Experiments

Data Science

How much traffic do you need to test AI features reliably?

Jun 8, 2026

min read

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

Your standard sample size calculator will give you a confident number for an AI feature experiment — and that number will probably be wrong.

Not because the math is broken, but because AI features violate the core assumptions the math depends on. They're non-deterministic: the same user can see meaningfully different outputs across sessions, which adds a second layer of variance on top of normal user behavior differences. That extra noise makes your pre-experiment variance estimates too low and your effect size assumptions unreliable.

The result is an experiment that looks properly designed but is structurally underpowered.

This article is for engineers, PMs, and data teams who are shipping or planning to ship AI features — LLM-powered recommendations, AI writing tools, conversational interfaces — and need to know how much traffic to plan for and why the usual approach falls short. Here's what you'll learn:

Why AI feature non-determinism inflates metric variance and breaks standard power calculations
How to estimate variance and effect size when you don't have clean historical analogs
How to calculate the AI feature test traffic you actually need, with a worked example
How to allocate and segment traffic so your sample is valid, not just large enough
What to do when you genuinely don't have enough traffic to run a fully powered experiment

The article moves in that order — from the root cause, through the math, through the operational details, and finally to practical fallbacks for low-traffic situations. Each section builds on the last, but you can also jump to whichever problem you're facing right now.

Why AI features break standard sample size assumptions

When you're planning an experiment on a new checkout flow or a redesigned navigation element, the statistical setup is relatively straightforward. You pull historical conversion data to estimate variance, make a reasonable assumption about the lift you're trying to detect, and run the numbers through a power calculator.

The result is a sample size you can trust — because the feature you're testing behaves the same way for every user who sees it.

AI features don't work that way. And if you apply the same planning process to an LLM-powered recommendation engine or an AI-generated content feature, you'll end up with a sample size estimate that looks credible but is systematically too small.

The experiment will run, reach your calculated threshold, and return a result that may be structurally underpowered — without any obvious sign that something went wrong.

Non-determinism is the root cause

The defining characteristic of AI features is that they are non-deterministic: the same input can produce different outputs across sessions. A button color is identical for every user who sees it. An LLM-generated product recommendation shown to User A on Monday may be substantively different from what that same user sees on Wednesday — different phrasing, different items surfaced, different implicit framing.

The "treatment" in your experiment is not a fixed stimulus. It's a distribution of stimuli, varying across users and across time.

This isn't a bug or an implementation detail. It's the nature of how these systems work. Traditional software features "can be thoroughly tested with unit tests and integration tests". Once those tests pass, teams can trust the feature will continue working as expected.

AI features require quality thresholds instead — and as Stack Overflow discovered when it rolled back its conversational search feature for failing to reach a 70% accuracy threshold, a feature can pass every functional test while still failing to deliver consistent user-facing value.

How non-determinism inflates metric variance

In a standard A/B test, metric variance is driven primarily by differences in user behavior. For a conversion rate experiment, some users convert and some don't — that behavioral heterogeneity is the variance you're estimating from historical data.

When you add an AI feature, you introduce a second source of variance: output variability. Now your metric reflects user behavior differences plus the variability in what the AI actually showed each user plus the interaction between those two factors.

The practical consequence is that pre-experiment variance estimates based on historical metric data will be too low. Your historical engagement data on "recommendation clicks" doesn't capture the additional noise introduced by non-deterministic outputs.

Standard power calculators assume variance is a stable, knowable input — for AI features, it isn't. The calculator doesn't warn you. It just returns a number that's smaller than it should be.

Effect sizes are harder to predict

Sample size calculations also require an assumed minimum detectable effect. For a UI change, you can often anchor this estimate to analogous past experiments. For an AI feature, the effect size depends on model quality, prompt design, user segment, and context — none of which have clean historical analogs.

The relationship between a model's technical performance metrics and its actual user-facing impact is non-linear and difficult to anticipate in advance.

Effect sizes for AI features also tend to be highly heterogeneous across user segments. Power users and casual users often respond very differently to AI-generated suggestions, which means a single MDE assumption can mask a situation where the feature has a strong effect on one segment and a null or negative effect on another.

