Experiments

How to run experiments on AI features without breaking UX

Jun 3, 2026

min read

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

Shipping an AI feature to 100% of your users on day one is the fastest way to find out your model behaves very differently in production than it did in your evals.

The failure mode isn't always dramatic — it's often subtle. A model that performs well on benchmarks starts producing outputs that feel slightly off to real users. Trust erodes quietly, and by the time aggregate metrics surface the problem, the damage is already done.

The core argument of this article is simple: you can experiment on AI features without breaking UX, but only if you treat AI features as fundamentally different from UI changes — because the failure modes are different, the metrics that catch them are different, and the cost of getting it wrong is higher.

This guide is for engineers, PMs, and data teams who are shipping AI features and want a practical system for testing changes safely. It covers the four building blocks you need before any AI experiment goes live:

Feature flags — how to gate AI features at runtime so you can control exposure without a deploy
Incremental rollouts — how to ramp traffic gradually and catch quality regressions before they reach your full user base
Guardrail metrics and fallback behaviors — what to monitor, what to commit to not harming, and how to define a safe fallback before launch
Measurement — which metrics actually capture AI feature quality, and what to do when you realize mid-experiment you tracked the wrong thing

Each section builds on the last. By the end, you'll have a concrete framework for how to experiment on AI features without breaking UX — one that gives you the speed to ship and the control to pull back fast when something goes wrong.

Why AI features demand a different approach to experimentation

When a Google Pixel Watch update locked core watch functionality behind an AI data-sharing consent screen, users didn't file bug reports — they posted to Hacker News. A thread titled "Don't push AI down our throats" accumulated 431 points and 358 comments, with practitioners describing AI updates that silently broke working features: music playback that stopped responding, assistant behaviors that changed without warning, functionality that disappeared behind consent walls.

None of these failures would have shown up as a visible UI regression in a standard A/B test. All of them eroded trust in ways that a miscolored button never could.

That asymmetry is the starting point for understanding why AI features require a fundamentally different approach to experimentation.

The non-determinism problem

Standard A/B testing rests on a foundational assumption: the treatment is fixed. Variant B is always the same button color, the same headline copy, the same checkout flow. Every user in the variant group sees the same thing, which is what makes the statistical comparison valid.

AI features break this assumption by design. The same user, the same prompt, and the same context can produce meaningfully different outputs across requests. When you run an experiment on a conversational interface or a generative assistance feature, you're not testing a fixed treatment — you're testing a distribution of possible outputs.

Standard A/B frameworks assume the treatment is fixed — Variant B is always the same button, the same copy, the same flow. They measure the average difference between two stable things. AI breaks that assumption: Variant B is a model that might produce ten different responses to the same prompt, depending on context, phrasing, or random sampling.

A standard A/B test can tell you the average difference between the control and the AI variant, but it can't tell you whether that AI variant is behaving consistently across the full range of what it might say — or whether it's occasionally producing outputs that are wrong, off-topic, or damaging to user trust. A statistically significant result on the average can coexist with serious quality problems in the tail — and that's precisely why internal evals alone are insufficient. Internal evals can tell you how a model performs on a benchmark. They cannot tell you how users actually experience the variance in its outputs across millions of real interactions.

Trust erosion is asymmetric and sticky

A layout bug gets noticed and forgiven. A bad AI response gets remembered. The Pixel Watch incident and the broken Android assistant examples from that Hacker News thread share a common structure: an AI feature change degraded a non-AI core experience, with no visible interface regression to catch.

Users didn't experience a broken UI — they experienced a broken product that used to work. The backlash was disproportionate to the technical severity of the failure because it felt like a betrayal, not a bug.

This asymmetry matters for how you think about experiment risk. The downside of a bad UI experiment is a temporary conversion dip you can reverse with a rollback. The downside of a bad AI experiment can be lasting user resentment, public backlash, and churn from users who simply stop trusting the product.

There is no quantified recovery timeline for AI-driven trust erosion, but the community data makes clear that the recovery curve is longer and less predictable than it is for visual regressions.

Standard metrics can mask downstream harm

Even when teams do run controlled experiments on AI features, the metrics they reach for — click-through rate, session length, conversion — are poorly suited to detecting the failure modes that matter most. GrowthBook's experimentation documentation makes this explicit: "Relying solely on short-term metrics can encourage dark patterns in A/B testing, where you inadvertently exploit user trust to boost numbers temporarily at the expense of long-term retention."

