Experiments

Data Science

How to A/B test AI-powered search or recommendations

May 27, 2026

min read

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

A/B testing an AI search or recommendation system with the same setup you'd use for a button color test will produce results you can't trust — and nothing in your experiment pipeline will warn you.

The model learns from user behavior during the experiment, which means control-group clicks can quietly shape the treatment model's training, and vice versa. By the time you call a winner, the two groups may no longer be independent.

This guide is for engineers, PMs, and data teams who are building or iterating on AI-powered search or recommendation features and need a reliable way to measure what's actually working. Here's what it covers:

Why AI search experiments break the assumptions standard A/B testing is built on
Which metrics to track — relevance, click-through rate, conversion, session success, and deflection — and why each one catches something the others miss
How to structure the experiment correctly: user bucketing, experiment ID tagging, and group isolation
The statistical pitfalls specific to AI search tests, including novelty effects, sample size requirements, and false positives from multiple metrics
What to do after the first test, including how to iterate on ranking configs and personalization parameters over time

The article moves in that order — from why this is different, to what to measure, to how to set it up, to how to keep improving. Each section is self-contained, so if you already have a test running and just need the metric guidance or the statistical safeguards, you can jump straight there.

Why A/B testing AI-powered search and recommendations is fundamentally different

Traditional A/B testing rests on a deceptively simple assumption: the control and treatment are static, independent, and do not influence each other. Show half your users a blue button and half a green one, measure clicks, declare a winner. The variants don't change based on who interacts with them. That assumption breaks completely when the thing you're testing is an AI model.

AI-powered search and recommendation systems learn from user behavior in real time. The model you're testing today is not the same model you'll be measuring next week — it has been shaped by the behavioral signals your experiment generated.

This creates a category of failure mode that doesn't exist in conventional experimentation: your test can appear to be running correctly while producing results that are structurally uninterpretable.

The contamination risk: how feedback loops make missing experiment IDs a silent failure

When a user submits a query, clicks a result, or abandons a session, that event becomes training signal. In a standard web experiment, a button click is just a data point. In an AI search experiment, that same click can influence how the model ranks results for the next user.

Google's documentation for AI Commerce Search makes this explicit: when running an A/B experiment, you must include which group a user belongs to when recording user events, because that information is used to refine the model and provide metrics. Both the control group and the experimental group must log events — not just the treatment arm. This is a direct acknowledgment that the model is actively learning during your experiment, and that behavioral data from both groups is feeding back into model state.

This is the core structural difference from a button-color test. The button does not update itself based on who clicks it. The AI model does.

If experiment IDs are not attached to every user event log, the model cannot distinguish which behavioral signals came from which group. Control-group clicks can influence the treatment model's training. Treatment-group behavior can bleed into the baseline. The two groups are no longer independent — but nothing in your experiment infrastructure will tell you that.

This failure mode is silent. There is no sample ratio mismatch alert, no error in your event pipeline, no anomaly in your dashboards. The experiment continues to run, metrics continue to accumulate, and at the end you have a result that looks like a result but is not interpretable as one.

Google's solution is precise: every user event must include the experiment ID in a dedicated field, traffic splitting must be handled by a third-party platform (not the AI system itself), and the two groups must be isolated at the model-training level, not just at the UI layer. The requirement exists because the risk is real and the failure mode is invisible.

The moving target problem

Even setting aside contamination, AI search results are probabilistic and personalized in a way that static variants are not. The same query from the same user can return different results depending on session context, prior behavior, and current model state. "The treatment" in an AI search experiment is not a fixed thing you can point to — it is a distribution of outputs that shifts over the duration of the experiment.

This matters for measurement. Tools designed to compare static variants assume the thing being measured stays constant between exposure and outcome. In practice, this means that if your AI search model updates its ranking weights on Tuesday and your experiment doesn't end until Friday, the "treatment" your Thursday users experienced is measurably different from the one your Monday users saw — but your metrics treat them as the same. That inconsistency introduces error you cannot correct retroactively by reprocessing the data.

The infrastructure requirements that have no equivalent in standard experimentation

The design requirements that follow from these dynamics have no equivalent in standard experimentation. Every event from both groups must be tagged with an experiment ID. Traffic assignment belongs outside the AI system entirely. And group isolation has to be enforced at the model-training layer, not just the serving layer.

