How to add feature flags to a Go service, step by step

A feature flag in a Go service looks like a small conditional. The hard part starts when two requests arrive together, configuration cannot load, or a deployment shuts down while requests are still using the SDK. The conditional needs an owner, an identity, and a defined failure behavior.
This guide builds a small HTTP service around those decisions. A Boolean flag called new-greeting selects one of two responses. Definitions load once under a service-owned client and refresh through polling. Each request creates its own attribute context, so one visitor's targeting information does not become another visitor's state.
The result is deliberately narrow: an ordinary rollout flag, an executable service, and tests for targeting isolation and fallback behavior. It gives you a foundation for a real application without treating a successful evaluation as authorization, a delivered user experience, or an A/B test result.
Define the flag contract before adding the SDK
Choose the behavior and its baseline
Start with a behavior that can tolerate the old implementation remaining active. Our baseline returns Hello from the original greeting; the enabled branch returns Hello from the new greeting. Missing definitions, unavailable startup configuration, and values of the wrong type all select the original message.
That policy is appropriate for a cosmetic greeting. A payment calculation, entitlement, or destructive operation needs its own failure policy and authorization checks. A remotely controlled Boolean cannot decide whether a caller has permission to perform an operation. Keep those checks in the application's trusted access-control path.
Write the contract down before creating the flag:
| Decision | Contract for this service |
|---|---|
| Flag key and type | new-greeting, Boolean |
| Baseline | Original greeting, enabled value false |
| Targeting unit | A stable visitor ID for this demonstration |
| Invalid configuration | Original greeting plus a diagnostic reason |
| Initial configuration failure | Start with baseline; retry by restarting after repair |
| Flag ownership | One shared definitions owner; a fresh child for every request |
The same discipline applies to a feature flag rollout controlling a larger implementation. Decide whether users, accounts, or organizations should receive a consistent experience. An organization-level setting often needs an organization identifier as its targeting unit; switching it to a session identifier changes the population that can receive different behavior.
Also decide where the new implementation is available. If the enabled branch depends on a new database column or downstream service, deploy those prerequisites before enabling the flag. A configuration change cannot make an old binary understand a new response shape. During a rolling deployment, both old and new processes may be serving traffic, so your rollback plan must account for code versions as well as flag values.
Keep the first flag specific to one behavior. Combining a greeting redesign, a new storage system, and a changed API contract under one Boolean makes every rollout problem harder to isolate. Separate flags may be appropriate for independent changes; mutually dependent settings may instead need a reviewed coherent configuration object and a stronger runtime schema.
Our query parameter is only a repeatable targeting probe. A caller can choose visitor_id=qa-on, so it is unsuitable as evidence of a logged-in identity. In production, derive attributes from authenticated application state or an established anonymous identity mechanism. Avoid sending unnecessary personal information when an opaque identifier serves the same targeting purpose.
Create a module with a specific SDK version
You need a supported Go toolchain, terminal access, and a GrowthBook SDK connection for the intended environment. The example uses Go 1.27.4 and pins the Go SDK to v0.7.0. The published module's requirements specify Go 1.22; use a currently supported compatible toolchain rather than treating that minimum as an operational recommendation.
Create an empty directory and initialize its module. These commands use a POSIX-style shell; the Go commands themselves work in PowerShell too.
The final command should report v0.7.0. Commit the resulting go.mod and go.sum with the application. Go's module dependency workflow makes the selected version explicit; copying an unversioned snippet does not do that.
Create a Boolean feature named exactly new-greeting, with default false, in the environment delivered by your SDK connection. For initial verification, add a targeting rule that returns true when attribute id equals qa-on. Leave qa-off on the default. Publish the reviewed change to that environment before expecting the service to receive it.
Use an SDK connection key, not an administrative API credential. SDK connection endpoints deliver read-only configuration scoped to an environment and optionally a project. The name of a key does not prove it points at the intended environment; inspect the connection and its delivered payload. Keep the key and configuration host in server configuration rather than returning them in the HTTP response.
Load definitions under one service-owned client
Separate startup readiness from the SDK lifetime
The Go SDK client can share definitions while creating a separate user context for each request. Its constructor starts a configured data source asynchronously. A returned client therefore does not, by itself, prove that the initial definitions loaded.
Three contexts have different jobs. The service owner keeps polling alive. A startup timeout limits how long this service waits for the first load. Each incoming request supplies its own cancellation signal for evaluation. Using a two-second startup context as the SDK owner would cancel future work as soon as that context expires or its cancel function runs.
Create flags.go with the complete loader below. An empty key intentionally leaves the demonstration in baseline mode. A configured connection must finish loading within the startup budget or this process also uses baseline mode.
EnsureLoaded checks the configured data source's startup. Its deadline bounds the wait, while the finite HTTP client timeout bounds individual HTTP operations. The extra owner cancellation on failure matters: an initial request can still be in progress when the wait expires. This loader cancels that work and closes the root instead of leaving a discarded client refreshing in the background.
The distinction follows the data source lifecycle. Do not call EnsureLoaded on a manually populated client with no data source and expect it to validate an in-memory fixture. The tests later use that separate mode and evaluate directly.
Make initial failure an explicit application choice
Here, configuration availability does not prevent the baseline greeting service from starting. A failed initial load produces a diagnostic and a disconnected baseline for the rest of that process. Correct the connection or network problem and restart to rejoin configuration updates. The example does not silently schedule recovery for the discarded client.
You might instead refuse startup when configuration is essential to the application's contract. That is a different operational decision: return an error to the process supervisor, define readiness accordingly, and test the failure. Neither policy should hide behind an unchecked constructor error or an assumed usable empty client.
Once initial loading succeeds, polling retrieves later definitions. A one-minute interval is a refresh schedule, not a one-minute rollback guarantee. Network failures, delivery caches, and process health still affect when a specific instance sees a change. Later refresh failure can leave previously usable definitions active; distinguish that condition from a successful refresh that actually changes the value.
Connect your Go rollout
Create a GrowthBook account to configure the Boolean flag and environment connection this service reads, then verify both targeted responses before expanding traffic.
Get StartedEvaluate with fresh attributes for each request
Keep request identity out of the shared root
The shared client owns configuration, not the current visitor. Calling WithAttributes returns a child with the supplied attribute snapshot. Keep that child local, and treat its fresh attribute map as immutable after construction. Replacing a global variable with the newest request's child would reintroduce the identity leak the design is meant to avoid.
The client attribute options distinguish replacement from overrides. This guide supplies the complete request attributes, containing only id. If your application also targets account tier or region, derive and validate those fields before creating the child; do not assume old attributes are automatically carried forward.
Create handler.go. It validates the probe ID before evaluation and accepts only a real Boolean result. A string such as "true" remains invalid configuration even if it looks convincing in an editor.