With AI, this risk is structural. A model that produces more sensational or emotionally provocative responses may drive short-term engagement while accelerating long-term churn.

Aggregate metrics compound the problem further: a model that performs well on average may be producing degraded or harmful outputs for specific user segments — demographic, behavioral, or contextual — that the aggregate numbers simply don't surface.

The compounding cost of a poorly planned AI experiment

When the feature under test is an AI model, the costs of a bad experiment stack in ways they don't for UI tests. There are inference costs for a model that is underperforming at scale, user churn from trust erosion during the exposure window, and measurement costs when teams discover mid-experiment that they tracked the wrong proxies entirely.

The point is not that A/B testing is the wrong tool for AI features. The point is that standard A/B testing, applied naively to AI features without guardrails, gradual exposure, and metrics designed for probabilistic outputs, is insufficient. And the cost of getting that wrong is not proportional to the cost of getting a UI experiment wrong.

Feature flags are the control layer that makes everything else in AI rollouts possible

Before gradual rollouts, guardrail metrics, or fallback behaviors can do their job, you need a control layer that operates faster than your deploy pipeline. For AI features, that layer is feature flags — and the reason they matter isn't philosophical. It's operational.

A hardcoded branch can't close a four-minute incident window. A feature flag can.

A conditional if statement in your application code is static the moment it ships. Changing it requires a pull request, a review, a build, and a deploy — a cycle that, under normal circumstances, might take an hour. Under incident conditions, it takes longer.

A feature flag is evaluated at runtime, which means the individual feature — not the deployment — becomes the unit of control. You can change who sees what, and how much of your traffic is exposed, from a dashboard, in seconds, without touching the codebase.

For AI features, this distinction is not a convenience. Model behavior can degrade unpredictably. A prompt change, a model version update, or an upstream API shift can produce outputs that are wrong, slow, or actively harmful to user trust. The window between "something is wrong" and "users are affected" is measured in minutes. A deploy cycle cannot close that window.

One practical concern worth naming: GrowthBook SDKs download flag rules as a locally cached JSON payload and evaluate every flag check in-process with zero network latency, handling over 100 billion flag lookups per day. Adding a flag layer to an AI feature that already carries latency variance does not compound that variance — the evaluation overhead is negligible.

Controlled rollout mechanics and JSON flag payloads

Feature flags support more than on/off states. Flag types include Boolean, String, Number, and JSON — and for AI features, JSON flags are particularly useful. A single flag can carry a full model configuration as a structured payload: model name, temperature, retry count, fallback behavior.

That means a team can swap model parameters across user segments without deploying new code, and can version those configurations independently of the application.

Targeting rules let you define exactly who receives a given flag state — by user ID, account type, geography, plan tier, or any custom attribute your application surfaces. This is how you expose an AI feature to internal users first, then a beta cohort, then a controlled percentage of production traffic, each stage gated by explicit criteria rather than a coin flip.

Kill switches and instant rollback

When a GrowthBook flag is disabled, it is excluded from the API response entirely. The feature evaluates to null, targeting rules are ignored, and the SDK renders the fallback value — the safe, deterministic default you defined when you created the flag. No deploy required.

This is the kill switch mechanism in practice, and it is the reason kill switches are a near-universal requirement for AI feature management rather than a premium capability.

The fallback value matters as much as the kill switch itself. If your AI feature flag is off and your application has no defined default, users see a broken experience. If the fallback is a deterministic, non-AI version of the same functionality, users see a degraded but functional one. Defining that fallback before launch is not optional — it is the difference between a contained incident and a visible outage.

Prerequisite targeting for AI variants

GrowthBook's prerequisite flags capability allows you to gate one feature on the state of another. In practice, this means an advanced AI variant — say, a multi-turn reasoning interface — can be restricted to users who already have the base AI feature enabled and have demonstrated stable engagement with it.

You prevent exposing users to a complex model behavior before they've had a stable experience with the simpler version — one of the more common sources of UX breakage in AI rollouts.

Separating deployment from exposure

The principle underlying all of this is straightforward: code can be deployed to production and validated before any user sees it. This separating deployment from exposure is what makes gradual rollouts, guardrail metrics, and fallback behaviors operationally viable. Without it, the team is choosing between shipping to everyone and not shipping at all.

