Experiments

How to test multiple prompts in production (without chaos)

May 31, 2026

min read

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

Shipping a prompt change to production without a rollback plan is the LLM equivalent of deploying code with no way to revert — except the failure mode is quieter, slower to surface, and harder to trace back to the exact change that caused it.

Most teams discover this the hard way: a wording tweak that looked fine in testing quietly degrades quality for a long tail of real users, and by the time support tickets arrive, the damage is already done. The fix isn't more careful manual review. It's infrastructure — the same kind engineers already use for code, applied to prompts.

This article is for engineers and PMs shipping AI features who need a repeatable, safe way to test prompt changes with real users. If you've outgrown "run it a few times and ship it" but haven't yet built the scaffolding to do better, this is the practical foundation you need. Here's what the article covers:

Prompt versioning — how to track changes, promote across environments, and answer "which prompt is running right now?" in seconds instead of hours
Traffic splitting — how to route users to prompt variants consistently using deterministic bucketing, so your experiment data stays clean and your users don't see incoherent behavior
Quality metrics and event tracking — how to define what "good" looks like before an experiment goes live and instrument the signals that tell you whether a variant is actually better
Rollback — how to stop a bad prompt variant instantly using feature flags, guardrail metrics, and gradual ramp-up schedules that limit how many users are affected before you catch a regression

Each section builds on the last. Versioning makes rollback possible. Traffic splitting makes measurement valid. Measurement makes rollback automatic. None of it works in isolation, and none of it requires building from scratch — the infrastructure patterns here map directly onto tools your team likely already has or can adopt quickly.

Why prompt testing in production is fundamentally different from traditional software testing

Most engineers shipping AI features start in the same place: write a prompt, run it a few times, check that the output looks reasonable, and ship. It's the same instinct that drives unit testing — define an input, assert an output, move on.

The problem is that this instinct, which serves you well across almost every other layer of the stack, breaks completely when applied to LLM prompts. Not partially. Not in edge cases. Completely.

Understanding why requires a different mental model before any conversation about tooling makes sense.

The determinism problem — why unit tests don't work on LLM outputs

Traditional software testing is built on a single assumption: given input X, you will always get output Y. That assumption is so foundational that it's rarely stated explicitly. It's just how software works.

LLMs violate it at the most basic level. The same prompt, sent to the same model, with the same parameters, can return meaningfully different responses across invocations. This isn't a bug you can fix or a configuration you can tune away — it's intrinsic to how these systems generate text.

The practical consequence is that string comparison, the simplest and most common form of automated testing, fails even when the model is doing exactly what you want. "The answer is 42" and "42 is the answer" are semantically identical but string-different.

Any test that asserts an exact output will produce false failures constantly, and any test loose enough to avoid false failures will miss real regressions.

There's no middle ground that makes unit tests viable here. This isn't a tooling gap — it's a fundamental mismatch between the testing paradigm and the system being tested.

Why manual "try and tweak" fails at scale

If unit tests don't work, the natural fallback is manual iteration: run the prompt, read the output, adjust the wording, repeat. This works fine for demos.

It fails spectacularly when you're deploying prompts that thousands of users depend on, where a subtle regression can cause support ticket floods overnight before any engineer notices.

The failure mode isn't that manual testing is slow — it's that it's invisible. Production inputs differ from the inputs you tested against in ways that are impossible to anticipate manually. A prompt change that improves results on your curated test cases can quietly degrade quality on the long tail of real user queries. Degraded responses often remain hidden until users report them, and by the time that signal reaches your team, the damage is already done.

Quality is multi-dimensional — you can't optimize for one axis

Even if you could solve the determinism and scale problems, there's a third challenge that traditional pass/fail testing has no mechanism for handling: prompt quality isn't a single thing.

A response can be accurate but poorly formatted. It can be well-structured but off-topic. It can be relevant and complete but unsafe in certain contexts. Correctness, relevance, safety, and task completion are distinct quality dimensions that often exist in tension with each other — improving one can degrade another.

A test that checks whether the model answered the question tells you nothing about whether the answer was appropriate, complete, or formatted in a way users can act on.

This is why a single "good/bad" evaluation is insufficient. Prompt quality requires scoring across multiple axes simultaneously, and those axes have to be defined before deployment — not inferred from user complaints after the fact.

