Documentation version: 2026-11-16.2
Canonical: https://docs.creator.gg/guides/email-experiments.md

# Prepare and inspect email A/B tests

Your own agent can write email variants and prepare a supported campaign experiment. This uses Creator Agents campaign tools, not a separate AI editor or a new model connection. Read guides/email, guides/copywriting and guides/email-content first.

## Prepare without sending

1. Identify the existing draft with email_campaigns_list and email_campaigns_get. Agree one hypothesis, the audience, sender, what changes between variants and the success measure. Do not silently replace an existing campaign or audience.
2. For a subject test, use abTest.testType subjectLine with two to four distinct variations containing subject. The email content is shared.
3. For a content test, use testType emailContent and two to four distinct existing HTML or plain-text templateId values. Read each template and prepare any requested changes separately. Each variation can also have its own subject. Visual-builder variations require the native editor to preserve their structure.
4. Supply testDuration in seconds, testSize as the audience percentage and winningCriteria openRate or clickRate. Creator Agents currently accepts a positive duration up to seven days and a test size from 1 to 100 percent. These are this tool’s bounds, not a promise about every native editor setting.
5. Call email_campaigns_preflight with campaignId, recipients, emailMeta, optional timeZone, abTest and an optional future sendAt. It checks the current draft, sender, audience references and variant templates. The result is an unsaved plan with saved, sent and scheduled all false. No delivery request is made. Read all warnings.

A preflight does not save A/B settings into a draft, reserve a version, verify domain readiness, score spam risk, send a test email or render an actual Gmail/Outlook inbox preview. It can become stale. There is no claimed API for the separate AI email-sequence canvas, inbox preview, domain warm-up or native visual mobile controls. Keep their actual Beta/Labs/access requirements and use the matching account screens when needed. An agent can prepare responsive HTML and review content without claiming those native checks ran.

## Send or schedule only within the requested scope

Pass the same reviewed abTest to email_campaigns_send or email_campaigns_schedule with a stable requestId. A schedule also requires a future sendAt. The service repeats current-state and audience/template checks before submitting the documented campaign request.

Both actions have real effects and native usage charges. After the test duration, the delivery service can automatically send the winning variation to the remaining audience. A request to draft, compare variants or run preflight is not permission to send the test or its winner. Resolve any genuine audience, content or timing ambiguity first. This is not a dry run.

The native AI editor’s Labs availability does not establish this connection’s campaign permissions. A discovered tool, successful preflight or accepted request is not proof of delivery. Missing access remains an access error, not a request to replace a working connection.

## Read variant results

Read the campaign again. Its variations contain sourceId and isWinner where the delivery service reports them. Call email_campaigns_stats with campaignId and variationSourceId copied from that same campaign. Another campaign’s source ID is rejected. Omitting variationSourceId preserves the normal campaign-level statistics read.

Keep variant and aggregate metrics separate. Do not add them together or infer winner status from incomplete counts. Open rates can be distorted by mail privacy features; choose the measure that fits the customer job. A missing source, unavailable category or failed read is not zero engagement.

## Recovery

Reuse a requestId only for an identical retry. After an uncertain send or schedule, inspect the campaign and operation receipt before any repeat; never generate a new identity simply to retry. True test emails and native pause/cancel controls remain account-screen actions. This guide does not authorize a test send, customer enrollment or automated follow-up.