Averaging across that heterogeneity produces an effect size estimate that may not represent any real subpopulation accurately.

Why your sample size calculator will fail you

The failure mode here is subtle. Standard sample size calculators aren't broken — their math is correct. The problem is that the inputs they require (stable variance, a single predictable effect size, a deterministic treatment) are assumptions that AI features violate.

The calculator produces a confident, plausible-looking number, and nothing in the output signals that the inputs were wrong.

Teams run the experiment, hit the calculated sample size, and declare a result. What they've actually run is an underpowered experiment built on variance estimates that were too optimistic and effect size assumptions that may not reflect the feature's actual behavior in production.

GrowthBook's framing captures the failure pattern well: the industry has a recognized tendency to shipping AI features based on "vibe checks" rather than rigorous measurement — and applying a standard sample size formula to a non-standard problem is a more sophisticated version of the same mistake.

The rest of this article addresses how to fix the inputs, not just the formula.

High variance and uncertain effect sizes: the core AI feature test traffic problem

Every sample size calculation reduces to two inputs: variance and effect size. Get either one wrong, and your AI feature test traffic estimate is wrong — and with AI features, both are genuinely hard to pin down before the experiment runs.

This isn't a minor calibration issue you can correct with a slightly larger safety margin. It's a structural problem that makes standard pre-experiment planning unreliable.

Why AI outputs inflate metric variance

When the treatment itself is heterogeneous across users and sessions — as it is for any LLM-powered feature — the metric distribution widens, and the sample size required to detect a given effect grows with it. Different user segments — power users versus casual users, early adopters versus mainstream — react to AI-generated content in ways that diverge substantially from each other.

GrowthBook's AI testing playbook frames the challenge as "tuning AI responses across thousands of use cases." That breadth isn't incidental — it's precisely what drives variance. Add the novelty effect on top: tech-savvy early adopters tend to respond more positively to AI features than mainstream users, which means the variance you observe early in an experiment often understates the variance you'll see once the full user population is exposed.

Estimating variance before the experiment runs: three defensible approaches

The honest answer is that you can't know your AI feature's metric variance until you've run the experiment — but you can make a defensible estimate. Three approaches are worth considering.

The most reliable starting point is historical data on the closest analogous metric you have. A power analysis tool that uses the last eight weeks of historical data to estimate baseline metric values and expected traffic for a given population gives you a grounded starting point rather than a guess.

If you're launching an AI-powered recommendation feature, look at the variance in engagement metrics for your existing recommendation system, even if it's rule-based. The distributions won't be identical, but they'll be closer than starting from nothing.

If historical analogues are weak, a limited pilot rollout — even to a small percentage of traffic — gives you empirical variance data before you commit to a full experiment design. The variance estimate from a pilot will be noisy, but it's real signal.

When neither option is available, apply a conservative multiplier to your initial estimate. The direction of the error matters: underestimating variance produces an underpowered experiment, which is a worse outcome than overestimating it and running slightly longer than necessary.

Why effect sizes are harder to predict for AI features

For a conventional product change, you often have a reasonable prior on expected lift — historical experiments on similar UI changes, industry benchmarks, or a tight relationship between the change and the metric. For an AI feature, the effect on user behavior depends on model quality, prompt design, user context, and the interaction between all three. These factors don't produce a stable prior.

GrowthBook's power analysis documentation puts it directly: "Obviously you do not have complete information about the true lift, otherwise you would not be running the experiment." Their explicit recommendation is to run power analysis across a range of effect sizes — optimistic, expected, and pessimistic — rather than anchoring to a single point estimate.

For AI features, the pessimistic scenario deserves particular attention, because the population-average effect is often smaller than the effect observed in your most engaged users, who tend to be overrepresented in early data.

The cost of getting either wrong

Underestimating variance means your minimum detectable effect is larger than you planned for. In GrowthBook's documented power analysis example, the MDE at Week 1 is 34.5% — only very large effects are detectable that early. If your AI feature produces a real but modest improvement, an underpowered experiment will simply miss it.

