measurements within each pair can be correlated. Analyze each user's difference or use an appropriate paired method; do not apply the independent-arm formula automatically. The paired-means framework treats those differences as the observations of interest. Pairing also does not remove time trends or establish causality by itself; the design still determines which comparison is defensible.

Inspect tails instead of deleting them casually

Large observations influence both variance and standard deviation. A few unusually long tasks can dominate a time metric; a few large purchasers can dominate revenue. Determine whether those values reflect valid behavior, instrumentation defects, retries, or unit errors.

A valid expensive user is not automatically an outlier to discard. Removing high values after seeing an experiment result can change the estimand and introduce selective analysis. Decide justified exclusions or transformations before interpreting the test.

Report complementary summaries when they answer useful questions: a histogram, median, or a relevant tail percentile can show what a single spread measure hides. Those summaries do not replace the mean effect if the mean remains the decision metric.

Keep units consistent across pipelines

Converting seconds to milliseconds multiplies every value by 1,000. Standard deviation also multiplies by 1,000, but variance multiplies by 1,000,000. A dramatic numerical variance increase can therefore be a unit conversion rather than a product regression.

For the four-value example, sample standard deviation becomes approximately 1,633 milliseconds and variance becomes approximately 2,666,667 milliseconds squared. The underlying task times have not changed. Store units with metric definitions and inspect them when comparing dashboards.

A lower variance caused solely by rescaling a metric does not create more statistical information. The effect size rescales too. Separate a numerical change in units from an actual improvement in the precision of your estimate.

Also avoid averaging standard deviations from daily reports and labeling the result a weekly standard deviation. The complete-period spread depends on both variation within days and differences between daily means. Keep the underlying observations or sufficient summaries that support the intended calculation. A simple average of displayed daily values generally loses information needed to reconstruct the full distribution.

Turn spread into a better testing plan

Define the outcome before planning the sample

Start with the metric definition: population, identity, aggregation, window, units, exclusions, and desired effect. Estimate variability from representative historical data at that same grain.

Historical variance is an input, not a guarantee. New traffic, a different user mix, or a changed measurement window can alter it. Explore plausible variance and effect-size assumptions rather than treating one historical number as exact.

Keep the two arm variances separate when appropriate. A treatment can change variability as well as its mean. The simple equal-standard-deviation illustration above is a teaching assumption, not a requirement that production groups have identical spread.

Reduce estimation noise without hiding behavior

CUPED and post-stratification use suitable pre-experiment information and covariates to improve effect estimation. CUPED stands for Controlled-experiment Using Pre-Experiment Data. Its value depends on useful predictive information and a valid implementation.

Adjustment can reduce uncertainty in the treatment effect without making real users' outcomes less variable. That distinction prevents a reporting error: a narrower adjusted interval is not evidence that customers now experience a more consistent product.

Also separate missing data from genuine zeros. A user with no purchase may correctly contribute zero revenue; a missing instrumentation event might represent an unknown outcome. Treating both identically can distort the mean and its variance before any statistical engine sees the data.

Report the right quantity for the decision

Reader's question	Useful quantity
How much do individual outcomes differ?	Standard deviation, with distribution context
What enters this uncertainty calculation?	Variance at the correct analysis unit
How precisely is the effect estimated?	Standard error and an appropriate interval
Is the measured difference worth acting on?	Effect size, uncertainty, guardrails, and decision context

For a clear experiment readout, show the metric's unit and analysis grain, the estimated effect, and its appropriate uncertainty interval. Add outcome spread when it helps readers understand the population or diagnose a noisy metric.

GrowthBook's experimentation platform brings defined metrics and statistical analysis together. A trustworthy interpretation still starts with your measurement choices: describe spread with standard deviation, calculate with variance where required, and judge precision using the estimator's uncertainty.

Build a clearer experiment readout

Connect your metrics to GrowthBook and evaluate effects with the analysis method and measurement definitions your team has chosen.

Get Started With Experimentation

Experiments

Analytics

How to calculate true positive rate (TPR)

Oct 21, 2026

min read

Catching 80 defects means little until you know how many defects existed.

Suppose a detector flags 125 suspicious software changes. Engineers inspect them and confirm 80 real defects. That tells you something about alert quality, but it does not tell you how many defects the detector missed. To calculate true positive rate, you need the actual positives outside the alert queue too.

True positive rate, or TPR, measures the share of actual positive cases that a system correctly identifies. It is also called recall or sensitivity. The formula is TP / (TP + FN), where TP means true positives and FN means false negatives.