There is no child Close call. Children share the root's definitions and background resources; closing a child can close the shared data source. The service will close the root once, after HTTP work drains. This ownership rule also applies when a framework or middleware makes it tempting to attach cleanup to every request.
Use runtime types and diagnostics deliberately
The type assertion is the application's contract boundary. An SDK result can hold a Boolean, number, string, object, or missing value. Compilation verifies the assertion is legal Go; it cannot prove a remotely edited value will be a Boolean at runtime. The ok result gives you an explicit invalid-value branch.
The response's reason is an application diagnostic, not an invented SDK status field. It distinguishes startup baseline from a successfully evaluated Boolean and malformed or missing values. A valid false result is not an outage, and a baseline caused by invalid configuration should not be counted as a successful rollout decision. In a real service, aggregate those categories without attaching raw personal identifiers to logs.
Passing r.Context() follows Go's request-context propagation model. It does not make a CPU-local flag evaluation perform a network request, and it does not guarantee every application operation is interruptible. Your downstream database, HTTP, or inference calls must also receive suitable contexts. Keep those time budgets separate from the lifetime of shared configuration.
The personalized response has private, no-store so an HTTP cache should not reuse one visitor's greeting for another. That header governs this response; it does not control the SDK's definition refresh. Application caches, upstream configuration delivery, and request evaluation are different boundaries that need separate verification.
Adapt the handler to a real application boundary
When replacing the probe with your application's identity, create one adapter near authentication middleware that produces the attribute map. Give it a documented input contract: an authenticated user ID, a verified account ID, and any trusted targeting fields. A query parameter saying plan=enterprise should not become a trusted subscription attribute merely because it is easy to pass to the SDK.
Choose the decision once for an operation when consistency matters. If one request evaluates the same flag before and after a background definition update, a later evaluation can see new configuration. Retaining the validated Boolean in a local variable gives that operation one chosen behavior. It does not create a transaction across multiple flags, services, or requests; those workflows need an explicit version or coordination strategy.
This is also the place to define logging limits. The demonstration returns a diagnostic reason for inspection, while an application may emit a counter tagged by flag and reason instead. Keep the selected behavior available for troubleshooting, but avoid turning a feature decision into a log of every identity and all its attributes. Configuration observability should answer an operational question with the minimum information needed.
Wire the service and prove both branches
Drain requests before retiring shared resources
Create main.go to own the server and SDK lifetime. The server listens only on localhost for this demonstration. The shutdown path waits for HTTP draining before deferred flag cleanup runs; it does not close the shared client as soon as a signal arrives.
Go's graceful server shutdown stops accepting work and waits for active connections to become idle. If the budget expires, this example closes the server and returns the error. A deployed service needs its process supervisor's termination window to exceed its drain budget, and long-lived connections require additional design.
The independent SDK owner remains alive during normal draining. Reusing the already canceled signal context for shutdown would make the drain fail immediately. Reusing it as the SDK owner could also stop background work before in-flight requests finish. These are small context choices with service-wide consequences.
Test targeting without an external connection
Add handler_test.go. This fixture uses an actual SDK client populated with local JSON definitions, including the same qa-on targeting rule. It does not start polling or call EnsureLoaded, and it closes only its root after all goroutines finish.
The HTTP testing package lets you exercise the handler without a deployed endpoint. Alternating identities in concurrent goroutines tests the specific isolation property we need: enabled requests cannot overwrite the default visitor's attributes. It does not prove that your production authentication system supplies the right identity.
Extend this test around your actual failure contract. Supply definitions with a missing key, a null default, a string default, and a numeric default; each should produce the original greeting with the invalid-value reason. Construct the handler with a nil root to verify startup baseline independently. An invalid visitor ID should produce HTTP 400 regardless of whether configuration is available, because validation precedes evaluation.
Keep the returned message in those assertions too. A test that checks only the Boolean can miss a later refactor that reports the enabled decision but executes the old branch. Conversely, checking only a successful HTTP status misses incorrect targeting. For a larger operation, assert both the decision and the externally meaningful behavior, using isolated dependencies where the operation would otherwise send messages or charge money.
Format, inspect, test, and compile all three application files plus the test. Run these inside the module directory.
All commands should succeed. When your platform has the required race-detector and C-toolchain support, also run go test -race ./.... The race detector checks concurrent paths that execute; ordinary concurrent tests and a clean race run are different evidence. Neither establishes that every future modification is race-free.
Verify real delivery separately from local targeting
Run the service with your reviewed SDK connection. Replace the placeholder with its actual key; for self-hosting, also set GB_API_HOST to the SDK API base URL. The following environment syntax is for a POSIX-style shell.
In a second terminal, check the two targeted identities and one invalid request:
Expect HTTP 200 with enabled:true for qa-on, HTTP 200 with enabled:false for qa-off, and HTTP 400 for the missing ID. The successful responses should say evaluated-boolean. If both say startup-baseline, inspect the startup diagnostic before changing targeting. If they say invalid-or-missing-value, inspect the flag key, type, environment, and delivered definitions.
Finally change the reviewed targeting value, publish it, and verify the running instance after refresh without restarting. Preserve the same IDs throughout. This check exercises delivery and application behavior together; a passing in-memory unit test cannot establish that the hosted connection or environment is configured correctly.
Repeat the checks with the SDK host deliberately unavailable in a disposable local environment. A new process should report its initial-load failure and serve the startup baseline. A process that already loaded usable definitions can behave differently during a later outage. Test both conditions, then restore connectivity and verify the recovery policy you actually implemented. They are distinct failure modes, even though both start with an unreachable host.
Operate the rollout and retire the temporary branch
Distinguish available configuration from current configuration
A service can successfully return a greeting while its definitions are stale. Initial readiness, later refresh success, evaluated values, and HTTP response success answer different questions. Monitor them separately rather than reducing them to a single SDK-is-working metric.
For gradual feature rollouts, verify the target population and identity attribute before widening allocation. A repeatable result assumes a compatible identity and configuration; arbitrary changes to rules, seeds, or targeting units can change assignment. Do not promise that a percentage rollout permanently pins a person across devices or future rule changes.
Use the following failure checks before expanding exposure:
| Failure to exercise | Expected application behavior |
|---|---|
| No connection key or failed first load | Original greeting, startup-baseline, diagnostic |
| Missing flag, null, string, or number | Original greeting, invalid-or-missing-value |
| Valid Boolean false | Original greeting, evaluated-boolean |
| Invalid or missing visitor probe | HTTP 400 before evaluation |
| Refresh outage after successful load | Previously usable definitions can remain; monitor refresh separately |
| Shutdown with active requests | Drain within budget, then cancel and close the shared root |
The default is code you maintain, not an automatic proof of safety. Feature flag defaults need the same review as the enabled branch. For sensitive behaviors, consider whether continued use of a last-known value is acceptable or whether an application-level expiry or separate availability policy is required.