Underestimating effect size produces the same outcome through a different path: you design for a lift that doesn't materialize, and the experiment lacks the power to detect what's actually there.

The standard 80% power threshold means a correctly specified experiment still misses real effects one time in five. Underspecification makes that rate worse in ways that are hard to recover from mid-experiment without inflating your false positive risk.

Setting conservative assumptions as a feature, not a bug

The practical implication is straightforward: when you're uncertain about variance or effect size for an AI feature — which is most of the time — assume higher variance and smaller effects than your intuition suggests. GrowthBook's guidance frames this as asking, "Suppose your feature impact is smaller than you think — how small could it be?" and designing the experiment to have adequate power at that pessimistic estimate.

Conservative assumptions translate into longer runtimes or larger traffic allocations. That's not a failure of planning — it's the correct response to genuine uncertainty. An experiment designed for realistic conditions is far more valuable than one that reaches a conclusion quickly but can't be trusted.

Calculating the AI feature test traffic you actually need

The inputs that make AI feature experiments hard to plan — elevated variance, uncertain effect sizes, non-deterministic treatments — don't change the mechanics of the sample size calculation itself. The formula still requires four inputs: your baseline metric value, the effect size you want to detect, the variance of your metric, and your desired statistical power.

What changes is how difficult it is to specify those inputs honestly. The goal of this section is to walk through those mechanics with AI-specific inputs, then show concrete techniques that can reduce the traffic requirement when your numbers look daunting.

The standard sample size formula and its inputs

Power is the probability of detecting a real effect when one exists. The conventional threshold is 80% — meaning if you ran the same experiment 100 times with different randomizations, you'd expect to observe a statistically significant result in at least 80 of them.

The minimum detectable effect (MDE) is the smallest effect size for which your experiment achieves that 80% power threshold. Smaller MDEs are better: they mean you can detect subtler improvements.

The relationship between these inputs is straightforward but unforgiving. Sample size grows as effect size shrinks, as variance increases, or as you raise your power requirement. For a simple revenue metric, if your control averages $100 and your treatment averages $102, your effect size is 2%.

Detecting a 2% lift requires substantially more traffic than detecting a 10% lift at the same variance and power level.

For AI features, the practical challenge is that you often don't know your effect size going in, and your metric variance is likely higher than historical baselines suggest. The right response is to run your power analysis across a range of effect sizes — an optimistic scenario, a realistic best guess, and a pessimistic smaller-than-expected case — so you understand the runtime implications under each assumption rather than anchoring on a single number.

CUPED reduces traffic requirements by absorbing pre-existing behavioral variance

CUPED is a statistical technique that strips out the noise that was already present before your experiment started. Here's the intuition: some users are just heavier users of your product than others, and that pre-existing difference shows up in your metrics regardless of which variant they're assigned to.

CUPED identifies those pre-existing behavioral patterns and removes their contribution from the result — leaving behind a cleaner signal of what the treatment actually caused. Less noise in the estimate means you need fewer users to reach the same level of confidence.

For AI features specifically, CUPED is particularly valuable because user engagement level is one of the biggest drivers of heterogeneous reactions to AI-generated content. A power user who already relies heavily on the product will respond to an AI suggestion feature differently than a casual user who opens the app twice a month.

CUPED can absorb much of that pre-existing behavioral difference, tightening your estimate.

Post-stratification serves a similar variance-reduction purpose and is reported to run experiments approximately 20% faster — both techniques are available natively without requiring custom statistical infrastructure.

Sequential testing for AI experiments

Sequential testing lets you check results at multiple points during an experiment without inflating your false positive rate — a meaningful capability for AI experiments where you may want to stop early if a model is causing harm, producing degraded outputs, or clearly underperforming.

The trade-off is direct: enabling sequential testing reduces power, which means you need more traffic or a longer runtime to reach the same 80% power threshold you'd achieve with a fixed-horizon test. This isn't a reason to avoid sequential testing — it's a reason to model the power impact before you commit.