Because AI search experiments typically require multiple metrics — relevance, engagement, conversion, session success — the statistical risk compounds: running five unrelated metrics at standard significance thresholds produces a 41% chance of at least one false positive. None of this is optional. A contaminated AI search experiment doesn't just produce a wrong answer — it can actively degrade the model for users in both groups without producing any visible signal that something went wrong. That's a higher-stakes failure mode than a misconfigured button test, and it requires a correspondingly more deliberate approach to every layer of the experiment.

Which metrics actually matter when testing AI search: relevance, click-through, conversion, and beyond

Metric selection is the single most consequential decision in an AI search or recommendations experiment. Get it wrong, and you'll ship regressions you can't see — optimizing a number that looks healthy while the actual user experience quietly deteriorates. This isn't a hypothetical risk. Practitioners discussing LLM A/B testing have noted exactly this failure mode: "Any negative result can now be thrown straight into the trash because of the chance [the platform] put you on the wrong side of an A/B test." That's what happens when your metric coverage has gaps.

The GrowthBook experimentation docs put it plainly: "The simplest way this can go wrong is if your metrics are tracking the wrong things, in which case you'll have garbage in and garbage out." AI search experiments need a layered metric stack — not because more metrics are always better, but because each layer captures something the others structurally cannot.

Relevance: the upstream signal everything else depends on

Relevance is the foundation metric, and it's the one teams most often skip because it's the hardest to instrument. In AI search, relevance has two components: finding matching records, and ranking them so the best results appear first. But it extends further — into personalization signals, popularity weighting, and merchandising rules that shape what "best" means for a given user in a given context.

The reason relevance matters upstream is that if your ranking is wrong, every downstream metric is measuring the wrong thing. A model that surfaces irrelevant results with compelling titles will generate clicks. Those clicks will not convert. And by the time you see the conversion signal, you've already run the experiment and shipped the change.

Relevance is qualitative by nature, which is why teams default to proxies. The right approach is to instrument it directly — through human relevance judgments, offline evaluation sets, or model-level signals — and treat it as a leading indicator that validates what your engagement metrics are actually measuring.

Engagement metrics: what CTR and dwell time capture (and miss)

Click-through rate is the most commonly used proxy for search quality, and it has real value — but it has a specific failure mode that matters in AI search. A result set with sensational but irrelevant titles can drive high CTR with low conversion. Goodhart's Law applies directly here: when CTR becomes the target, it ceases to be a good measure of search quality.

Dwell time adds a dimension CTR misses. If a user clicks a result and immediately returns to the search page, that's a signal of failure that CTR records as a success. The GrowthBook docs flag this pattern explicitly, warning that relying on short-term engagement 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." In AI search, that dynamic is especially dangerous because the model can learn from those clicks and reinforce the pattern.

Use CTR and dwell time as engagement signals, not as success criteria. They belong in your metric stack, but not at the top of it.

Conversion and revenue: the ground-truth validation layer

Conversion rate and revenue are the metrics that validate whether relevance and engagement improvements are real. Frameworks for evaluating search quality consistently name these as the most business-meaningful signals, and GrowthBook's approach to AI experimentation is built around linking model performance directly to user outcomes rather than stopping at surface engagement.

Because conversion is high-stakes, it warrants stricter statistical controls than softer engagement metrics. GrowthBook supports per-metric significance thresholds — applying tighter significance requirements to revenue metrics than to CTR — which is the right configuration for an AI search experiment where a false positive on conversion is far more costly than a false positive on click rate.

Session success and deflection: the metrics that catch what others miss

Session-level metrics capture whether the user actually accomplished their goal — something CTR and conversion can both miss in multi-step journeys. Query reformulation rate is the clearest signal: if a user submits a second query immediately after the first, the first result set failed. Pogo-sticking back to search after clicking a result is the same signal at the click level.

For support and help-center search, zero-result rate is the most direct indicator of model failure. A user who gets zero results doesn't click, doesn't convert, and often doesn't come back. This metric belongs configured as a guardrail — a threshold that flags or halts an experiment even when primary metrics look positive. GrowthBook's guardrail metric system is designed for exactly this use case: ensuring that a win on engagement doesn't mask a simultaneous degradation in core experience quality.

The concern about AI systems degrading UX without detection is precisely the problem these session-level signals are built to catch. Instrument them from the start, not as an afterthought when something looks wrong.