Size refresh behavior for the service fleet rather than one laptop. A positive polling interval gives each process a recurring configuration-delivery task; adding replicas changes the aggregate request pattern. Measure configuration fetch volume, failure rate, and refresh duration alongside application latency before adjusting the interval. Local evaluation avoids requiring a configuration HTTP request for every greeting, but initialization, refresh, callbacks, and application code still have costs.
If a change must stop faster than your observed configuration propagation allows, that is an architecture requirement to resolve. Evaluate supported delivery strategies and deployment controls against a measured recovery target. Shortening a timer alone does not prove that an unhealthy instance will receive or apply an emergency change. Keep the current baseline behavior testable even after most traffic uses the new path.
Add measurement only when the decision requires it
This example deliberately reports application routing, not experiment lift. Returning the new greeting proves that the handler selected that branch and attempted to write a response. It does not prove a browser displayed it, a user noticed it, or a conversion resulted from it. Even a response write can fail after some bytes have left the server.
If you later turn the flag into an A/B experiment, define the randomization unit, exposure boundary, conversion join, guardrails, and analysis plan first. An evaluation callback is not automatically a delivered-experience event. Keep assignment and delivery failures observable rather than silently counting a fallback as successful treatment delivery.
Operationally, start with a targeted internal population, inspect actual responses, expand deliberately, and keep an owner for the flag. Retain the old implementation only while you need the rollback or comparison. After the new path is accepted and required compatibility windows close, remove the conditional, obsolete branch, test fixture, and remote flag through reviewed changes.
A useful first milestone is a service that can explain every response: which trusted identity it used, whether configuration loaded, whether the value met the type contract, and which implementation ran. Establish those facts before adding more flags or a more elaborate rollout policy.
Retire the flag with a plan
Open the flag cleanup workflow to inspect call sites, review behavior changes, and coordinate code removal with archiving. Running it requires configured credentials and review before mutations; reading the instructions changes nothing.
View the Cleanup WorkflowRelated Articles
T-statistic explained: What the t-value tells you about your results
An experiment report shows a t-value of 2.4. Is the product change large? Is it beneficial? Does it have a 95% chance of working? That single number cannot answer those questions.
A t-statistic compares an estimated quantity with a hypothesized value, using the estimate’s standard error as the measuring stick. It compresses the size of the observed departure and the precision of the estimate into one signed number. Understanding the two parts is more useful than memorizing a cutoff.
For experiment readers, the essential habit is to unpack the t-value before interpreting it. Establish what was subtracted, which uncertainty estimate sits underneath it, and what reference distribution applies. Then return to the effect in product units and the decision it needs to support.
Read t as a difference divided by uncertainty
Identify the numerator and its direction
The general structure is:
In an independent treatment–control comparison, the estimated quantity might be the treatment mean minus the control mean. With a null difference of zero, a positive t means the estimated treatment mean is higher. Reverse the subtraction order and the sign reverses.
“Higher” does not necessarily mean “better.” More purchases may be desirable; more errors usually are not. A negative latency difference can support an improvement, while a positive error-rate difference can indicate harm. Read the direction in the metric’s own terms before assigning a business meaning to the sign.
The hypothesized value also matters. The usual zero-difference comparison asks whether the population mean difference is zero. Testing another specified difference changes the numerator. Neither calculation independently establishes an equivalence or non-inferiority claim; those require the appropriate hypotheses, margins and decision procedure.
Distinguish standard error from standard deviation
An observation standard deviation describes spread among observed unit outcomes. A standard error describes uncertainty in an estimator, such as a sample mean or a difference between means. They can share units while answering different questions.
For an ordinary mean of independent, identically distributed observations with finite variance, the estimated standard error is the sample standard deviation divided by the square root of the sample size. The standard-error calculation therefore changes when the number of independent observations changes, even if their measured spread stays the same.
This is why dividing a mean difference by the observation standard deviation does not give the usual two-sample t-statistic. It omits how precisely the means were estimated. The distinction between standard error and standard deviation is the first denominator check to make when a t-value looks surprising.
Unit-free does not mean effect size
Both numerator and standard error have the same measurement units, which cancel in their ratio. Converting a revenue analysis from dollars to cents multiplies both by 100 and leaves t unchanged. That is a useful consistency check for a positive linear change of units.
However, t does not describe the size of the product effect independently of precision. An observed difference of one dollar can have a small or large t-value depending on the standard error. Comparing t-values across two experiments is therefore not a reliable way to rank their business impact.
The statistic measures departure from the null in estimated standard-error units. Keep the raw difference beside it so the unit-free summary does not displace the quantity the team actually needs.
The calculation also needs a usable, positive estimated standard error. A zero or undefined denominator is a reason to inspect the data and model, rather than manufacture a finite t-value or call the result certain. Software may reject such inputs; for example, the R t-test reference documents an error for effectively constant input data.
Match the denominator to the study design
Different designs estimate different uncertainty
The formula’s short form hides a substantial choice: how was the standard error calculated? Three common cases illustrate why the design must come first.
| Design | Estimated standard error | Reference degrees of freedom |
|---|---|---|
| Ordinary one-sample mean | Sample SD divided by √n | n − 1 under the classical model |
| Ordinary paired mean difference | SD of within-pair differences divided by √number of pairs | Number of independent complete pairs − 1 under the classical model |
| Independent Welch comparison | √(s² treatment / n treatment + s² control / n control) | Welch–Satterthwaite approximation using both group sizes and variances |
In the paired case, the relevant spread is the spread of differences within pairs. Substituting the two arms’ separate standard deviations discards the relationship the design uses. The paired t calculation treats the differences as its observations; the independent information comes from the pairs.
For two independent groups, Welch’s calculation does not require equal population variances. The pooled Student calculation adds that assumption and uses a different variance estimate. The Welch-versus-Student comparison helps distinguish those cases. A label reading only “t-test” can conceal the choice.
Separate a reference model from an arithmetic ratio
You can calculate a difference divided by an estimated standard error without proving that the resulting statistic has the reference distribution you intend to use. Exact classical one-sample and paired t inference relies on the corresponding independent normal-observation model; in the paired case, this concerns the differences.
Welch’s reference distribution is an approximation with estimated degrees of freedom. It is not generally exact even for unequal-variance normal populations. Non-normal outcomes can support useful large-sample mean inference under appropriate conditions, but there is no sample-size threshold that makes every heavy tail or dependence problem disappear.
If accounts are randomized, individual events from one account are not automatically independent units. A standard error that ignores the account structure may understate uncertainty. More rows in an exported file do not by themselves justify a smaller denominator or larger degrees of freedom.
Plan for more precise estimates
Explore variance-reduction methods and their tradeoffs before your next test. Match each method to the outcome you intend to estimate.