Infrastructure is not optional — it's the prerequisite

These three problems together — non-determinism, invisible scale failures, and multi-dimensional quality — mean that responsible prompt testing in production requires purpose-built infrastructure. A test case repository. An execution engine that can run prompts against representative inputs at scale.

Output scoring across quality dimensions requires its own evaluation pipeline. And without a results dashboard, regressions stay invisible until users find them.

Without this, teams aren't testing prompts — they're guessing and hoping. The rest of this article covers the specific infrastructure layers that make production prompt testing viable: versioning prompts like code, splitting traffic between variants with consistent user bucketing, tracking the right quality metrics, and maintaining the rollback capability that makes any of this safe enough to run on real users. Each of those pieces exists because the problems described here are real, and the instinctive shortcuts don't survive contact with production.

Unversioned prompts are undebuggable prompts

When an AI feature breaks in production, the first question an engineer asks is: which prompt version is actually running right now? For most teams, that question has no clean answer.

The prompt lives somewhere — maybe hardcoded in application logic, maybe in an environment variable, maybe in a Notion doc someone updated last Tuesday — and reconstructing what changed and when becomes an archaeology project rather than a debugging session.

This is the same class of problem as unmanaged feature flags: invisible changes, untraceable regressions, and rollbacks that require guesswork instead of a revert command. The fix is the same too. Prompts need to be treated as first-class infrastructure artifacts, not as strings that get tweaked in a source file between deploys.

Why hardcoded prompts create technical debt

The failure mode is predictable and it compounds quietly. Most LLM applications start with a developer writing a prompt, testing it against a handful of inputs, and shipping it because it works well enough. Over time, small wording changes accumulate — handling edge cases, improving tone, supporting new user scenarios. Each change seems minor. But because nothing is tracked, teams lose visibility into what changed and why.

When output quality drops, rolling back means guessing what the previous prompt looked like rather than reverting to a known version. The deeper problem is proliferation. Without a single source of truth, prompt versions scatter across repositories, environment variables, dashboards, and notebooks.

Engineers hesitate to make even small edits because they can't predict what else might break. That hesitation is the clearest sign that a system has accumulated real technical debt — not in the code, but in the prompts driving it.

Git-like versioning for prompts

The mental model that resolves this is straightforward: apply the same discipline to prompts that you already apply to code. Every change should be recorded with what changed and why.

Versions should be immutable and linked to the model and parameter settings they run with, so teams can reproduce past behavior and understand the impact of any update. A centralized prompt registry with clear environment assignments means every output can be traced back to the exact prompt that produced it.

The operational implication is that a prompt change should go through the same review and traceability mechanisms as a code change — not because the process is bureaucratic, but because it's what makes safe iteration possible. You wouldn't push code straight to production without version control, testing, and proper deployment processes; your prompts deserve the same treatment.

Parameterized prompt templates

Static prompt strings break down quickly in production systems. A prompt that works for one user context needs to accept dynamic inputs — session data, user preferences, retrieved context, feature state — without spawning a new version for every variation.

Parameterized templates solve this by separating the stable structure of a prompt from the variable content injected at runtime.

A template defines the logic and tone of the prompt as a versioned artifact. At runtime, variables fill in the dynamic slots. This means the versioning system tracks changes to the prompt's structure and intent, while runtime flexibility is handled through injection — not through proliferating near-identical prompt strings across a codebase.

Environment promotion and decoupling deploys

A mature prompt versioning workflow mirrors the dev/staging/prod promotion model that engineering teams already use for code. A prompt change is authored and tested in development, validated against representative inputs in staging, and only then promoted to production — explicitly, not automatically. Each environment gate is an opportunity to catch regressions before they reach users.

The operational payoff of this model is decoupling. When prompts are embedded in application code, every wording change requires a full code redeploy. A prompt registry backed by feature flag infrastructure breaks that dependency: prompt changes can be deployed independently, allowing product teams to iterate on prompt behavior without waiting for an engineering release cycle.

GrowthBook SDKs evaluate flags locally from a cached JSON payload with sub-millisecond evaluation and no per-request API calls, providing exactly this kind of runtime update capability, with prompt variant selection happening at the SDK layer without adding latency to the LLM call path.