Feature flags create the middle ground — and in GrowthBook, any flag can be converted into a structured A/B test with one click, bridging the gap between safe deployment and rigorous experimentation without requiring a separate instrumentation pass.

Roll out AI changes incrementally to catch quality regressions early

The instinct to ship an AI feature to all users at once is understandable — you've tested it internally, the eval scores look good, and the team is confident. But internal testing doesn't surface the long tail of real-world inputs that will stress your model in production, and by the time a quality regression shows up in aggregate metrics, it may have already damaged trust with thousands of users.

Incremental rollouts are the practical answer: limit initial exposure, monitor what actually happens, and only widen access when the data supports it.

This isn't just operational caution. When done with the right tooling, a staged rollout generates statistical signal at each stage — you're not just hoping nothing breaks, you're running a controlled comparison between the new AI behavior and the baseline.

The staged rollout pattern for AI features

The mechanics are straightforward in principle: expose a small percentage of users to the new AI variant, monitor guardrail metrics, and advance to the next tier only when the data is clean. GrowthBook's Safe Rollout feature formalizes this into a documented ramp-up schedule — 10% → 25% → 50% → 75% → 100% — where the entire ramp completes within the first 25% of the configured monitoring duration.

If you set a four-day monitoring window, traffic reaches 100% by the end of day one; the remaining three days monitor the fully-deployed feature for guardrail metric failures before the rollout is marked complete.

Under the hood, Safe Rollout runs as a short-term A/B test: the control group receives the existing feature value, the rollout group receives the new AI behavior. This means you're not just watching error dashboards — you're getting statistically grounded comparisons between the two experiences at every stage of the ramp.

For teams that want gradual exposure without automated monitoring, a simpler Percentage Rollout rule (available on all plans) supports manual ramp-ups like 10% → 50% → 100%. The tradeoff is that go/no-go decisions become entirely manual, which requires pre-defined criteria and the discipline to actually enforce them.

Defining go/no-go criteria at each stage

The most common question from teams new to staged rollouts is: how do I know when I'm ready to move from 10% to 25%? For teams using automated guardrail monitoring, the answer is built into the tooling.

GrowthBook's Safe Rollout analyzes guardrail metrics using one-sided sequential testing, which means it can flag a regression as soon as statistical significance is reached — without inflating false positive rates from repeated looks at the data. The status dashboard surfaces clear signals: "Ready to ship" means no regressions detected and the monitoring duration has completed; "Guardrails Failing" means a regression has been detected and the rollout should be reconsidered.

For teams on manual rollout rules, the go/no-go decision requires explicit pre-commitment: define which metrics constitute a regression, set the threshold before the rollout begins, and treat a breach as a hard stop rather than a discussion point. The discipline to hold that line — especially under product pressure — is what separates a staged rollout from a staged rollout in name only.

Instant deactivation as a non-negotiable for AI

AI regressions don't always announce themselves immediately. A model that produces subtly lower-quality responses may not spike your error rate — it will surface gradually in session depth, task completion, or return behavior. By the time the regression is obvious in aggregate, the damage is already distributed across a significant portion of your user base.

This is why the ability to instantly deactivate a rollout without a code deploy is a hard requirement for AI features, not a nice-to-have. GrowthBook's Auto Rollback toggle makes this automated: if any guardrail metric fails at statistical significance during monitoring, the rollout rule is disabled automatically. Teams that prefer to retain manual control can leave Auto Rollback off and act on the "Guardrails Failing" status themselves.

Using engagement data to drive rollout decisions

Error rates and latency are necessary signals, but they're not sufficient for AI features. A model can respond quickly and without errors while still producing outputs that users find unhelpful, confusing, or off-topic — and that kind of quality regression will show up in engagement signals before it shows up in infrastructure metrics.

Session depth, task completion rates, and return visit frequency in the days following initial exposure are the leading indicators worth watching as you advance through rollout stages.

GrowthBook's Safe Rollout monitoring surfaces metric boundary trends as a time series — not just a single snapshot — which means you can see whether a guardrail metric is drifting toward failure or stabilizing as the rollout progresses. That trajectory matters as much as the current value, particularly for AI features where output quality can shift as the model encounters a wider distribution of real-world inputs.

Define guardrail metrics and fallback behaviors before you launch any AI experiment

Incremental rollouts tell you when to slow down. Guardrail metrics tell you why. Without the second piece, a staged rollout is just a slower way to ship a feature you don't fully understand.