Explore Variance ReductionHold the observed effect fixed and change precision
Work through an independent mean comparison
Consider a hypothetical fixed-horizon experiment with one completed revenue outcome per independent user. Control has mean revenue of $10; treatment has $11. Each arm has sample standard deviation $4 and 100 users. The target is the arithmetic mean difference, treatment minus control, and the null difference is zero.
The estimated effect is $1 per user. Using the independent unequal-variance standard error, square each group’s standard deviation, divide by its sample size, add those terms and take the square root. Here that is √(16/100 + 16/100), or about $0.5657.
Divide the $1 difference by $0.5657 to get t ≈ 1.7678. The estimated difference lies about 1.77 estimated standard errors above the zero-difference null. The sample standard deviation is still $4; the $0.5657 figure measures uncertainty in the difference between the two estimated means.
For these equal sample sizes and equal observed variances, the Welch–Satterthwaite calculation gives 198 degrees of freedom. Under the conventional Welch approximation, the two-sided p-value is approximately 0.07864 and the 95% confidence interval for the mean difference is approximately −$0.1155 to $2.1155.
Compare two hypothetical studies
Now consider a separate hypothetical study with the same group means and standard deviations, but 400 independent users per arm. The observed $1 difference has not grown. Its estimated standard error is smaller.
| Summary | 100 users per arm | 400 users per arm |
|---|---|---|
| Observed mean difference | $1 | $1 |
| Sample SD in each arm | $4 | $4 |
| Estimated SE of the difference | $0.5657 | $0.2828 |
| t-statistic | 1.7678 | 3.5355 |
| Welch reference df | 198 | 798 |
| Conventional two-sided p-value | 0.07864 | 0.0004304 |
| Conventional 95% CI for the difference | [−$0.1155, $2.1155] | [$0.4448, $1.5552] |
Quadrupling each group’s sample size halves the estimated standard error when these observed standard deviations stay fixed. It doubles t for the same observed effect. The larger statistic reflects greater estimated precision here, not a larger observed revenue gain.
These are teaching summaries for separate studies. They are not an instruction to keep extending one experiment until significance appears. In a real extension, means and variances can change, and the stopping rule affects how the inference should be calibrated. An observed t-value also does not directly tell you the power of a future study.
Translate t using the appropriate reference distribution
Degrees of freedom and tails determine the threshold
A t-value needs a reference distribution before you can obtain a p-value or critical value. The Student t family includes a degrees-of-freedom parameter; smaller values generally give heavier tails. The same absolute statistic can therefore produce different tail probabilities at different df.
For a prespecified two-sided 5% test, the relevant critical value is the 97.5th percentile of the reference t distribution. It approaches about 1.96 as df becomes large, but 1.96 is not a universal t cutoff. Nor does every test use n − 1. Degrees of freedom depend on the calculation and may be fractional for Welch.
A one-sided alternative uses a specified tail and a different critical value. A negative statistic is not evidence in favor of an upper-tail alternative simply because its absolute value is large. Choose the direction before seeing outcomes; taking whichever tail looks favorable changes the decision rule.
A p-value conditions on the null model
For a conventional two-sided t-test, the p-value is the probability, under the null model and assumptions, of a statistic at least as extreme in either direction as the observed t. It is not the probability that the null is true, that chance caused the result, or that a future experiment will succeed.
The ASA’s p-value principles also distinguish statistical significance from effect magnitude and practical importance. A threshold crossing does not determine whether the expected benefit pays for the change. If you need the broader conditional-probability interpretation, the t-test p-value guide develops it separately.
Read the interval in the original units
The corresponding conventional interval is the estimated difference plus or minus the relevant critical t-value times its estimated standard error. For the same two-sided procedure at 5%, a 95% interval excluding zero agrees with rejecting a zero-difference null. That agreement requires compatible test, interval and adjustment settings.
Confidence describes the repeated-use coverage of the method under its assumptions. It does not assign a 95% posterior probability to the fixed true effect lying inside the particular computed interval. The confidence-interval construction makes the relationship between standard error, critical value and width explicit.
In the first study, the interval includes a small loss and a substantial gain. In the second, it excludes zero but still spans different levels of business value. A minimum worthwhile effect helps interpret that range. Neither a large p-value nor an interval containing zero proves equality; equivalence requires an analysis designed for that claim.
Audit the inputs before acting on a t-value
Trace an unexpected statistic to its components
Start with the estimated effect and null value. Confirm the subtraction order, currency or time units, eligible population and observation window. A swapped group order explains a sign change; it does not explain a change in a correctly calculated two-sided p-value.
Then inspect the denominator. Check group sizes and sample variances at the actual independent unit. Look for duplicated joins, repeated events counted as independent observations, incomplete follow-up and treatment-dependent missingness. The arithmetic can be internally consistent while the data fail to represent the intended experiment.
Inspect the analysis settings too. A ratio metric, regression adjustment or variance-reduction method can have a standard error different from the simple independent-mean formula above. Do not attempt to reproduce a dashboard by substituting top-line traffic counts into the wrong calculation. GrowthBook’s statistical effect calculations describe its estimators and frequentist reference methods, including distinct metric types.
Keep the decision rule attached to the evidence
The worked examples use conventional fixed-horizon inference. Repeatedly inspecting that calculation and stopping at a favorable threshold changes its false-positive behavior. A planned sequential analysis is a different inferential procedure; GrowthBook documents an asymptotic confidence-sequence approach with its own settings. It does not repair incorrect units or missing outcomes.
Likewise, selecting the largest t among many metrics, variants or segments creates a testing problem beyond a single comparison. Preserve the planned comparisons and applicable corrections rather than treating the most impressive number as an isolated result.
A useful readout names the estimated effect, interval, test, df, direction and sample sizes, followed by the measurement checks and decision threshold. The team can then assess what the evidence supports and whether that effect is worth shipping. Use the t-value to understand precision relative to the null; keep the launch decision grounded in effects on users and the business.
Review the result in context
Open the official experiment-analysis workflow to inspect saved results, health checks and analysis settings with configured API access. It can refresh a result snapshot when needed; stopping or rollout is separate.
View the Analysis WorkflowA published Strapi document is content. An A/B test also needs an assignment, a delivered experience, and an outcome you can attribute.
The integration becomes unreliable when those pieces disagree. An older example may read data.attributes, while your Strapi 5 endpoint returns flat fields. A numeric database ID may be mistaken for a document ID. An internal service call may return a draft even though a similar REST request defaults to published content.
This guide builds an explicit Strapi 5 content test using a small server-rendered Node application and GrowthBook. It validates the complete published content pair before enrolling a visitor, preserves a stable browser identifier, and records an exposure only after browser execution reaches the selected page.