The audit log infrastructure that comes with a proper feature flag system also closes the traceability loop. When a prompt regression surfaces, the answer to "which version is running right now?" becomes a lookup, not a forensic investigation.

Splitting traffic between prompt variants without disrupting users or contaminating experiment data

The most common mistake teams make when running their first prompt A/B test is treating it like a coin flip — randomly assigning each incoming request to one variant or another. This feels reasonable until you realize that the same user might receive Prompt A on Monday and Prompt B on Wednesday, with no awareness that anything changed.

Two things break simultaneously when this happens: the user experience becomes incoherent, and the experiment data becomes uninterpretable. You can't attribute a support ticket, a retry, or a thumbs-down to a specific prompt variant if you don't know which variant that user actually experienced.

This isn't a theoretical concern. It's the most direct path to Sample Ratio Mismatch — a condition where the distribution of users across variants diverges from what you intended, making any statistical comparison between them unreliable. Mid-experiment targeting changes made without re-randomization are a documented cause of this kind of contamination. The fix isn't more careful manual tracking. It's consistent bucketing from the start.

Why consistent bucketing is non-negotiable

Consistent bucketing means that a given user always receives the same prompt variant for the duration of an experiment. This requires assigning variants based on a stable user identifier — a user ID, a device ID, a session token — rather than on the state of each individual request.

The mechanism that makes this work reliably is deterministic hashing: a hash function takes the user identifier as input and maps it to a value between 0 and 1. That value determines which bucket the user falls into, and because the hash function is deterministic, the same identifier always produces the same bucket assignment.

There's no database lookup, no session state to maintain, and no risk of a user drifting between variants as traffic conditions change. This is meaningfully different from random-per-request assignment, and conflating the two is the most common traffic-splitting mistake in prompt testing in production.

Feature flags as the traffic-splitting layer

The right abstraction for prompt routing isn't custom middleware, environment variables, or a database flag you toggle manually. Feature flag infrastructure already solves the exact set of problems prompt experiments require: consistent bucketing, real-time percentage adjustment without redeployment, audit logging, and kill switches. Building a bespoke routing layer means rebuilding all of that from scratch — and getting the edge cases wrong.

Feature flag systems that implement local evaluation are particularly well-suited here. GrowthBook SDKs download flag rules as a locally cached JSON payload and evaluate every flag check in-process with zero network latency, so there are no per-request API calls per flag check. For prompt routing, where you're already paying the latency cost of an LLM call, adding a synchronous network round-trip to resolve a flag assignment compounds quickly across thousands of daily requests. Local evaluation eliminates that dependency entirely — the app keeps routing correctly even if the flag service is temporarily unreachable.

GrowthBook's SDKs, for example, evaluate flags locally with sub-millisecond evaluation from a cached JSON payload. The JS SDK exposes the underlying bucketing mechanics directly: a user assigned to a 60/40 split would be evaluated against ranges: [[0, 0.6]] for the control and ranges: [[0.6, 1.0]] for the variant — a concrete illustration of how percentage-based routing maps onto consistent hash values.

Gradual rollouts and targeted segments

Two practical scenarios come up repeatedly in production prompt testing. The first is a gradual rollout: you want to expose a new prompt variant to a small percentage of traffic initially — say 5% or 10% — and increase that percentage over time as confidence grows, without touching application code.

The second is a targeted rollout: you want to route only a specific segment of users to the new variant, such as beta users, users on a specific plan tier, or users in a particular region.

Both scenarios depend on the hash attribute being correctly set at SDK instantiation. If the identifier isn't present when the SDK initializes, users can't be assigned to a variant and will default to the control — silently, without error. This is worth testing explicitly before any experiment goes live.

One additional edge case matters for teams running multiple prompt experiments simultaneously: ensuring those experiments don't overlap. If two experiments can assign the same user to different variants independently, the interaction effects between those assignments contaminate both experiments. Mutually exclusive bucketing — where user populations are partitioned so no user can appear in more than one experiment — is the correct solution, and it's a feature worth verifying your flag infrastructure supports before you need it.

A final consideration for long-running experiments: if you need to update targeting rules mid-experiment — adding a new user segment, adjusting percentage splits — doing so without re-randomizing can cause the variant distribution to shift in ways that invalidate prior data.