A team ships a new AI-powered recommendation feature. Primary engagement metrics climb past their goal within the first week. Everyone celebrates. Three months later, someone notices infrastructure costs have tripled — a direct consequence of the feature's unintended behavior at scale. This scenario, as Statsig's engineering team has documented, is not unusual. It's what happens when teams optimize for the metric they're watching and ignore the ones they're not.

For AI features, the stakes compound. A UI change that performs poorly on engagement metrics wastes engineering time. An AI feature that performs well on engagement metrics while silently degrading error rates, inflating latency, or generating harmful outputs can erode user trust before any monitoring system catches it.

The probabilistic nature of AI — where the same input can produce different outputs across sessions, user segments, and time — means damage can be inconsistent and hard to attribute without systematic monitoring in place.

Guardrail metrics are the answer to this problem. They are not the metrics you're trying to improve. They are the metrics you commit to not harming, regardless of what your primary goal metric does.

What guardrail metrics are and why they're non-negotiable for AI

The clearest way to think about guardrail metrics is as a safety net that runs in parallel to your primary success metric. You might be optimizing for task completion rate on an AI assistant feature. Your guardrail metrics are the signals that would tell you the feature is winning in a way that's unsustainable or harmful — error rates spiking, response latency degrading, downstream conversion dropping.

For AI specifically, the guardrail set should include metrics that reflect model health, not just business outcomes. Error rates and API failure rates capture model reliability. Latency tracks inference cost and user experience degradation simultaneously. Conversion rates downstream of the AI interaction capture whether the feature is actually serving users or just capturing their attention. GrowthBook's documentation on Safe Rollouts names exactly these three — error rates, latency, and conversions — as the canonical starting point.

One critical constraint: keep the guardrail set focused. Adding too many guardrail metrics increases false positive rates, which means you'll trigger rollbacks on noise rather than signal. Define the smallest set of metrics that would tell you the feature is genuinely causing harm.

Defining fallback behaviors in code before launch

A fallback is not an error message. It is a pre-engineered alternative experience that activates when the AI component fails, regresses, or gets rolled back — and it must exist in code before the experiment launches, not after a guardrail fires.

testRigor's framing of AI unpredictability is worth sitting with here: in traditional software, 2 + 2 always equals 4. In AI systems, it usually equals 4, but occasionally equals "purple," and every once in a while the model replies that it's not authorized to discuss mathematics. That unpredictability is not a bug you can patch — it's a property of the system you have to design around. The fallback is how you design around it.

The practical implication: your fallback behavior should be the control experience — a known-good, deterministic baseline that users can always be returned to. GrowthBook's Safe Rollout architecture enforces this structurally. The control receives the existing value; the rollout receives the new value.

When Auto Rollback is enabled and a guardrail metric fails, GrowthBook automatically disables the rollout rule and returns all users to the control. The failure threshold is set at zero — as soon as there is statistical certainty that any guardrail metric is being harmed, even by a small amount, the rollout is flagged for rollback. That design choice reflects a deliberate safety-first stance: the cost of a false positive rollback is lower than the cost of a harmful AI output reaching your full user base.

The ethics of short-term metric optimization

There is a practical ethical dimension to guardrail metrics that teams often miss until it's too late. An AI feature can appear to win on aggregate engagement metrics while producing harmful or degraded outputs for specific user subgroups. Probabilistic AI outputs don't fail uniformly — they can fail disproportionately for users with certain query patterns, languages, or contexts. Aggregate metrics won't surface this. Only guardrail metrics designed to catch distributional harm will.

This is not an abstract concern. It's the concrete reason why optimizing for short-term engagement in an AI experiment is a different kind of risk than optimizing for short-term engagement in a button color test. AI models must constantly adapt to evolving policies and ethical guidelines, and experimentation without guardrails is not responsible shipping — it's just fast shipping.

The practical takeaway is straightforward: before any AI experiment launches, define your guardrail metrics, configure automated monitoring against them, write your fallback behavior into the codebase, and ensure your control variant is a stable, deterministic baseline. If a guardrail fires, you want the system to act — not wait for someone to notice.

AI experiments produce the wrong answers when you ask the wrong questions

Getting the rollout mechanics right — flags, gradual exposure, guardrails — only solves half the problem. Once users are actually seeing your AI variant, you still need to know whether it's working. And for AI features, "working" is harder to define than most teams expect.