One practical advantage worth noting: warehouse-native experimentation architectures allow teams to define new metrics retroactively against existing experiment data. As Merritt Aho, Digital Analytics Lead at Breeze Airways, put it: "Being able to spin up new metrics mid-experiment is a game changer. This was simply never possible before." In AI search experiments — where you often discover mid-run that you're missing a critical signal — that flexibility matters more than it does in a standard UI test.

Three non-negotiable infrastructure requirements for AI search experiments

Clean experiment infrastructure is what separates interpretable results from noise in AI search testing — and the consequences of getting it wrong are worse than in a standard feature experiment. When a button-color test is misconfigured, you get unreliable data. When an AI search experiment is misconfigured, you corrupt the behavioral signals the model uses to train itself, and the damage compounds silently over the experiment's lifetime.

There are three non-negotiable requirements: consistent user bucketing, experiment ID tagging on every logged event, and strict group isolation. Each one is load-bearing.

Why inconsistent user assignment corrupts AI model training, not just metrics

Every user must be deterministically assigned to the same group — control or treatment — across every session and every surface where the search experience appears. This sounds obvious, but AI search creates a specific failure mode that simpler experiments don't face: if a user sees AI search results on one session and legacy results on another, their behavioral signals (clicks, dwell time, purchases) get attributed to the wrong model. The AI system learns from contaminated data, and your metric comparison becomes meaningless.

The choice of randomization unit matters and must be made deliberately. For authenticated products, user ID is the right anchor. For anonymous or session-heavy experiences, session ID or device ID may be more appropriate. The key is that the assignment logic is deterministic — given the same user identifier and experiment configuration, it always produces the same group assignment. GrowthBook supports flexible randomization units (user, location, postal code, URL path) precisely because different product architectures require different bucketing strategies.

Experiment ID tagging in every user event log

This is the most technically specific requirement in the section, and it comes directly from how AI search systems use behavioral data. Google Cloud's AI Commerce Search documentation states explicitly that when recording a user event, teams must specify which group the user belongs to by including the experiment ID in the experimentIds field. This is a hard requirement, not a suggestion.

The reason is dual-purpose. First, the experiment ID enables metric comparison between groups — without it, you cannot attribute a click or conversion to a specific treatment. Second, it allows the AI model to refine itself using correctly attributed behavioral data. If events are logged without experiment IDs, the model cannot distinguish which behavioral signals came from which treatment. Without attribution, results are unrecoverable.

Group isolation must happen at the model-training layer, not the UI layer

Both groups must receive identical application experiences except for the single variable under test — in this case, which search system generates the results. Any other difference, whether a UI change, a pricing update, or a promotional banner that only one group sees, invalidates the comparison. Google Cloud's documentation is explicit on this point: both versions of the application must be the same, except that users in the experimental group see results generated by AI Commerce Search.

One practical implication that surprises some teams: Google Cloud's AI Commerce Search does not handle traffic splitting internally. The documentation specifically calls out third-party experiment platforms — naming VWO and AB Tasty as examples — as the mechanism for group assignment. The third-party platform assigns unique experiment IDs to each group, which then flow into the event logs. Any experiment platform that can handle server-side assignment and pass experiment IDs into your event stream can fill this role.

Why keeping experiment data in your own warehouse makes the event tagging requirement tractable

The event logging requirement — tagging every user event with an experiment ID — only works if those events flow somewhere queryable. GrowthBook's experimentation platform analyzes data directly in your existing data warehouse — Snowflake, BigQuery, Redshift, or Postgres — rather than in a proprietary system. This matters for three practical reasons.

First, all tagged events can be joined with other product data and queried with SQL, making metric definitions transparent and reproducible. Second, metrics can be created or modified retroactively without re-running the experiment — a capability that matters specifically for AI search tests, where teams often discover mid-experiment that they should have been tracking query reformulation rate or zero-result rate from the start. Third, keeping experiment data in your own infrastructure means no end-user PII needs to leave your systems — a meaningful consideration for teams logging detailed search behavior at scale.

The warehouse-native approach is what makes the experiment ID tagging requirement tractable operationally. Tag every event, store it where you already keep your data, and define metrics against it in SQL. That's the implementation loop.