Sticky bucketing addresses this by locking each user's variant assignment at the point of first exposure, so subsequent targeting changes don't cause users to switch buckets. It's the mechanism that makes gradual rollout adjustments safe rather than dangerous.

Measurement architecture has to exist before the experiment, not after the regression

"Without a clear way to measure changes in output quality across inputs, degraded responses often remain hidden until users report them." That observation captures the central risk of prompt testing in production — and it points directly to the solution: measurement architecture has to be built before the experiment goes live, not assembled reactively after something breaks.

This section is specifically about prompt evaluation — measuring outcomes from live prompt variants — not prompt engineering or version management, which are covered elsewhere in this article. The distinction matters because evaluation requires a different kind of infrastructure: event schemas, scoring pipelines, and metric definitions that connect what a user received to what they did next.

Define quality criteria before the experiment goes live

Teams that decide what "good" looks like after reviewing early results aren't running experiments — they're doing post-hoc rationalization. Quality criteria need to be specified before a single user is bucketed into a variant.

The most useful framework for this comes from how the evaluation community defines prompt quality dimensions: correctness (did the response answer the question accurately?), relevance (was it appropriate to the user's context?), safety (did it avoid harmful or policy-violating content?), and task completion (did it actually accomplish what the user needed?). These four dimensions cover the vast majority of failure modes teams encounter in production LLM features, and they're concrete enough to instrument.

The practical implication is that before deploying a prompt experiment, you should be able to answer: which of these dimensions are we testing, how will each be measured, and what threshold constitutes a meaningful regression? If you can't answer those questions, the experiment isn't ready to run.

Design an event tracking schema that links exposure to outcome

The most common instrumentation failure in prompt experiments isn't missing data — it's data that exists but can't be joined. GrowthBook's troubleshooting documentation describes a specific failure mode that illustrates this exactly: an experiment showing 38,919 total users but empty columns for Baseline, Variation, and Chance to Win.

The cause is mismatched identifier types between the exposure event (which records that a user was assigned to a prompt variant) and the metric events (which record what that user did afterward). The data is there; it just can't be linked.

The fix is schema design that treats the exposure event as the anchor for all downstream analysis. When a user is assigned to a prompt variant, that event needs to fire with a stable identifier — a user ID, session ID, or device ID — that will also appear on every outcome event you care about. The outcome events themselves (a task completion signal, a quality score, a feedback submission) don't need to be complex, but they must share that identifier or be linkable through a join table.

GrowthBook's two-query model — an Experiment Assignment Query for exposure and a separate Metric query for outcomes — makes this architecture explicit. GrowthBook's warehouse-native model runs the statistical analysis directly against your existing data warehouse, which means you're instrumenting into your existing event tracker rather than exporting data to a separate analytics pipeline. The experiment results live where your data already lives. But the identifier alignment requirement is yours to get right upfront.

Automated scoring and behavioral proxies

Once the schema is in place, you need signals to put into it. Automated scoring converts the quality dimensions defined earlier into numeric outputs that can be tracked as experiment metrics. Rule-based scoring handles the deterministic cases — response length bounds, format compliance, keyword presence for safety filtering.

LLM-as-judge approaches, where a separate model evaluates response quality against a rubric, handle the more subjective dimensions like relevance and task completion. Both produce scores that can be stored as binomial or count metrics and analyzed per variant.

Behavioral proxies fill in where automated scoring isn't feasible at scale. When users regenerate a response, submit a thumbs-down rating, or abandon a session immediately after receiving a reply, those signals correlate with response quality even without direct scoring. The identifier alignment requirement described in the previous section applies here too.

GrowthBook's metric window configuration — controlling how long after exposure a behavioral signal is counted — lets teams account for the fact that some quality signals, like a user returning to rephrase their question, may surface minutes or hours after the initial interaction.

The reusable metric library approach matters here for teams running multiple concurrent prompt experiments. Defining a "task completion" binomial metric or a "session abandonment" rate metric once and applying it across experiments reduces instrumentation overhead and keeps quality definitions consistent across comparisons — which is the only way cross-experiment learnings stay meaningful over time.