Why standard metrics fall short

Click-through rate tells you whether a user acted. It doesn't tell you whether the AI response that prompted that click was accurate, helpful, or quietly eroding their trust in your product. Conversion rate has the same blind spot. These metrics measure downstream behavior, but AI quality problems often don't show up in downstream behavior immediately — they accumulate.

A user who gets a misleading AI-generated answer might still complete their session. They just don't come back next week.

There's also the averaging problem. Experiment results are aggregate statistics, and aggregates can hide harm to specific user subsets. An AI feature that performs well on average might be producing poor outputs for users in a particular context, language, or use case — and your top-line metrics won't surface that. GrowthBook's own experimentation documentation flags this directly: standard metrics "might not capture harm that is being done to some subsets of your population."

Metrics that actually capture AI feature quality

The metrics worth tracking for AI experiments are ones that connect model behavior to user outcomes, not just user actions. Task completion rate — did the user accomplish what they came to do? — is more informative than whether they clicked something. Session depth, measured as meaningful engagement rather than raw page views, can signal whether an AI response moved the conversation forward or dead-ended it.

User correction rate is a direct signal of output quality that click-based metrics completely miss. When users frequently rephrase, retry, or override an AI suggestion, that behavior tells you the model's first response wasn't good enough — a signal that aggregate engagement numbers will never surface.

Response relevance, where you can instrument it, captures whether the model's output was appropriate to the query at all.

These aren't exotic metrics. They're the kind of signals that require you to define what "success" means for your specific AI feature before you run the experiment — which is exactly the discipline that AI experimentation demands.

Matching metric types to what you're measuring

Not every AI signal fits a standard "did this event happen?" measurement. Depending on what you're trying to capture, you need different metric structures — and choosing the wrong one will give you a misleading result.

Latency is the clearest example. If you measure AI response time as an average, a handful of fast responses will pull the number down and hide the fact that 5% of your users are waiting four seconds for every reply. A percentile metric — specifically p95 or p99 — shows you what the worst-case user is actually experiencing, which is where trust breaks.

For task completion per session, a ratio metric normalizes for session length, so a user who completed three tasks in a long session and a user who completed one task in a short session aren't treated identically. For AI signals that don't map to any standard event type, SQL-defined metrics let you write the measurement logic directly against your data warehouse — which is the practical escape hatch for novel AI features where the output doesn't fit a button click or a page view.

When you realize mid-experiment you measured the wrong thing

It happens. You launch an AI experiment tracking one metric, and two weeks in you realize the signal you actually needed was something you didn't instrument. On most experimentation platforms, that's a sunk experiment. You either live with the wrong data or restart.

GrowthBook supports retroactive metric addition — you can add metrics to a past experiment and compute results against historical exposure data without re-running the test. Merritt Aho, Digital Analytics Lead at Breeze Airways, put it plainly: "Being able to spin up new metrics mid-experiment is a game changer. This was simply never possible before." For AI experiments, where unexpected behaviors are the norm rather than the exception, this capability is a meaningful safety net.

Tracking cumulative impact across many AI model variants

AI teams don't run one experiment. They run a continuous stream of model comparisons — different prompts, different retrieval strategies, different fine-tuning approaches — and the learning compounds over time. Measuring each experiment in isolation misses the larger picture.

This is exactly the use case Character.AI describes. Landon Smith, Head of Post-Training at Character.AI, explains: "GrowthBook has been an invaluable tool for Character.AI, helping us develop our models into a great consumer experience. We can compare different modeling techniques from the perspective of our users — guiding our research in the direction that best serves our product."

The key phrase is "from the perspective of our users." Model-side metrics — loss, perplexity, benchmark scores — don't tell you whether users are better served. An experimentation layer that surfaces user-outcome data across many model variants does.

Cumulative impact dashboards, win rate tracking across experiments, and a longitudinal learning library let teams find what worked and what didn't over time. For AI teams running frequent model iterations, that longitudinal view is often more valuable than any single experiment result.

The four pieces only work as a stack — pull one out and the others lose their teeth

The four pieces covered in this article — feature flags, incremental rollouts, guardrail metrics, and measurement — aren't independent best practices you can adopt in any order. They're a stack. Flags make gradual rollouts possible. Gradual rollouts make guardrail monitoring meaningful. Guardrail monitoring is only useful if your metrics are actually connected to model quality, not just downstream clicks. Pull one piece out and the others lose their teeth.