The arithmetic is simple. Choosing the right population, labels, threshold, and denominator is the difficult part. This guide works through an engineering example, distinguishes neighboring metrics, and explains what changes when the “test” is an A/B experiment rather than an item classifier.

Start with actual positives, not alerts

Define the positive class

“Positive” means the condition you want to detect. It does not necessarily mean a good outcome. In a defect detector, positive means a real defect. In a spam filter, it means spam. Write that definition before calculating anything.

A true positive is an item that is actually positive and predicted positive. A false negative is actually positive but predicted negative. Those two groups together form the denominator: all actual positives in the evaluation sample.

The recall definition is therefore conditional on actual class membership. TPR asks, “Among the real defects, what proportion did we catch?” It does not ask, “Among the alerts, what proportion were real?”

Fix the unit and observation window

Choose a consistent evaluation unit. A code change, an individual finding, and a repository are different units. If one change generates 12 duplicate alerts, counting alerts in the numerator and changes in the denominator produces a meaningless rate.

For the example, each unit is one code change, evaluated once by one fixed detector version at one fixed threshold. An independent review process eventually labels each change as defective or not defective. The observation window is long enough for those labels to mature.

This resembles the discipline needed when defining experiment metrics: the event, identity, time window, and denominator must describe the same question. A correct formula applied to incompatible counts still gives the wrong answer.

Calculate TPR from the confusion matrix

Put every evaluated item in one cell

Consider an original synthetic evaluation of 1,000 code changes. Reviewers establish that 100 contain a defect and 900 do not. The detector catches 80 defects, misses 20, and incorrectly flags 45 nondefective changes.

Actual label	Predicted positive	Predicted negative	Total
Defective	80 true positives	20 false negatives	100
Not defective	45 false positives	855 true negatives	900
Total	125	875	1,000

Read the labels rather than relying on cell position. Libraries and reports can arrange axes differently. The scikit-learn confusion matrix convention uses actual classes on rows and predicted classes on columns, with class order also affecting where each value appears.

Check the totals before calculating rates. The four cells sum to 1,000. True positives plus false negatives sum to 100 actual defects. Predicted positives total 125, which is a different denominator.

Divide detected positives by all actual positives

The calculation is:

TPR = TP / (TP + FN) = 80 / (80 + 20) = 80 / 100 = 0.80 = 80%.

Interpret it in words: the detector caught 80% of the actual defects in this evaluation sample. It missed 20%, so the false negative rate is 20%. For a binary task with complete labels, TPR and false negative rate sum to 100%.

Do not divide 80 by 1,000. That would measure true positives as a share of all evaluated changes, not recall. Do not divide 80 by 125 either. That is precision, which answers the alert-quality question.

The result describes this sample and threshold. It does not establish that every future workload will have the same recall, or that all defect categories are equally detectable. Keep the counts beside the percentage so readers can see the support behind it.

Handle zero positive support explicitly

If the sample contains no actual defects, TP + FN equals zero. Sample TPR is undefined. Returning 0% would imply the detector missed known defects; returning 100% would imply it successfully found them. Neither interpretation is supported.

Software libraries need a practical zero-division convention. That convention can return a placeholder or warning, but it does not create evidence. Report “TPR unavailable: zero actual positives” and preserve the support count. A pipeline can then distinguish an unevaluable slice from measured poor performance.

Check the experiment error risks

See how power, repeated checks, and multiple comparisons affect the reliability of experimental decisions beyond a single classifier score.

Read the False Positives Guide

Compare TPR with precision and false positives

Keep the denominators visible

For the same 1,000 changes, the other common metrics tell different stories:

Metric	Calculation	Result
True positive rate / recall	80 / 100 actual positives	80%
Precision	80 / 125 predicted positives	64%
False positive rate	45 / 900 actual negatives	5%
Specificity	855 / 900 actual negatives	95%
Accuracy	(80 + 855) / 1,000 changes	93.5%

The classification metric definitions differ because they condition on different populations. Here, 36% of alerts are false, but the false positive rate is 5%. Calling both numbers “the false positive rate” hides the operational distinction.

Accuracy can also look reassuring in an imbalanced population. A detector that labels all 1,000 changes nondefective would be 90% accurate while detecting none of the 100 defects. TPR exposes that failure, though it still needs the false-positive side of the picture.

Choose a threshold for the decision

Suppose a stricter threshold, evaluated on the same labeled changes, catches 60 defects and incorrectly flags 10 nondefective changes. Its TPR is 60%, its precision is about 85.7%, and its false positive rate is about 1.11%.