Rollback capability is the prerequisite for running experiments on real users

Every other part of a production prompt testing workflow — versioning, traffic splitting, event tracking — only works if your team is willing to actually run experiments with real users. And teams won't do that unless they trust they can stop something that's going wrong. Rollback capability isn't a nice-to-have at the end of the checklist. It's the prerequisite that makes the whole system safe enough to use.

Without a reliable rollback path, teams default to the safest possible behavior: testing only in staging, capping experiments at tiny traffic percentages, or avoiding production tests entirely. The result is a team that "iterates without fear" only in environments that don't reflect real usage.

When a bad prompt variant does reach production — and eventually one will — the question isn't whether you'll need to revert it, but whether you can do it in seconds or hours.

Kill switches and flag-based rollback

The mechanics here are straightforward but worth being explicit about. When a prompt variant is served through a feature flag, reverting it requires no code deploy. You disable or revert the flag, and all traffic immediately routes back to the known-good prompt. That's the entire operation.

The practical implementation detail engineers often overlook is what happens when the flag evaluation service is unreachable. Your system needs a defined default — the prompt your application falls back to when it can't reach the flag service. That default should always be your last known-good variant, not an empty string or an error state.

Rollback speed also depends on your evaluation architecture: GrowthBook SDKs evaluate flags locally from a cached payload, so flag changes propagate without a network round-trip during an incident, and your application continues to function correctly even if GrowthBook's servers are temporarily unavailable.

GrowthBook's Safe Rollouts feature adds an automated layer on top of this — an Auto Rollback toggle that disables the rollout rule automatically when a guardrail metric fails, without requiring an engineer to be watching a dashboard at the moment something goes wrong.

Setting guardrail metrics that trigger automatic rollback

Guardrail metrics are the specific signals you've designated as regression indicators — the ones where any degradation warrants stopping the rollout. For prompt testing in production, these typically include LLM error rates, response latency, and behavioral proxies from the event tracking layer you've already instrumented: retry rates, thumbs-down signals, session abandonment after a response.

One design decision worth understanding: GrowthBook's guardrail threshold is always set to zero. The system triggers a failure as soon as there is statistical certainty that a metric is being harmed at all — even by very small amounts. This is more aggressive than most teams expect, and intentionally so. The goal isn't to measure how bad the regression is; it's to catch any regression as early as possible.

The practical implication is that you should keep your guardrail metric set small and focused. Choosing too many guardrail metrics increases the chance of false positives — a metric that fluctuates naturally triggering a rollback on a perfectly good prompt variant. Error rates, latency, and one or two behavioral proxies is a reasonable starting point.

Guardrail metrics in this architecture are pulled directly from your existing data warehouse rather than a separate analytics pipeline, which means the signal is the same source of truth your data team already uses.

Capping blast radius with gradual ramp-up

Structural traffic limits are a guardrail in their own right. Starting a rollout at 10% exposure rather than 50% means that if something goes wrong before the monitoring system catches it, the number of affected users stays small.

GrowthBook's Safe Rollouts use a fixed ramp-up schedule — 10% to 25% to 50% to 75% to 100% — that completes within the first quarter of the configured monitoring duration. A four-day monitoring window, for example, ramps to full traffic by the end of day one, then spends the remaining three days monitoring at full rollout with guardrails active.

This structure is intentional: the ramp-up phase limits blast radius while the system accumulates enough data to detect regressions. The extended monitoring phase at full traffic catches slower-moving degradations that wouldn't be visible at 10%.

Sequential testing lets you stop early without corrupting your false positive rate

Stopping an experiment early based on interim results is statistically dangerous under standard frequentist testing. If you check results continuously and stop when p < 0.05, you'll see false positives far more often than your significance level implies — because you're effectively running multiple tests on the same data.

Sequential testing solves this with a different kind of math. Instead of calculating significance once at the end of a fixed experiment window, sequential testing uses a method where checking results early doesn't corrupt the final answer — the confidence intervals stay valid no matter when you look. The practical effect: the system can trigger a rollback the moment it's statistically certain something is getting worse, without the false alarm rate increasing because you checked early.

GrowthBook's Safe Rollouts use a one-sided version of this — meaning the system is only watching for harm, not for positive impact. That's intentional. For operational safety decisions, the question isn't "is this prompt better?" It's "is this prompt hurting anyone?" Those are different questions and they warrant different statistical tools.