Consider the failure mode when guardrail metrics are in place but the fallback behavior was never defined. A Safe Rollout fires Auto Rollback because error rate crosses the threshold. The flag disables. The SDK returns null. Your application has no fallback value configured, so users see a broken interface — not a degraded one, a broken one.

The rollout system worked exactly as designed. The outcome was still an incident. That's what "pull one piece out" looks like in practice: each remaining piece performs correctly and the system still fails.

The minimum viable guardrail stack: flags, fallbacks, and metrics you need before launch

Before any AI experiment goes live, three things need to exist in your system: a feature flag with a defined fallback value, at least one guardrail metric that reflects model health rather than just business outcome, and a control variant that is a stable, deterministic baseline. That's the floor. Everything else — staged ramp schedules, automated rollback, retroactive metric addition — builds on top of it.

Sequencing matters more than speed: what to do before you open traffic

The sequencing matters more than the speed. Write your fallback behavior before you write your experiment config. Define your guardrail thresholds before you open traffic. Start with internal users, then a small cohort, then a percentage of production — and treat the go/no-go decision at each stage as a real decision, not a formality.

The teams that skip these steps aren't moving faster; they're just deferring the cost to a moment when it's harder to absorb.

What AI experimentation requires that general-purpose analytics platforms weren't built for

The tooling question matters because AI experimentation has specific requirements that general-purpose analytics platforms weren't built for: runtime flag evaluation without latency overhead, sequential testing that lets you look at guardrail metrics without inflating false positive rates, retroactive metric addition for when your instrumentation doesn't match what the model actually surfaced, and longitudinal tracking across many model variants.

GrowthBook was built with these requirements in mind, which is why teams like Character.AI use it specifically to compare modeling techniques from the user's perspective — not just from benchmark scores.

One honest tension worth naming: the system described here takes real upfront investment. Defining fallbacks, instrumenting the right metrics, and building the discipline to hold go/no-go criteria under product pressure all require work before you see any benefit. The temptation is to skip the setup and ship.

That tradeoff is real — but so is the asymmetry this article opened with. Trust erosion from a bad AI experiment doesn't recover on the same timeline as a bad button color test.

If you're early in this process, the goal isn't to build the full system on your first experiment. It's to build enough of it that you can pull back fast if something goes wrong. That's a much more achievable starting point than it sounds.

What to do next

If you don't yet have feature flags on your AI features, that's where to start — not because it's the most interesting piece, but because nothing else in this system works without it. Create a flag for your next AI feature, define the fallback value before you enable it for anyone, and make sure your team knows how to disable it without a deploy.

Once that's in place, add one guardrail metric — error rate or latency, not engagement — and configure monitoring against it before the first user sees the variant. That's a complete, functional first version of this system. Build from there.

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.

Shipping an AI feature to 100% of your users on day one is the fastest way to find out your model behaves very differently in production than it did in your evals.

Why AI features demand a different approach to experimentation

The non-determinism problem

Trust erosion is asymmetric and sticky

Standard metrics can mask downstream harm

The compounding cost of a poorly planned AI experiment

Feature flags are the control layer that makes everything else in AI rollouts possible

A hardcoded branch can't close a four-minute incident window. A feature flag can.

Controlled rollout mechanics and JSON flag payloads

Kill switches and instant rollback

Prerequisite targeting for AI variants

Separating deployment from exposure

Roll out AI changes incrementally to catch quality regressions early

The staged rollout pattern for AI features

Defining go/no-go criteria at each stage

Instant deactivation as a non-negotiable for AI

Using engagement data to drive rollout decisions

Define guardrail metrics and fallback behaviors before you launch any AI experiment

What guardrail metrics are and why they're non-negotiable for AI

Defining fallback behaviors in code before launch

The ethics of short-term metric optimization

AI experiments produce the wrong answers when you ask the wrong questions

Why standard metrics fall short

Metrics that actually capture AI feature quality

Matching metric types to what you're measuring

When you realize mid-experiment you measured the wrong thing

Tracking cumulative impact across many AI model variants

The four pieces only work as a stack — pull one out and the others lose their teeth

The minimum viable guardrail stack: flags, fallbacks, and metrics you need before launch

Sequencing matters more than speed: what to do before you open traffic

What AI experimentation requires that general-purpose analytics platforms weren't built for

What to do next

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