The code also fingerprints the reviewed strings so a later edit can be rejected before assignment. The local event collector demonstrates the measurement boundary; a real experiment still needs your existing signup flow, durable analytics delivery and verified warehouse joins.
Define the Strapi 5 delivery contract
Pick a content model you can inspect
The existing GrowthBook Strapi integration guide demonstrates Dynamic Zone components selected through a frontend SDK. The reference here uses two ordinary string fields and server-side evaluation. It does not require a separately installed Strapi experimentation plugin or depend on an unverified plugin compatibility claim.
Start with a hypothesis about a meaningful outcome: an outcome-focused headline will increase completed trial signups per eligible offer-page visitor. Keep the offer and signup destination consistent if headline copy is the intervention. If the change includes a larger component package, describe that package as the treatment.
A plain-text pair makes the first integration easier to inspect. You can verify the exact response, render the selected value safely, and compare the approved strings. When extending the model to components or Dynamic Zones, validate their full structure and use a reviewed variation mapping. Reordering an array must not silently redefine which experience belongs to control.
For example, an editorial interface might place an alternative card first because it is easier to find, while the frontend still interprets index zero as control. The content can remain valid JSON and render perfectly while the experiment labels have effectively switched. Use explicit variation keys or a constrained, reviewed ordering contract when extending the simple pair. Check the rendered package against the experiment mapping, not just that every component has a title.
Use the document API that your code actually calls
Strapi 5 REST responses expose normal content fields directly under data. The document endpoint uses documentId; the numeric id in a response is not the identifier this guide puts in the URL. Some plugin endpoints have different response conventions, so this contract applies to an ordinary generated collection endpoint.
The production request explicitly sets REST status to published. This is also the REST default when Draft & Publish is enabled. The internal Document Service API defaults differently, to draft. Do not transfer that internal default—or an older v4 parameter—into this REST implementation.
Keep locale explicit as well. The requested document and locale together identify the content version being delivered. This guide checks the returned locale instead of borrowing another CMS's fallback behavior or assuming a valid document ID means every translation exists.
The document concept groups localized and draft/published variations around a document identifier. That does not make the document identifier an immutable snapshot of the text. Keep source identity, publication state and approved content as separate fields in the experiment record. Changing one is a deployment decision even when the others remain familiar.
Define assignment and measurement before launch
Use one stable assignment unit. The example creates an anonymous first-party browser ID; account-level interactions may require account assignment instead. Plan the anonymous-to-authenticated transition so signup measurement does not lose the identifier or enroll one person under several unrelated IDs.
Choose a primary business outcome, a small diagnostic funnel and relevant guardrails. A headline can increase button clicks without increasing completed or qualified signups. Estimate eligible traffic and the effect that would change your decision before starting a test that cannot resolve the question in a useful period.
Separate content ownership from experiment operation. An editor approves the published pair, an engineer owns delivery and fallback behavior, and the analytics owner verifies exposure and outcome data. Name who can change the tested content and how a change starts a new reviewed comparison.
Model, publish and inspect the content pair
Create a localized collection in development
Use an existing Strapi 5 test application with Draft & Publish enabled for the collection. Configure internationalization and the intended locale if the content is localized. The Content-type Builder belongs to the development workflow; deploy schema changes through your normal application process.
Create a collection named Experiment page with singular API ID experiment-page and plural API ID experiment-pages. A minimal schema at src/api/experiment-page/content-types/experiment-page/schema.json is:
For an existing model, merge the relevant changes rather than replacing its other fields. Restart or rebuild according to your development workflow, create a document, enter both approved headlines and publish the intended locale. Copy its actual documentId from the API response or your document inspection workflow.
Review model changes and content changes through their appropriate paths. Adding a field affects application schema and deployment; editing a headline changes a content entry. Give the editor clear field descriptions so control means the approved baseline and variant means the approved experimental alternative. A required field helps authoring, but the renderer still validates the actual response from the deployed endpoint.
With internationalized content, inspect the specific localized pair that will be eligible for the experiment. Translating control but leaving treatment in an unintended language changes more than the hypothesis describes. Keep locale targeting and content approval aligned, and decide in advance whether results across languages belong to one planned analysis or separate comparisons.
Because both alternatives are plain strings, this request needs no relation population. Components, relations and media require an explicit population strategy when added later; populate=* is not a substitute for defining the treatment's dependencies and permissions.
Use a read token on the rendering server
Create a scoped Content API token for the renderer. A read-only token supports find and findOne; a Custom token can narrow the relevant endpoint access. Keep it on the application server. You do not need to expose an admin token or broadly enable public-role access for this authenticated server request.
Read-only describes allowed operations, not a universal guarantee that a caller can never request drafts. Fix production status and locale in trusted code. Keep editor preview routes and any draft-capable configuration separate from live enrollment and analytics.
When extending the endpoint with a language switcher, select from an approved locale set rather than forwarding every browser query parameter to Strapi. A request parameter named status should not turn the production renderer into a draft-content proxy. Keep the collection path and publication choice fixed, and validate the returned document and locale even after constructing a valid request.
Check permissions on the actual route you use. A working admin interface does not prove the rendering token can read the generated Content API endpoint. Conversely, an access problem is not a reason to grant broad public access to the entire model. Confirm the token's scope and the endpoint response independently of experiment assignment.
Create a reference Node project, or integrate the files into your existing application. Use a modern Node runtime compatible with the locked dependencies, such as Node 22 or later:
Add server configuration. STRAPI_ORIGIN is the CMS origin, including a port when needed, without /admin or /api appended:
Use the appropriate HTTPS origin outside your local development setup. Store credentials through your deployment's configuration mechanism, not in public HTML or query parameters.
Validate the published response and reviewed fingerprint
The following reader fetches one document by documentId, with explicit locale selection. It rejects a missing, draft, wrong-locale or malformed pair before the SDK can assign anyone:
Leave the optional expected hash blank for the first read-only inspection, with Strapi running and the document published. Run:
Confirm the actual approved pair, source origin, document and locale. Set STRAPI_EXPECTED_SHA256 to the returned digest before enrollment. The SHA-256 calculation fingerprints those exact strings and source dimensions. It does not retrieve an older Strapi version or grant access to anything.
documentId, updatedAt and publishedAt are useful context, but they do not freeze a treatment. The fingerprint rejects changes to the reviewed pair before assignment. Identical strings can retain the same digest after republication; publication metadata is recorded separately. Extend the fingerprint when additional content becomes part of the treatment.
Treat updating the expected digest as a review step. First inspect the new delivered pair, confirm why it changed, and decide whether the current experiment can continue under a documented phase or needs a new experiment. Copying the latest digest merely to clear an unavailable response defeats the check. Preserve the approved strings alongside the digest so a reviewer can understand the change without reconstructing it from a hash.