Statistical pitfalls in AI search experiments: novelty effects, sample size, and false positives

AI search experiments inherit all the statistical hazards of standard A/B tests and then add several of their own. Teams that borrow their statistical practices from UI experiments — where the treatment is a static change to a button or headline — will find those practices break down when the treatment is a probabilistic, adaptive system with high output variance and multiple correlated metrics. The failure mode is not always obvious: you can ship a false positive on an AI search test and spend months optimizing a model that was never actually better.

Novelty effects and why your first two weeks of data lie

When users encounter a new search interface or recommendation system, their behavior changes simply because it's new. They explore more, click more, and engage differently than they will after the experience becomes familiar. This inflates early engagement metrics — CTR especially — and can push a mediocre AI model over a significance threshold it doesn't deserve to cross.

The standard mitigations are straightforward but frequently skipped: run the experiment long enough for novelty effects to decay (a minimum of two to four weeks is a common baseline, though the right duration depends on your traffic volume and how frequently your users search), or restrict the experiment to new users who have no prior behavior to compare against. AI search compounds this problem because the model itself may be warming up on new behavioral data during the early experiment window, creating a double novelty effect — the user is adjusting to the interface while the model is adjusting to the user.

Sample size, variance, and the low-traffic reality

Conversion and session success are the metrics that matter most for AI search, and they are also the hardest to measure precisely. They are low-frequency, high-variance events. A power calculation sized for a button-color test will dramatically underestimate what you need for a search relevance experiment.

This is not a theoretical concern. The practitioner community is direct about it: most teams simply don't have enough traffic to run statistically valid tests on low-frequency outcomes, and running an underpowered test is not a neutral act — it produces noise you might mistake for signal. GrowthBook's Minimum Data Thresholds guardrail addresses this at the platform level by flagging conclusions drawn from implausibly small samples (the docs give the example of 5 vs. 2 conversions). But the more powerful technique is CUPED — Controlled-experiment Using Pre-Experiment Data — which reduces metric variance by controlling for pre-experiment user behavior. CUPED effectively increases statistical power without requiring more users, which matters enormously when your conversion events are sparse.

Multiple comparisons and the false positive stack

AI search tests don't evaluate a single metric. A properly instrumented experiment tracks relevance signals, CTR, conversion, session success, and deflection rates simultaneously — five or more metrics evaluated at the same significance threshold. The problem is that false positive probability compounds across metrics. GrowthBook's documentation states it directly: if you test the same hypothesis at a 5% significance level across 20 different metrics, the probability of finding at least one statistically significant result by chance alone is approximately 64%. For a more realistic AI search stack of five metrics, the math still produces a false positive risk that should make any team uncomfortable — and that's before accounting for the correlation between metrics, which tends to make things worse rather than better.

Multiple-comparison corrections work by raising the significance bar for individual metrics when you're evaluating several at once — the more metrics you test, the higher each individual metric's bar needs to be to keep your overall false positive rate under control. The standard approaches are Bonferroni correction (conservative, raises the bar uniformly across all metrics), False Discovery Rate correction, and the Benjamini-Hochberg procedure (more permissive, controls the expected proportion of false positives rather than eliminating them entirely). GrowthBook ships both Benjamini-Hochberg and Bonferroni as built-in corrections.

The practical implication is that you should decide which metrics are primary before the experiment runs, apply corrections to your full metric stack, and resist the temptation to go hunting through segments and time windows after the fact — that path leads to the Texas Sharpshooter problem, where you find significance by looking in enough places.

Sequential testing and SRM detection

Two additional tools deserve specific attention for AI search experiments. Sequential testing allows valid inference at any point during an experiment without inflating false positive rates — it addresses the peeking problem, which is acute for AI search tests where teams feel pressure to call results early when a new model appears to be performing well. Standard fixed-horizon tests are not designed for mid-experiment reads; sequential testing is.

Sample Ratio Mismatch detection catches a different class of failure: traffic splits that don't match what you configured. A 48/52 split when you expected 50/50 is a signal that something in your bucketing or event logging is broken, and results from a test with an SRM are not trustworthy regardless of what the metrics show. In AI search experiments, SRM can occur if experiment ID tagging is inconsistent across event types or if the model itself influences which user interactions get logged. GrowthBook runs SRM detection automatically as a data quality check, alongside suspicious uplift detection — which flags metric changes that are implausibly large and likely indicate an instrumentation bug rather than a real effect. For AI search, where a model error can produce dramatic metric swings, that guardrail is worth having.