This threshold reduces the review queue from 125 changes to 70. It also misses 40 defects instead of 20. Whether that tradeoff is acceptable depends on the consequences of a missed defect, the review team's capacity, and what action follows an alert.

A ROC curve plots TPR against false positive rate across thresholds. It helps compare operating points. It does not pick the operational threshold for you, and an area-under-curve summary does not specify the queue size at the threshold you intend to deploy.

If you tune thresholds using the same labels you later report as a final evaluation, the score can be optimistic. Use separate validation and held-out evaluation data, or another defensible evaluation design, and freeze the decision rule before the final assessment.

Specify the action at each threshold too. A low-confidence finding might create a suggestion, while a high-confidence finding might block a merge. Those are different interventions with different error costs. One aggregate recall figure should not imply that every detected item received the same treatment.

Make sure the labels can support the claim

Audit predicted negatives too

If engineers review only alerts, you can estimate precision within the reviewed alert population. You cannot directly count missed defects among unreviewed negatives. Assuming those changes are clean makes FN artificially small and TPR artificially high.

An evaluation may need independent review of a representative sample of predicted negatives, delayed incident labels, or a curated benchmark with known defects. Each approach has limits. Incident labels miss undetected failures; curated examples can differ from ordinary production work.

If the evaluation oversamples suspicious changes or certain repositories, report that design. Use appropriate weighting when estimating performance for the real workload. A simple unweighted confusion matrix describes the sampled dataset, which may not represent the deployed population.

Preserve uncertainty and segment support

An 80% estimate based on 4 detected defects out of 5 is much less precise than one based on 800 out of 1,000. The important sample size for recall is the count of actual positives, not just the total number of changes evaluated.

Report positive support by important slices, such as language, defect category, or repository type. A high pooled TPR can hide weak performance on a rare but expensive failure mode. At the same time, avoid declaring tiny slices definitively better or worse from a few examples.

Counts alone do not make observations independent. Many similar changes from one repository can share a common pattern. Use uncertainty methods that fit the evaluation design rather than assuming every row is an unrelated trial.

Separate label delay from label absence

A recently deployed change might appear nondefective because nobody has had time to find the bug. Treating it as a confirmed negative gives the newest cohort a different labeling standard from older cohorts.

Define when an item becomes eligible for evaluation and track unresolved labels separately. Keep model version, threshold, label version, and observation dates with the result. That makes a later score change interpretable: you can tell whether the detector changed, the workload changed, or the labels matured.

For a warehouse-native measurement workflow, these definitions belong in the underlying data model. Joining model decisions to review outcomes still requires identity and time alignment; a queryable warehouse does not automatically resolve ambiguous ground truth.

Keep comparable populations together

When comparing detector versions, evaluate them on the same held-out cases where possible. Otherwise, an apparent recall improvement could come from an easier mix of defects. Preserve a breakdown of new, removed, and unchanged cases so reviewers can see what changed between evaluations.

When combining reports, add the underlying TP and FN counts for nonoverlapping, consistently defined populations, then calculate the pooled rate. An unweighted average of repository percentages answers a different question because it gives a tiny repository the same influence as a large one. Either summary can be useful, but label the weighting explicitly.

If the same change appears in several reports, deduplicate or model that repeated observation before pooling. Repeated scans do not create additional independent defects. Keep an audit trail from the final counts back to the evaluated item identifiers so a suspicious rate can be investigated.

Connect detection quality to product decisions

Distinguish classifier recall from statistical power

In hypothesis testing, statistical power is the probability that a testing procedure rejects the null when a specified alternative is true. It is a repeated-test detection probability, depending on assumptions such as effect size, variance, sample size, and decision rule.

That is related to the idea of detecting a real signal, but it is not the same dataset as a labeled-item confusion matrix. In an ordinary A/B program, you do not observe the true causal effect of every experiment with certainty. Dividing declared winners by all experiments does not reveal a ground-truth TPR.

Power planning asks how often your procedure would detect a particular real effect under its assumptions. A classifier evaluation asks how many labeled positives were detected in the evaluated sample. Keep those questions separate when discussing GrowthBook's statistical methods.

Evaluate the intervention as well as the detector

An improved offline detector can still create a worse product workflow. More alerts might overwhelm engineers, delay releases, or catch defects too late to matter. Conversely, a modest recall change might be valuable if it catches the failures with the highest operational cost.