A power calculator with sequential testing support lets you toggle this in analysis settings and see the effect on your projected runtime before the experiment launches.

Bayesian vs. frequentist — a practical choice for AI tests

Bayesian vs. frequentist — Modern experimentation platforms support both Bayesian (with custom priors) and frequentist statistical engines, and the choice between them is a practical one, not a philosophical one. The relevant question is: do you have reliable data about how the AI model performs before it goes live?

This might come from internal model evaluations, from running the model in the background without showing results to users (sometimes called shadow mode), or from benchmark tests on historical data.

If you have that kind of data, a Bayesian approach lets you feed it into the analysis as a starting assumption — which can shorten the time to a confident conclusion because you're not starting from scratch. If you don't have reliable pre-launch data, or if your model behaves differently in production than it did in testing, the frequentist approach is the safer default.

The practitioner community debates this actively, with some arguing Bayesian is strictly superior and others treating frequentist as a more defensible standard for organizational buy-in. The honest answer is that both are valid, and the right choice depends on the quality of your prior data.

A worked example: AI reply suggestions for customer support

Consider a concrete AI feature test scenario: an AI reply-suggestion tool for customer support agents. The primary metric is agent adoption rate — the share of sessions where an agent accepts at least one AI suggestion. Baseline adoption is 18%. The team wants to detect a 3 percentage point absolute improvement (to 21%), with α = 0.05 and 80% power. Eligible sessions per week: 8,400, split 50/50 between control and treatment.

Using the standard normal approximation for a two-sided test of two proportions with p₁ = 0.18 and p₂ = 0.21, the required sample size works out to approximately 1,732 sessions per group — roughly 3,464 total. At 8,400 eligible sessions per week with a 50/50 split, that's about 4,200 sessions per arm per week, meaning the experiment could reach 80% power in under a week under these assumptions. That's a favorable scenario.

Now compare it to a tighter traffic situation. GrowthBook's own power calculator documentation illustrates an experiment with approximately 2,195 users per week, a 14.41% baseline conversion rate, and a 20% expected relative effect size. At that traffic level, the experiment reaches 80% power at three weeks.

In week one, the MDE is 34.5% — meaning you can only detect very large effects early on. As users accumulate, the MDE shrinks and power increases. If your weekly traffic is lower than this, or your expected effect size is smaller, you may be looking at runtimes measured in months rather than weeks.

The practical takeaway: run your power analysis before you launch, use realistic variance estimates from analogous historical metrics, and model multiple effect size scenarios. If the numbers show a 12-week runtime to detect a plausible effect, that's not a calculation to ignore — it's a signal to either find variance reduction techniques, reconsider your primary metric, or set honest expectations about what the experiment can and cannot tell you.

Traffic allocation and segmentation for AI feature experiments

Hitting your target sample size is necessary but not sufficient. In AI feature experiments, the composition of your test population shapes your results as much as its size. A statistically adequate sample drawn from the wrong segment — or exposed in the wrong sequence — can produce results that are internally valid but externally meaningless, or worse, actively misleading.

Defining who actually belongs in the experiment

Not every user in your system is a meaningful participant in an AI feature test. The AI feature needs to produce substantive output for the experiment to measure anything real, which means participants generally need sufficient interaction history, relevant context, and technical eligibility — logged-in state, the right platform, enough session depth for the feature to engage.

Exposing users who barely trigger the AI feature dilutes your signal and inflates variance, because you're averaging over a large population where most of the treatment effect is zero.

Defining this eligible population precisely — and consistently — matters more than it does in a standard button-color test. Experiment targeting rules let you specify eligibility using attribute-based rules with AND/OR logic, saved groups, and org-level targeting for B2B contexts.

The key word is consistently: the same eligibility rules must apply to both control and treatment at assignment time. Eligibility criteria that are ambiguous, loosely defined, or applied differently across variations are a documented cause of Sample Ratio Mismatch.