The digest covers the two headline strings and source dimensions in this implementation. It does not cover an unrelated image, price, signup destination or application deployment. If those become part of the intervention, extend the reviewed package and its validation. A useful content contract is explicit about both its coverage and the changes that require another review.
Review the experiment before launch
Open the experiment-design workflow to turn your content hypothesis, assignment unit, and metrics into a reviewable specification. Configure API access as directed and review the plan before launch.
Open Experiment Design InstructionsRender the assigned page with request-local context
Link the content values to an experiment
Create a string feature strapi-offer-copy, defaulting to control. Link an experiment with tracking key strapi-offer-copy-v1, mapping the two string values to the intended variations. Set hash attribute id, the correct SDK environment, targeting and goals in the experiment configuration.
The content value chooses a string from the pair. The experiment and variation identifiers describe the assignment for analysis. Preserve experiment.key and result.key; do not substitute the headline or assume a variation key must equal the content value.
The GrowthBook Node SDK can share feature payload data while keeping each visitor's context in a scoped instance. The server initializes through the request path and checks success, using the SDK's caching and refresh policy rather than retaining one startup promise forever. Test stopping and refreshing on warm instances; a shared cache is not an instant global update guarantee.
Keep the browser identity stable and the output safe
Create the server below. The existing /signup application flow is a prerequisite: implement or replace that link before measuring completed signups. The collector is a local diagnostic endpoint until connected to durable analytics.
The cookie parser can produce non-string values, so identity is validated before assignment. Express cookie maxAge uses milliseconds. The redirect establishes the cookie before the offer request; its one-attempt marker avoids looping when persistence fails. Adapt identity and consent behavior to the application's existing policy.
The headline is escaped as HTML text. The JSON data element separately escapes less-than signs so a closing-script sequence cannot terminate it. These are different output contexts. Do not use this plain-string renderer as a general HTML or Rich Text renderer.
Keep failure behavior visible
A missing document, wrong locale, draft payload or changed fingerprint fails before enrollment. A failed feature-payload initialization can retain valid control content outside enrollment. Unexpected raw assignment values are logged and do not become misleading control-view events.
Fetching both alternatives in one validated payload makes this eligibility check independent of the assigned arm. A different architecture that first assigns treatment and then fetches only its content has another failure problem: a missing treatment resource can selectively remove treatment visitors from observed exposures. If larger packages require that architecture, preserve assignment and delivery failures explicitly and design the analysis population before launch. Do not infer that a missing view means the visitor never belonged to the comparison.
Preserve diagnostics and investigate the affected interval. Filtering failed treatment deliveries from analysis can select the comparison population after assignment. A safe display fallback does not make a broken experiment trustworthy; repair or halt it when delivery or configuration is unreliable.
The personalized response uses private, no-store. Shared caching of the complete validated pair is a separate choice from caching a visitor's selected HTML. Verify the actual reverse proxy and CDN behavior rather than assuming a response header alone describes every cache in the deployment.
When adding a shared content cache, include the CMS instance, document, locale, publication selection and reviewed fingerprint in the appropriate cache design. A cached English pair should not satisfy a request for a different locale, and a previously approved pair should not conceal a broken publication process indefinitely. Test invalidation and recovery using the same deployment configuration as the experiment.
The server selects one headline before producing HTML, so the browser does not need to rewrite that headline to implement this treatment. Measure the server response path under realistic load nonetheless: content fetching, payload initialization and application work still consume time. Keep loading and error guardrails appropriate to the page rather than promising that server-side evaluation has no performance cost.
Connect browser exposure to the outcome
Record a defined display boundary
The callback during evaluation queues assignments; it is not proof that HTML reached a browser. Use the deferred-tracking pattern to carry compact metadata into a client script that runs after parsing the page:
This boundary establishes that browser execution reached the rendered page. It does not prove viewport visibility or reading. Use a symmetric visibility condition if that is the intended exposure definition. A successful beacon queue operation also does not prove network acceptance or durable ingestion.
The data-element marker prevents duplicate execution within one document. Reloads can produce additional displays, and collector retries need their own event-ID policy. Define whether analysis uses first exposure or another consistent rule, then make the assignment query match that definition.
Replace local logging with verified ingestion
The diagnostic collector reads identity from the HttpOnly cookie, validates the event/source fields and discards unexpected fields. It still receives browser-supplied observations. Do not use that metadata as an authorization or billing decision, and do not treat a console row as a warehouse record.
Connect your supported analytics ingestion with appropriate validation, origin controls, delivery monitoring, retries and deduplication. Apply consent and tracker-readiness behavior consistently across both arms. Compare assignments, browser events, accepted requests and warehouse rows when investigating loss; those counts represent different boundaries.
Configure the GrowthBook data source to read experiment key, variation ID, timestamp and assignment unit from the actual stored schema. Preserve the document/locale/hash provenance in the event or a linked experiment record so the content behind a result remains identifiable.
Validate that mapping with one controlled event before increasing traffic. Compare the browser request, accepted collector fields and resulting warehouse row. Field names can change through an analytics transformation, timestamps may be stored in different zones, and a custom identifier can be dropped by a tag or export. Finding the row and joining its outcome is stronger evidence than seeing a successful HTTP response from the local collector.
Join the business outcome to the same identifier
Define completed signup or another chosen outcome with Fact Tables and metrics. Verify that it joins to the browser assignment identifier or a tested authenticated-identity mapping. A signup on another origin does not automatically carry this cookie.
Inspect an actual controlled exposure and outcome together, including timestamps, duplicate handling and follow-up window. Recent visitors may not yet have had enough time to complete the outcome. Keep diagnostic clicks and qualified conversions separate so faster but less useful signals do not replace the intended decision.
Verify publication and maintain a clean test
Check HTTP behavior before interpreting results
With the reviewed hash configured, run the application and establish a local cookie jar:
Expect a fixed redirect and consistent selected content for the same identifier while configuration remains stable. These requests inspect server HTML; they do not execute the browser exposure script. Use a real browser to inspect the beacon and collector, then verify the actual analytics and warehouse path separately.
Exercise both arms with controlled identities. A forced value can prove a rendering branch, but it does not prove ordinary randomized enrollment. Keep preview, forced and employee QA traffic outside production analysis.
Persistence is also narrower than sticky assignment across arbitrary configuration changes. The same browser ID can be used consistently while a changed experiment key, rule or allocation alters the experience. Freeze the intended configuration during the comparison, record deliberate changes, and use the platform's supported bucketing behavior when your requirements extend beyond a stable identifier.