When appropriate, use randomized experiments to evaluate the downstream intervention: which review experience helps teams resolve defects, with what delay and cost? Keep detector recall, review workload, and product outcomes as separate measures.

GrowthBook's experimentation platform can analyze defined outcomes from such tests. Your team supplies reliable assignment, exposure, labels, and metric definitions. Neither a high TPR nor a significant product result eliminates the need to inspect the other evidence.

Use a repeatable reporting checklist

Before sharing a TPR number, state the positive class, evaluation unit, detector version, threshold, and label process. Show TP and FN beside the rate. Include precision or false positive rate so readers can assess the workload created by detection.

Then ask whether predicted negatives were evaluated, whether labels had time to mature, and whether the sample reflects the population where you will act. Mark zero-support slices unavailable and explain material gaps in coverage.

The useful result is a statement someone can audit: “At this threshold, the detector found 80 of 100 labeled defects in these 1,000 evaluated changes, while producing 45 false alerts.” That is a clearer basis for a decision than “our model is 80% accurate.”

Measure what the change delivers

Connect your data and defined metrics to GrowthBook to evaluate how a product or workflow change affects the outcomes your team cares about.

Start Measuring With GrowthBook

Ready to ship faster?

No credit card required. Start with feature flags, experimentation, and product analytics—free.

Get Started

Book a Demo

Simplified white illustration of a right angle ruler or carpenter's square tool.

White checkmark symbol with a scattered pixelated effect around its edges on a transparent background.

An AI feature that looks successful in your aggregate metrics can be quietly failing a specific group of users at the same time.

Why AI features don't perform the same way for every user

AI outputs are probabilistic, not deterministic

The dimensions along which AI performance diverges

Why aggregate metrics hide segment-level failures

What "working" actually means depends on who you ask

Segment selection happens before the experiment, not after the results disappoint you

Which dimensions actually correlate with AI model performance

User expertise and use-case intent as high-signal dimensions

Language, locale, and plan tier — operationalizing them in practice

The cost of getting segmentation wrong

A single global A/B test is structurally incomplete for AI features

The aggregate-masking problem

Pre-specifying segment hypotheses before the experiment runs

Feature flag gating by user attribute

Warehouse-native analysis and interpreting heterogeneous treatment effects

Choosing the right success metrics for each user segment in AI experiments

Aligning metrics to segment behavior, not segment demographics

Guardrail metrics as early warning systems for high-value cohorts

Handling small per-segment sample sizes

Pre-specifying metrics before you launch

Turning segment testing from a one-off project into a default release practice

The repeatable workflow: define, gate, monitor, decide per cohort

Feature flag architecture for AI variants

Continuous monitoring and the kill-switch mechanic

Who owns each step, and why that division prevents the workflow from collapsing

Where to start when your last AI experiment didn't break out segment results

The minimum viable segment testing stack: what you need to get started

Prioritizing which AI features and segments to test first

From one-off experiment to continuous AI quality program

Related insights

Sign up for free

Table of Contents

Related Articles

CLI vs MCP: TLDR, it doesn’t matter as much as you think

Your agent should know how to do the job whichever interface it uses.

The debate combines three different questions

How can the agent reach the product?

What procedure should the agent follow?

What context makes the procedure relevant?

GrowthBook puts the durable parts below the interface

The API provides the operations

Skills provide the workflow guidance

MCP provides a thin bridge

Explore the shared flag workflows

The same disabled flag can use either route

Connect and verify before preparing the write

Resolve real IDs and construct one payload

Submit once, then read the object back

Choose the route that fits the runtime

Shell access favors terminal composition

Remote tool access favors MCP clients

Use a small decision table

Compare actual outcomes, not tool counts

Keep the evaluation task constant

Test failure behavior deliberately

Record the versions behind the result

Build workflows that survive the next client

Start with a common foundation

Standard deviation vs variance: Key differences for A/B testers

A metric can have wide user-to-user variation and still yield a precise experiment estimate.

Understand the difference in units

Variance uses squared distances

Standard deviation returns to the original scale

Calculate both from the same observations

Work through a small dataset

Choose sample or population formulas deliberately

Verify the calculation with explicit function names

Find the noise you can reduce

Use standard errors for experiment uncertainty

Distinguish users from estimated means

Combine arm variances before taking the square root

Avoid interpreting intervals as outcome ranges

Check what your metric actually measures

Match the variance to the randomization unit

Inspect tails instead of deleting them casually

Keep units consistent across pipelines

Turn spread into a better testing plan

Define the outcome before planning the sample

Reduce estimation noise without hiding behavior