Gradual rollout limits blast radius before full exposure

AI features carry higher risk than deterministic features precisely because their outputs are unpredictable at scale. A gradual rollout — exposing a small fraction of eligible users first, then expanding as you gain confidence — lets you catch unexpected behavior before it affects your entire user base.

The practitioner consensus on this is clear: as one engineer put it in a discussion on production testing, "every new change goes into your codebase behind a flag. You can flip the flag for a limited set of users".

The important implementation detail is that gradual rollout phases should be planned upfront, not adjusted reactively mid-experiment. Feature flags that use deterministic hashing ensure the same user consistently receives the same experience across sessions — a prerequisite for valid measurement. GrowthBook's percentage-based rollouts use deterministic hashing (MurmurHash3) so the same user always gets the same variant, without requiring server-side session storage.

When you're ready to move from rollout to formal experiment, converting a feature flag directly into an A/B test keeps the rollout and the measurement instrument unified rather than managing them as separate systems.

Avoiding sample ratio mismatch

Sample Ratio Mismatch is the failure mode where the observed traffic split doesn't match the configured split. It invalidates your results silently unless you're checking for it. Experiment platforms surface SRM as a warning in results views, but understanding why it happens in AI feature contexts is what lets you prevent it.

AI feature architectures create elevated SRM risk for several specific reasons. Server-side assignment combined with client-side tracking — common when an AI feature runs inference on the backend — creates opportunities for tracking to fire inconsistently.

Misconfigured Activation Metrics can trigger at different rates across variations if the AI feature itself changes session behavior (users who see an AI recommendation may stay longer, generating more events). Mid-experiment targeting changes are another documented cause: if you tighten or expand eligibility criteria after the experiment starts without creating a new phase with re-randomization, your traffic split will drift.

The most reliable mitigation is running an A/A test before the real experiment launches. This validates that traffic is flowing correctly and that your metrics are instrumented consistently across both groups — before any treatment effect is in play.

Managing the novelty effect

AI features are particularly susceptible to novelty effects. Users who are first exposed to an AI-powered experience often respond more positively simply because it's new and different, not because it's better.

If your initial rollout population skews toward power users or early adopters — which is common when you're doing a gradual rollout and your most engaged users are most likely to encounter the feature first — your early results may look stronger than they'll prove to be at scale.

The practical response is to run the experiment long enough for novelty to decay, and to monitor whether the treatment effect is stable over time or trending downward as users acclimate. Segment composition should also include a representative mix of user types, not just your most active cohort.

An experiment that proves an AI feature works for power users but was never exposed to casual users hasn't answered the question you actually need to answer before a full launch.

When traffic is genuinely insufficient: strategies that don't pretend otherwise

Correcting for the segmentation and allocation problems described in the previous section gets you a valid experiment — but it doesn't conjure traffic that doesn't exist. Insufficient AI feature test traffic is often a genuine structural constraint, not a planning failure you can engineer your way out of.

Your product may simply not have the volume needed to reach statistical significance in a reasonable timeframe — especially given the elevated variance that AI-generated outputs introduce. That's a real situation, and it deserves real strategies rather than a lecture about running your power calculations earlier.

What follows is a prioritized toolkit. Proxy metrics and sequential testing are the highest-leverage technical interventions. Holdout groups are best suited for AI features with compounding, longitudinal effects. Qualitative evaluation is the fallback when quantitative methods genuinely can't deliver signal. Feature flags are the operational layer that makes all of the above safer to execute.

Lower-variance proxy metrics reduce the traffic threshold without sacrificing causal validity

When your primary outcome metric — 30-day retention, revenue per user, long-term engagement — requires enormous sample sizes because of high variance or low base rates, a lower-variance leading indicator can dramatically reduce the traffic you need. For an AI-powered recommendation feature, that might mean measuring immediate click-through on AI suggestions rather than downstream conversion.

For an AI writing assistant, it might mean session depth or output acceptance rate rather than subscription renewal.

The critical design constraint here is causal validity. A proxy metric is only useful if it's genuinely connected to the outcome you care about — not just correlated in historical data, but plausibly on the causal path. Choosing a proxy for convenience rather than validity is how teams end up optimizing for something that doesn't move the needle downstream.