Test the modern contract and its failure cases
| Test condition | Expected behavior | What to investigate |
|---|---|---|
| v4-style nested response | Reject missing flat fields | Migration compatibility and parser version |
| Numeric row ID used as document ID | No successful content display | Actual documentId and generated endpoint |
| Draft or wrong-locale payload | Reject before assignment | Status, localization and preview configuration |
| Reviewed string changes | Hash mismatch before enrollment | Publication approval and treatment version |
| SDK, cookie or event failure | Explicit baseline/unavailable/diagnostic path | Identity continuity and delivery loss |
Also test HTML-significant characters and a closing-script sequence in approved QA content. They should remain text, and the data element should remain intact. Repeat cache and identity checks on the actual deployment: local success does not establish proxy behavior, public rendering or warehouse attribution.
Include a schema deployment in the regression check when field names or content structure change. A correct new schema in development can coexist with older published documents or a renderer expecting the previous shape. Verify the deployed model, the delivered payload and the rendering contract together. Keep a clear unavailable/diagnostic result for incompatible content rather than disguising it as a valid control observation.
Close the loop with an owned content decision
Before interpreting lift, investigate allocation anomalies, missing data and variation-specific failures. Agree on duration, outcome maturity, meaningful effect and a decision framework before seeing the result. An imprecise estimate can leave the content choice unresolved, and nonsignificance does not prove equivalence.
When the test ends, stop enrollment, verify payload refresh in deployed instances and publish the chosen headline as ordinary content. Preserve the approved pair, source document/locale, fingerprint and metric definition with the learning. Remove retired assignment and tracking branches through review so future tests begin with a clear owner and a clean content contract.
Coordinate the content promotion with the fingerprint gate. Turning the winning headline into the new control changes the reviewed pair, so approve that resulting baseline and its new digest deliberately. Test the publication/configuration sequence in staging; do not clear a mismatch by copying a new hash without checking what it represents. The final baseline should work independently of the retired experiment rule and its exposure code.
Run a clear content comparison
Connect published Strapi content to consistent assignments and the business outcomes your team already measures. Start with a small test and verify the complete delivery and measurement path.
Get Started With GrowthBookYour randomizer can work correctly while the final experiment comparison becomes unreliable. People can enter different reporting paths, lose their identifiers, or disappear from outcome collection after assignment. The result still looks like an A/B test, but the rows left in the analysis may no longer support the question you meant to answer.
Random assignment is a powerful protection against selection into treatment. It does not guarantee representative recruitment, complete follow-up, or valid post-treatment filters. Understanding that boundary helps you distinguish an allocation problem from a problem observing the people you allocated.
The most consequential question is often what randomization lets you learn when some outcomes remain unknown. A larger sample or a more sophisticated statistical engine cannot supply those missing facts without additional evidence or assumptions.
Understand the protection randomization provides
Assignment should not depend on potential outcomes
Selection bias concerns systematic distortions introduced by how observations enter or remain in the comparison relevant to your question. The selection-bias definition emphasizes the relationship between the studied group and the population of interest. Terminology varies across fields, so name the actual selection process rather than assuming the label explains it.
With a common assignment probability, proper random assignment makes treatment allocation independent of the units’ potential outcomes. When assignment probabilities vary across prespecified strata, that independence is conditional on those strata. Product teams do not have to identify every motivation, prior preference, or unmeasured characteristic before assigning users. The allocation mechanism avoids deliberately routing more promising users into treatment and less promising users into control.
This protection concerns the units eligible for that randomization. It also relies on the implemented experiment matching the design: the assigned treatment must be well defined, identity must support the assignment, outcomes must be measured appropriately, and interference between units must be addressed where relevant.
Balance is an expectation, not a perfect copy
Randomization does not create two identical lists of people. A finite experiment can have chance differences in baseline behavior even when the mechanism is sound. The methodological basis of randomized comparisons is comparability on average under the design, rather than exact equality of every measured or unmeasured characteristic in every sample.
Investigate evidence of implementation failure, but do not repeatedly rerun allocation until the observed outcome looks attractive. Prespecified blocking or stratification can be legitimate design choices; changing allocation in response to favorable results is a different procedure.
Randomization also does not mean the observed difference equals the true effect exactly. Sampling or randomization uncertainty remains even with complete measurement. A confidence interval addresses uncertainty under its assumptions; it does not certify that recruitment, joins, filters, or outcome observation were valid.
There is another limit to what is observed. Each unit receives one assigned experience, so you do not observe that same unit’s outcome under the alternative at the same time. Randomization supports an aggregate comparison under the design; it does not reveal an individual causal effect for every user. Keeping this distinction clear prevents a clean experiment average from becoming an unsupported claim that every customer benefits.
Name the treatment effect you want to estimate. The effect of offering an experience and the effect among people who actually use it are different questions. Random allocation directly supports the first comparison under its assumptions; a user’s decision to adopt is not another random allocation.
Define the population and unit the design covers
Random assignment is different from random sampling
Assignment decides which experience an enrolled unit receives. Sampling concerns which units enter the study from a population. Randomizing a volunteer beta group can support a comparison inside that beta, while leaving a separate question about whether its result applies to customers who did not volunteer.
A narrow study is not automatically a failed study. An Android-only feature may appropriately be tested among eligible Android users. A launch claim about all platforms needs additional support. State the population, setting, and time period in the decision before debating whether the sample is representative enough.
The broader selection-versus-sampling comparison explains these categories and post-treatment subgroup examples. For an experiment review, the immediate task is to identify which population your assignment actually covers and whether the final analysis preserves it.
Choose identity and eligibility before treatment changes them
Eligibility should describe who can meaningfully receive the tested experience, using information that the treatment cannot already have changed. A supported device or an account state measured at entry can define a relevant population. Successfully using the new experience is usually an outcome, not an equivalent entry criterion.
Choose the randomization unit to match the intervention. Assigning individual employees to different collaboration settings may create interference when they work together. Assigning accounts can address some shared-experience requirements, but changes the number of independent units and the analysis needed. Randomization-unit design guidance makes identity stability and network effects part of the experiment design, not merely SDK configuration.
In implementation, preserve the assigned unit and variation through the outcome join. A device identifier that disappears after login can create missing outcomes rather than genuine non-conversions. A changing identity can also expose one person to multiple variations. Record the identity and timing assumptions with the experiment configuration, then verify the delivered data follows them.
Define the denominator first
Use the KPI playbook to distinguish binary, continuous, and ratio metrics, and connect each measurement to the experiment decision it is meant to support.
Read the KPI PlaybookMissing outcomes can defeat an apparently favorable result
Complete-case rates describe the people still observed
Suppose a hypothetical experiment assigns 100 units to each arm. The binary outcome window has ended for everyone, but a collection problem leaves some outcomes unknown. There are 18 known successes in each arm:
| Outcome availability | Control | Treatment |
|---|---|---|
| Assigned units | 100 | 100 |
| Units with observed binary outcomes | 90 | 60 |
| Known successes among observed outcomes | 18 | 18 |
| Units with unknown outcomes | 10 | 40 |
| Success rate among observed outcomes | 20% | 30% |
The complete-case comparison is 18/90 versus 18/60, an observed difference of +10 percentage points. It compares the units whose outcomes were available. It does not establish the success-rate difference for all 100 assigned units in each arm.