One counterintuitive implication: if results are still inconclusive after the full monitoring duration, the recommended action is to ship. There's no statistical basis for extending monitoring indefinitely on a prompt that hasn't shown harm.

Safe Rollouts are designed for operational decision-making, not long-term learning — when the monitoring window closes without a regression signal, the absence of evidence is sufficient to proceed.

Turning one experiment into a repeatable prompt testing system

The four layers this article covers — versioning, traffic splitting, event tracking, and rollback — aren't independent improvements you can adopt in any order. They form a dependency chain. Versioning gives you something to roll back to. Consistent bucketing gives you experiment data you can trust. Event tracking gives you the signal that triggers rollback. Rollback gives your team the confidence to actually run experiments on real users instead of hiding behind staging. Pull out any one layer and the others lose their value.

The minimum viable stack for prompt testing in production

You don't need all of this in place before you run your first production prompt experiment. What you do need is: a versioned prompt with a known-good fallback, a feature flag that routes users with consistent bucketing, and at least one behavioral metric — a retry rate, a thumbs-down, a session abandonment signal — that shares an identifier with your exposure event. That's enough to run a safe, measurable experiment. Everything else is refinement.

Start with the prompt that has the most exposure and the least visibility

Start with the prompt that has the most user exposure and the least visibility into quality. That combination — high traffic, no measurement — is where silent regressions do the most damage and where the payoff from instrumentation is highest.

A prompt that handles 10 requests a day from internal users is not the right first experiment. A prompt that runs on every customer-facing query and has never had a guardrail metric attached to it is.

From one-off experiment to continuous prompt improvement loop

The goal isn't to run one clean experiment. It's to build the muscle so that prompt changes go through the same review, rollout, and measurement cycle as code changes — without requiring heroic effort each time. That happens when your metric definitions are reusable, your bucketing infrastructure is already in place, and your team has shipped at least one rollback successfully. The first experiment is mostly about proving the system works. The second one is where you start learning something useful.

One honest tension worth naming: the infrastructure described here takes real effort to set up correctly, and there's a temptation to shortcut the event schema or skip the guardrail metrics on the first pass. Resist that. The identifier alignment problem — exposure events and outcome events that can't be joined — is the most common reason prompt experiments produce no usable data. Getting the schema right once saves you from rebuilding it after every experiment.

The other tension is organizational. Prompt testing in production requires product, engineering, and data to agree on what "good" looks like before an experiment launches. That conversation is harder than the technical setup. But it's also the conversation that makes the results meaningful.

This article is meant to give you the foundation to have that conversation and build that infrastructure with confidence — not from scratch, but by applying patterns your team already knows to a problem that's newer than it looks.

What to do next: Identify one prompt currently running in production that has no version history and no quality metric attached to it. Before anything else, answer two questions about it: what identifier would you use to bucket users consistently, and what behavioral signal — a retry, a rating, an abandonment — would tell you if a new variant made things worse? If you can answer both, you have enough to instrument your first safe prompt experiment. If you can't answer either, that gap is your actual starting point — and GrowthBook's unified platform is a practical place to close it without building the bucketing and rollback mechanics from scratch.

Related insights

Sign up for free

Take Growthbook for a spin, no credit card required.

Create my account

Example H2

See All Articles

Experiments

Feature Flags

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

Oct 21, 2026

min read

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

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

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

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

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

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

The debate combines three different questions

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

Choosing an interface answers only part of that job.

How can the agent reach the product?

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

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

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

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

What procedure should the agent follow?

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

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

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

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

What context makes the procedure relevant?

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

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

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

GrowthBook puts the durable parts below the interface

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

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

The API provides the operations

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

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

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

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

Skills provide the workflow guidance

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

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

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

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

MCP provides a thin bridge

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

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

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

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

Explore the shared flag workflows

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

View the Flag Skill

The same disabled flag can use either route

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

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

Connect and verify before preparing the write

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

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

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

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

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

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

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

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

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

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

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

Resolve real IDs and construct one payload

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

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

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

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

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

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

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

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

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

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

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

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

Submit once, then read the object back

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

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

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

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

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

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

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

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

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

Verify through the CLI get command:

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

Or invoke growthbook_api_read with:

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

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

Choose the route that fits the runtime

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

Shell access favors terminal composition

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

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

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

Remote tool access favors MCP clients

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

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

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

Use a small decision table

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

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

Compare actual outcomes, not tool counts

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

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

Keep the evaluation task constant

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

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

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

Test failure behavior deliberately

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

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

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

Record the versions behind the result

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

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

Build workflows that survive the next client

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

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

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

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

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

Start with a common foundation

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

Get Started With GrowthBook

Experiments

Analytics

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

Oct 21, 2026

min read

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

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

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

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

Understand the difference in units

Variance uses squared distances

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

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

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

Standard deviation returns to the original scale

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

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

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

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

Calculate both from the same observations

Work through a small dataset

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

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

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

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

Choose sample or population formulas deliberately

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

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

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

Verify the calculation with explicit function names

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

from statistics import variance, stdev, pvariance, pstdev

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

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

Find the noise you can reduce

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

Read the Variance Reduction Guide

Use standard errors for experiment uncertainty

Distinguish users from estimated means

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

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

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

Combine arm variances before taking the square root

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

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

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

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

Avoid interpreting intervals as outcome ranges

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

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

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

Check what your metric actually measures

Match the variance to the randomization unit

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

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

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

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

Inspect tails instead of deleting them casually

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

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

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

Keep units consistent across pipelines

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

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

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

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

Turn spread into a better testing plan

Define the outcome before planning the sample

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

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

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

Reduce estimation noise without hiding behavior

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

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

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

Report the right quantity for the decision

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

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

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

Build a clearer experiment readout

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

Get Started With Experimentation

Experiments

Analytics

How to calculate true positive rate (TPR)

Oct 21, 2026

min read

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

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

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

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

Start with actual positives, not alerts

Define the positive class

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

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

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

Fix the unit and observation window

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

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

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

Calculate TPR from the confusion matrix

Put every evaluated item in one cell

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

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

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

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

Divide detected positives by all actual positives

The calculation is:

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

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

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

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

Handle zero positive support explicitly

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

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

Check the experiment error risks

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

Read the False Positives Guide

Compare TPR with precision and false positives

Keep the denominators visible

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

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

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

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

Choose a threshold for the decision

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

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

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

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

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

Make sure the labels can support the claim

Audit predicted negatives too

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

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

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

Preserve uncertainty and segment support

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

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

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

Separate label delay from label absence

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

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

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

Keep comparable populations together

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

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

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

Connect detection quality to product decisions

Distinguish classifier recall from statistical power

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

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

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

Evaluate the intervention as well as the detector

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

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

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

Use a repeatable reporting checklist

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

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

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

Measure what the change delivers

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

Start Measuring With GrowthBook

Ready to ship faster?

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

Get Started

Book a Demo

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

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

Shipping a prompt change to production without a rollback plan is the LLM equivalent of deploying code with no way to revert — except the failure mode is quieter, slower to surface, and harder to trace back to the exact change that caused it.

Why prompt testing in production is fundamentally different from traditional software testing

The determinism problem — why unit tests don't work on LLM outputs

Why manual "try and tweak" fails at scale

Quality is multi-dimensional — you can't optimize for one axis

Infrastructure is not optional — it's the prerequisite

Unversioned prompts are undebuggable prompts

Why hardcoded prompts create technical debt

Git-like versioning for prompts

Parameterized prompt templates

Environment promotion and decoupling deploys

Splitting traffic between prompt variants without disrupting users or contaminating experiment data

Why consistent bucketing is non-negotiable

Feature flags as the traffic-splitting layer

Gradual rollouts and targeted segments

Measurement architecture has to exist before the experiment, not after the regression

Define quality criteria before the experiment goes live

Design an event tracking schema that links exposure to outcome

Automated scoring and behavioral proxies

Rollback capability is the prerequisite for running experiments on real users

Kill switches and flag-based rollback

Setting guardrail metrics that trigger automatic rollback

Capping blast radius with gradual ramp-up

Sequential testing lets you stop early without corrupting your false positive rate

Turning one experiment into a repeatable prompt testing system

The minimum viable stack for prompt testing in production

Start with the prompt that has the most exposure and the least visibility

From one-off experiment to continuous prompt improvement loop

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