If your experimentation platform runs analysis directly against your data warehouse — querying BigQuery, Snowflake, Redshift, or similar sources where your event data already lives — you can often identify candidate proxy metrics from existing event streams without new instrumentation, which matters when you're already resource-constrained.

Sequential testing makes longer experiments statistically valid, not just longer

Running an experiment longer to accumulate traffic is legitimate. Peeking at results repeatedly while it runs is where the problem starts. Here's the issue: each time you check a metric for significance, there's a small chance you'll see a false positive just by chance — that's what a 95% confidence threshold means. It controls that chance for a single check.

But if you're tracking five separate metrics, the chances that at least one of them shows a false positive compound. The math works out to roughly a 23% false positive rate across five independent metrics at α = 0.05 — not because anything went wrong, but because you gave chance five opportunities instead of one. Even two metrics gets you to about 10%. Sequential testing is the tool that controls this inflation when you need to check results more than once.

Sequential testing provides the guardrails that make longer experiments statistically valid. Rather than applying a fixed threshold at a single decision point, sequential testing adjusts significance boundaries dynamically as data accumulates, preserving your error rates whether you stop early or run long.

The practical implication: commit to a maximum experiment duration upfront, use sequential testing to allow early stopping if a clear signal emerges, and don't treat "we'll just run it longer" as a free option. It isn't — not without the right statistical infrastructure in place.

Holdout groups measure cumulative AI impact that two-week tests miss

Some AI features don't produce their full effect in a two-week experiment window. A recommendation engine that adapts to user behavior, or a personalization layer that improves with interaction history, may show modest short-term lift but substantial long-term value. Standard A/B tests measure point-in-time lift; holdout groups measure cumulative, longitudinal impact.

A holdout group is a set of users permanently excluded from the AI feature who serve as a long-running control. The tradeoff is real: those users never receive the feature, which means you're permanently withholding something potentially valuable from a slice of your audience.

That cost is worth bearing when the AI feature is expected to compound over time and you need evidence of that compounding to justify continued investment. Experimentation platforms support holdouts as a named capability, and feature scheduling — the ability to activate and deactivate features at specific times — enables structured measurement windows within holdout designs.

Qualitative evaluation as a complement to quantitative testing

When traffic is genuinely insufficient for statistical significance, qualitative methods provide directional signal that quantitative testing cannot. User interviews, session recordings, human rater studies, and LLM evaluation frameworks (evals) can tell you whether the AI feature is producing coherent, useful outputs — and whether users are engaging with it in the way you intended.

This isn't a replacement for A/B testing. It's the evidence that justifies keeping a feature live while traffic accumulates, or killing it before you sink further resources into something that isn't working.

Feature flags enable safe incremental exposure, but don't replace experimental design

Feature flags let you expose an AI feature to a small percentage of traffic — say, 5 to 10% — monitor for regressions, and gradually increase exposure as confidence builds, without requiring a fully powered experiment from day one. Attribute-based targeting rules with AND/OR logic, saved groups, and org-level targeting for B2B contexts give you control over who sees the feature and when.

One distinction matters here: gradual rollout is not the same as a controlled experiment. Without randomization and a held-out control group, you can't attribute metric changes to the AI feature with any confidence. Feature flags enable safe exposure and operational control; they don't replace experimental design.

What they do provide — particularly when combined with warehouse-connected guardrail metrics — is an automated safety net. If a guardrail metric degrades during rollout, the exposure can be halted automatically before a problem compounds. That's a meaningful protection when you're shipping without full statistical power, which is exactly the situation this section is about.

Planning your AI feature test traffic: matching method to constraint

The through-line of this article is simple: the inputs matter more than the formula. Standard power calculators aren't broken — they just require stable variance and a predictable effect size, and AI features provide neither.

Non-deterministic outputs inflate your metric variance in ways that historical data won't capture, and effect sizes for AI features depend on model quality, prompt design, and user segment in ways that resist clean prior estimates. The result is that a plausible-looking sample size can mask a structurally underpowered experiment.