The larger missing fraction in treatment is a warning about observation. It is not proof of a particular direction or magnitude of bias. Equal missing fractions would not settle the issue either: equally many missing users could have different outcomes or different reasons for being unobserved.
Missingness does not inevitably create bias in every analysis. Purely random loss can mainly reduce precision; other missingness processes can select people with systematically different outcomes. The missing-value count alone does not identify which process occurred. Record the reasons for loss and what was known about affected units before treatment, rather than treating an arbitrary percentage threshold as a universal certificate of validity.
Intention-to-treat analysis with missing outcomes requires a defensible approach to those unknown outcomes. Keeping everyone’s original assignment preserves the target comparison, but the words “intention to treat” do not recover information that was never observed. Automatically assigning every missing value zero imposes an outcome assumption.
Ask what the unknown values could change
Because the example’s outcome is binary, we can calculate simple logical completion ranges. If none of control’s ten unknown outcomes were successes, its full-cohort sample rate would be 18%. If all ten were successes, it would be 28%.
Treatment has forty unknown outcomes, so its corresponding range is 18% to 58%. Across possible completions of these missing entries, the all-assigned sample rate difference could run from −10 percentage points to +40 points.
These are logical missing-outcome completion bounds, not a confidence interval and not exact bounds on a population causal effect. They hold the observed entries fixed and vary the unknown binary values. Even if every missing outcome were recovered, inference about an underlying effect would retain uncertainty from the experiment’s design.
The exercise reveals how little the complete-case +10-point result establishes without assumptions about observation. Filling all missing values with zero happens to produce 18% in both arms here. That zero difference is one completion, not a demonstrated absence of an effect.
Attrition analysis can use sensitivity checks to make missing-data assumptions visible. A useful next step is to recover outcomes from reliable records, investigate why collection failed, and assess conclusions under plausible alternatives. The most favorable completion is not a substitute for that investigation.
Preserve the comparison after assignment
Different exclusions require different repairs
Distinguish incomplete observation from a known zero. An account that was fully observed and did not purchase can legitimately contribute zero to a purchase metric. An account that no longer joins to the purchase system has an unknown outcome unless other evidence establishes what happened.
Telemetry-loss research shows why lost records can affect power and bias comparisons. Increasing traffic through the same broken collection path need not fix the problem. It may narrow the reported interval around a distorted complete-case comparison.
Inspect the metric and Fact Table definitions, including identifiers, join types, filters, timestamp alignment, and outcome availability. A left join followed by a filter on an outcome-table field can silently become an exclusion rule. A metric name does not reveal which people survived that rule.
Post-treatment filtering is another mechanism. Restricting analysis to purchasers or successful feature users can select different people in each arm when treatment changes those behaviors. Identical SQL applied to both arms does not make selected membership independent of treatment. Use such a view for a clearly limited descriptive purpose or an appropriate causal analysis; do not assume it inherits the original randomized comparison.
Triggering needs a causal and implementation argument
Analyzing a relevant triggered population can be defensible when the trigger identifies comparable eligibility across assignments. The criterion is not simply whether logging happens before or after the assignment timestamp. Ask whether the tested treatment can affect who satisfies the condition or whether that condition is recorded.
GrowthBook’s activation-metric settings explicitly warn that activation should be independent of assignment. Where a feature exists only in treatment, a carefully designed counterfactual logging approach can record who would have encountered it in control. That requires equivalent logic and validation, not just a treatment-only click event with a new name.
Passing a sample ratio mismatch check is useful evidence about allocation counts, not a guarantee that every inclusion rule is valid. Selection can change the composition of two equally sized groups. Review reasons for missingness and baseline characteristics alongside counts, rather than forcing arm sizes to match by discarding extra observations.
No statistical repair is universal. Weighting and imputation require assumptions and relevant observed information. A group absent from the data cannot be studied merely by assigning it a larger weight. A missingness mechanism related to unobserved outcomes cannot be declared harmless because a model returned a precise estimate.
For example, an imputation model may assume observed baseline information is sufficient to account for who has missing outcomes. That assumption needs a substantive argument; the unobserved values themselves cannot be inspected to verify it. Compare plausible departures from the assumption and report whether the practical decision survives. Do not choose the missing-data model because it produces the preferred launch recommendation.
Report the evidence randomization actually preserved
Track assignment separately from outcome observation
Maintain counts for eligible units, assigned units, recorded exposures, observed outcomes, and the final analysis population. Explain losses between stages by arm and by relevant baseline groups. The ledger should distinguish a known outcome from unavailable measurement, and an original exclusion from a later analytic restriction.
The data-quality requirements for online experiments include completeness, consistent assignment records, reliable joins, and timing. An allocation check alone covers only part of that path. A clean A/A test can reveal existing problems but cannot prove a future treatment will leave exposure and outcome collection unchanged.
Preserve the query or metric revision behind the result. If the team repairs a join or recovers delayed records, state whether the correction reconstructs the intended original population. If it does not, explain the resulting scope and whether a new run is needed. Avoid relabeling a selected survivor cohort as the original experiment.
Design outcome observation before increasing exposure. For an important transaction metric, a reliable backend record may allow recovery when a client event fails, provided identity and timing still join correctly. Verify that the recovery method is available in both arms and measures the intended outcome. A second event stream is useful evidence only when its own coverage and meaning are understood.
Make missing-data assumptions part of the launch decision
Alongside the experiment results, report outcome availability and the assumptions used for missing values. Transparent randomized-trial reporting treats participant flow, exclusions, and missing-data handling as information needed to judge a result. The same statistical discipline is useful for product experimentation when the participants are accounts rather than clinical-trial patients.
For the hypothetical experiment, the immediate conclusion is that the favorable complete-case result does not identify the all-assigned sample comparison without more information. The next action is to investigate observation, recover outcomes where possible, and evaluate the remaining uncertainty. It is not to declare a winner from 20% versus 30%, or a tie from zero-filled missing outcomes.
Randomization gives you a defensible starting contrast. Preserving it requires that the data and analysis continue to describe the units and outcomes the decision concerns. Before launching from a randomized result, be able to explain not only how people received a variation, but also whose outcomes are known, whose are missing, and what your conclusion assumes about the difference.
Inspect the metric definition
Open the read-only metric search workflow to find a fact metric and inspect its aggregation, filters, identifiers, and associated fact table. Running the workflow requires configured GrowthBook API access; the instructions do not execute an audit when opened.
View the Metric Search WorkflowReady to ship faster?
No credit card required. Start with feature flags, experimentation, and product analytics—free.