The broader point is that statistical rigor in AI search testing is not a single decision about significance thresholds. It's a layered system of countermeasures — novelty mitigation, variance reduction, multiple-comparison corrections, and traffic integrity checks — each targeting a distinct failure mode. Skipping any one of them doesn't just weaken your results; it can produce confident, wrong conclusions about which model to ship.

After the baseline wins: shifting from AI vs. legacy to config-vs-config iteration

"When it comes to AI, evals are just the tip of the iceberg. A/B testing is where real value gets created." That framing captures something important about where most teams go wrong: they treat their first AI search experiment as a decision, when it's actually the foundation for a program.

Once your AI-powered search or recommendation system has beaten the legacy baseline and been fully adopted, the experiment that got you there has served its purpose. The infrastructure you built — user bucketing, experiment ID tagging, a layered metric stack — is now the engine for something more valuable: continuous iteration on the system itself.

From baseline to config-vs-config

The shift in framing matters. You're no longer asking "does AI search beat keyword search?" You're asking "which version of AI search wins?" That's a fundamentally different experimental posture, and it opens up a much richer hypothesis space.

In practice, config-vs-config testing means running controlled experiments on the parameters that govern how your AI search system behaves: retrieval strategy weights, reranking model choices, embedding model versions, boost specifications for certain product categories, personalization signal inclusion or exclusion. Each of these is a testable hypothesis with measurable downstream effects on the metrics you've already instrumented. The experiment infrastructure doesn't change — you're still bucketing users consistently, tagging events with experiment IDs, and measuring against the same metric stack. What changes is that both variants are AI search configurations rather than AI versus legacy.

Landon Smith, Head of Post-Training at Character.AI, describes exactly this kind of iteration in practice: "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." That's the model: user-facing metrics, not internal evals, as the arbiter of which configuration wins.

One practical caution here, drawn from practitioner experience: config changes tested during iteration can silently harm users if you're not watching guardrail metrics closely. The same rigor that protected your initial experiment — pre-registered guardrails, kill switches, SRM detection — applies equally to follow-on tests. Iteration done carelessly is how teams ship regressions they can't explain.

Tracking cumulative impact across the program

Individual AI search experiments often produce modest effect sizes. A 1–2% improvement in click-through rate or session success rate is meaningful, but it doesn't look impressive in isolation. The compounding value of an iteration program only becomes visible when you measure across experiments, not within them.

This is the argument for tracking cumulative impact as a first-class metric for your experimentation program. Experiment analytics platforms built around exactly this pattern offer a dashboard that shows the cumulative impact of all experiments on any given metric, with North Star metric performance tracked over time so teams can see how each individual experiment contributed to the trend. Individual experiments rarely move the needle in isolation — the compounding value only becomes visible when you measure across the program.

Win rate and frequency tracking adds another dimension. Teams should monitor whether they're running the right mix of bold bets — new retrieval architectures, embedding model swaps — versus incremental config tweaks. A program that only runs safe, small experiments will compound slowly. One that only swings for large changes will have high variance in outcomes and learn less per cycle.

Building a learning library to prevent redundant testing

As an AI search experiment program matures, institutional knowledge becomes both a competitive asset and a liability if it isn't captured. Teams that don't document what they've tested will re-run experiments already answered — wasting cycles, confusing stakeholders, and potentially reversing decisions that were made for good reasons no one remembers.

A learning library solves this. GrowthBook's implementation surfaces experiments that worked and those that didn't, making past results searchable and usable for informing future hypotheses. Some platforms go further, using vector embeddings to flag potentially redundant experiments before they're launched — catching "we already tested this" before it becomes a wasted sprint. AI-generated result summaries help teams quickly process and archive learnings from completed tests, which matters when iteration velocity is high and the volume of completed experiments starts to accumulate.

The broader point is that a learning library is what transforms a collection of individual tests into a compounding organizational capability. Without it, every new team member, every new quarter, every new model release starts from scratch. With it, your experiment history becomes infrastructure — the same way your event logs and metric definitions are infrastructure.

The teams that will extract the most value from AI-powered search aren't the ones who ran the best single experiment. They're the ones who built the program.