Estimating traffic requirements before launch: three honest inputs

Before you run a single user through your experiment, you need three honest estimates: your baseline metric variance (drawn from the closest historical analog you have, not from the AI feature itself), a range of plausible effect sizes from pessimistic to optimistic, and a realistic weekly eligible user count that reflects your actual targeting criteria — not your total user base.

Run those numbers through a power analysis at each effect size scenario, and let the runtime implications tell you whether you need variance reduction techniques like CUPED, a lower-variance proxy metric, or a frank conversation about what the experiment can actually detect.

Matching your testing approach to traffic volume and feature type

The right testing method depends on your specific constraints. Use this decision framework to identify your situation:

If you have adequate traffic and a short-horizon metric: Run a standard fixed-horizon experiment with conservative variance assumptions. This is your cleanest path.
If your traffic is tight: Switch to a lower-variance proxy metric earlier in the causal chain — output acceptance rate rather than 30-day retention, for example.
If the AI feature compounds over time: Use a holdout group, not a two-week A/B test. Short-horizon experiments will understate the feature's longitudinal value.
If you're in genuinely low-traffic territory with no good proxy: Qualitative evaluation is not a consolation prize — it's the appropriate tool for the evidence level you can actually achieve.

The tooling layer: variance reduction and flag-to-experiment unification

A power calculator that models runtime across multiple effect size scenarios before you commit, combined with native support for CUPED, post-stratification, and sequential testing, means the variance reduction techniques described in this article don't require custom statistical infrastructure to use.

The ability to convert a feature flag directly into a formal experiment also keeps your gradual rollout and your measurement instrument unified — which matters when the gap between "we're testing this" and "we're measuring this" is where most AI experiment validity problems originate.

Where to start: power analysis before experiment design

Start with your power analysis, not your experiment design. Pull eight weeks of historical data on the closest metric analog you have, estimate your weekly eligible user count using your actual targeting criteria, and run the numbers at three effect size scenarios: the lift you're hoping for, half that lift, and a quarter of it.

If the pessimistic scenario shows a runtime you can't accept, that's your signal to either find a lower-variance proxy metric or apply CUPED before you launch — not after you've been running for six weeks and the results look inconclusive.

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.

Your standard sample size calculator will give you a confident number for an AI feature experiment — and that number will probably be wrong.

Why AI features break standard sample size assumptions

Non-determinism is the root cause

How non-determinism inflates metric variance

Effect sizes are harder to predict

Why your sample size calculator will fail you

High variance and uncertain effect sizes: the core AI feature test traffic problem

Why AI outputs inflate metric variance

Estimating variance before the experiment runs: three defensible approaches

Why effect sizes are harder to predict for AI features

The cost of getting either wrong

Setting conservative assumptions as a feature, not a bug

Calculating the AI feature test traffic you actually need

The standard sample size formula and its inputs

CUPED reduces traffic requirements by absorbing pre-existing behavioral variance

Sequential testing for AI experiments

Bayesian vs. frequentist — a practical choice for AI tests

A worked example: AI reply suggestions for customer support

Traffic allocation and segmentation for AI feature experiments

Defining who actually belongs in the experiment

Gradual rollout limits blast radius before full exposure

Avoiding sample ratio mismatch

Managing the novelty effect

When traffic is genuinely insufficient: strategies that don't pretend otherwise

Lower-variance proxy metrics reduce the traffic threshold without sacrificing causal validity

Sequential testing makes longer experiments statistically valid, not just longer

Holdout groups measure cumulative AI impact that two-week tests miss

Qualitative evaluation as a complement to quantitative testing

Feature flags enable safe incremental exposure, but don't replace experimental design

Planning your AI feature test traffic: matching method to constraint

Estimating traffic requirements before launch: three honest inputs

Matching your testing approach to traffic volume and feature type

The tooling layer: variance reduction and flag-to-experiment unification

Where to start: power analysis before experiment design

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