The infrastructure decisions that determine whether your AI search results are interpretable

The through-line of this article is that AI search experiments fail silently. The feedback loop problem, the contamination risk, the moving target — none of them produce an alert. They produce a result that looks like a result, and you ship something based on it. That's the actual risk, and it's why the infrastructure decisions covered here aren't optional overhead. They're what makes the output of your experiment mean something.

The non-negotiables: experiment design decisions you can't skip

The three things that will break an AI search experiment before it starts are inconsistent user bucketing, missing experiment IDs on event logs, and traffic splitting handled by the AI system itself. Get any one of these wrong and your groups are no longer independent — and you won't know. These aren't configuration preferences; they're the structural requirements that make everything else in the experiment interpretable.

Pre-registering your metric stack is what separates an interpretable result from a number you'll second-guess

The temptation is to start with CTR because it's easy to instrument and moves quickly. Resist it. Relevance is the upstream signal that validates whether your engagement metrics are measuring anything real. Conversion and session success are the ground-truth layer. Zero-result rate and query reformulation rate are the guardrails that catch regressions your primary metrics will miss.

Decide which metrics are primary before the experiment runs — not after you've seen the data — and apply multiple-comparison corrections to the full stack. Per-metric significance thresholds and built-in multiple-comparison corrections make this tractable without requiring a statistics degree.

The first experiment is an infrastructure investment, not a decision

The tension worth holding onto: your first experiment is not the goal. It's the infrastructure investment that makes every subsequent experiment cheaper and more reliable. The teams that compound value from AI search aren't the ones who ran one careful test — they're the ones who built the program, documented what they learned, and kept iterating on ranking configs and personalization parameters with the same rigor they applied to the baseline comparison.

The honest caveat is that this takes longer than most teams expect. Novelty effects alone require two to four weeks of runtime before your data stabilizes. Low-traffic products may need CUPED variance reduction just to reach adequate power on conversion metrics. That's not a reason to skip the work — it's a reason to size the timeline correctly from the start.

This article was written to give you a complete picture of what reliable AI search experimentation actually requires — not to make it sound simpler than it is, but to make it approachable enough that you can start with confidence.

What to do next: If you don't have a test running yet, your first action is to audit your event logging. Check whether every user event your search system generates can be tagged with an experiment ID and stored somewhere queryable. If it can't, that's your blocking dependency — fix it before you design the experiment. If your logging is already in place, define your metric stack before you start: a relevance signal as your upstream validator, CTR and dwell time as engagement proxies, conversion as your primary success criterion, and zero-result rate as a guardrail. The specific metrics matter less than the commitment to pre-registering them — deciding what counts as a win before you see the data is what makes the result interpretable.

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.

A/B testing an AI search or recommendation system with the same setup you'd use for a button color test will produce results you can't trust — and nothing in your experiment pipeline will warn you.

Why A/B testing AI-powered search and recommendations is fundamentally different

The contamination risk: how feedback loops make missing experiment IDs a silent failure

The moving target problem

The infrastructure requirements that have no equivalent in standard experimentation

Which metrics actually matter when testing AI search: relevance, click-through, conversion, and beyond

Relevance: the upstream signal everything else depends on

Engagement metrics: what CTR and dwell time capture (and miss)

Conversion and revenue: the ground-truth validation layer

Session success and deflection: the metrics that catch what others miss

Three non-negotiable infrastructure requirements for AI search experiments

Why inconsistent user assignment corrupts AI model training, not just metrics

Experiment ID tagging in every user event log

Group isolation must happen at the model-training layer, not the UI layer

Why keeping experiment data in your own warehouse makes the event tagging requirement tractable

Statistical pitfalls in AI search experiments: novelty effects, sample size, and false positives

Novelty effects and why your first two weeks of data lie

Sample size, variance, and the low-traffic reality

Multiple comparisons and the false positive stack

Sequential testing and SRM detection

After the baseline wins: shifting from AI vs. legacy to config-vs-config iteration

From baseline to config-vs-config

Tracking cumulative impact across the program

Building a learning library to prevent redundant testing

The infrastructure decisions that determine whether your AI search results are interpretable

The non-negotiables: experiment design decisions you can't skip

Pre-registering your metric stack is what separates an interpretable result from a number you'll second-guess

The first experiment is an infrastructure investment, not a decision

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

Report the right quantity for the decision

Build a clearer experiment readout