Launch a split test from one sentence in ChatGPT

At the end, a split test exists in Adrails with two to four variants: the ad set or ad as it runs, and one to three new ones created on Meta or Google Ads, paused. The sentence took a minute to write; the review card took two to read. Launching the test is one more confirmation on its page in Adrails, and from then on Adrails measures it about once an hour.

The winner is not a judgement. It is the variant that leads on the metric you chose, CPA or conversion rate, at 95% confidence against every other variant, once each of them has at least 10 conversions. When the end date passes first, the test ends with no winner and nothing is paused.

Connect Adrails to ChatGPT through Developer mode, as the setup article explains, and select Prepare advertising changes on the consent screen. The proposal needs an owner or admin on an active plan. The launch and the end are confirmed in Adrails, by an owner or admin too; any member can prepare them there.

The ad set or ad you start from must be active on the platform: a paused parent is refused with the reason. A Meta split test on an ad set has three more conditions. The campaign runs on Ad set budget without budget sharing, the ad set has its own daily budget, and the account belongs to a Meta business portfolio. A test on an ad has none of these conditions.

• Permissions: Read workspace data, Prepare advertising changes. Owner or admin, active plan.
• Evidence fresher than 24 hours: the tool refuses a stale read of the account.
• For a creative test: images or videos already in the account's library, by hash or video id, or asset id on Google.

campaigns_structure returns the account as of the last sync: campaigns with type, status, budget and bidding, ad sets or ad groups, and ads with their format, copy, final URL and review status. The test starts from one of these ids, exactly as returned. It becomes variant A and keeps running as it is.

Pick the entity by the question. An audience, placement or bidding question starts from a Meta ad set. A copy question starts from an ad: a Meta ad for ad_text, a Google responsive search ad for search_ad_text. A creative question starts from a Meta ad with one image or video, or a Google display or Demand Gen image ad.

• Prompt: Read campaigns_structure for account <id>, campaign <id>. Which ads run in it, with their primary text and headline?
• Returned: the ad set and its ads with ids, copy, status and review.

The sentence names the entity, the variable and the variants. Metric, duration and budget split are yours to add: when you leave them out, the card shows the defaults and marks them as such, CPA and 14 days. Duration runs from 4 to 30 days. Shares apply to a Meta split test of ad sets only, one percentage per variant, A first, each at least 10, adding up to 100.

ChatGPT calls campaigns_split_test with a name in your convention, the variable, the entity id and one to three variants. The reason, up to 300 characters, cites the observed figure with its source and period. Each variant is complete: a new ad's full text, a new ad set's full audience. Each changes only the tested variable; the tool refuses a variant that changes two.

• Sentence, Meta ad: Test two new primary texts against ad <id> on account <id>, decided on CPA, for 14 days: one that opens on the price, one that opens on delivery time.
• Sentence, Meta ad set: Split test ad set <id> on audience, 25 to 44 against the current 18 to 65, 50/50 of the daily budget, decided on conversion rate, 10 days.
• Sentence, Google: Test a new set of 10 headlines and 3 descriptions against responsive search ad <id> on account <id>, decided on CPA.
• Returned: a pending status, the card with each variant's change, the metric and duration marked default or chosen, and the review link.

Open the review link within 30 minutes. The Review advertising change page shows the test's name, the variable, each variant's change and the budget split when there is one. Confirm the change there, or decline it. Only the person who authorized the connection can confirm it, as a current owner or admin on an active plan.

Once you confirm, Adrails creates the new ad sets or ads on the platform, paused, and records the test as not launched. Nothing spends. The test now has a page in Adrails, reachable from Campaigns, with the variants, the metric, the duration and a Launch test button.

Launching is not an MCP tool. On the test's page in Adrails, choose Launch test. Adrails checks the platform first: the campaign and the ad set must be active and the variants ready to switch on. The launch is one change waiting in Priorities on Home for 24 hours until someone confirms it. Confirming it switches the variants on and, for a Meta split test, creates Meta's study.

If the platform changed since the launch was prepared, the confirmation is refused and you prepare it again. Before launch, Cancel test stops everything; what was created stays on the platform, paused.

Adrails checks running tests about once an hour, counting only the days since the launch. Conversion rate compares conversions per click, per link click on Meta, or per impression when a variant has no clicks. CPA compares each variant's conversions with its share of the spend.

The leader is tested against every other variant at 95% confidence, corrected for the number of variants. No winner is named before every variant has at least 10 conversions. When the leader is reliable, the status becomes Winner found and the test appears in Priorities. When the end date passes first, the status becomes Ended and nothing is paused.

Once a winner is found, choose End test: pause the others on the test's page. Adrails prepares a pause for each losing variant, waiting in Priorities. For a Meta split test, confirming one of them also stops Meta's study. From ChatGPT, campaigns_set_status can pause a loser too, as its own proposal, but it does not stop the study.

Keep the losing variants paused rather than deleted: actions_recent keeps their pause with its 24-hour, 3-day and 7-day observations, and a paused ad can be resumed from the same tool.

The common failures are a test that cannot exist and a test that cannot finish. The tool refuses the first in words. The second is yours to avoid: ten conversions per variant on a low-volume ad set can take longer than 30 days.

• Carousels, dynamic creative and existing posts cannot be tested. Google ad groups cannot either: test the ads inside them.
• A creative variant uses only media a tool returned for this account, by hash or video id. An invented one is refused.
• A reason without its figure, source and period is a weak card. The tool asks for the observed reason, not the wish.
• Rate limits: 60 calls and 10 proposals a minute per connection, one call at a time.
• A proposal expires after 30 minutes; a launch waiting in Priorities after 24 hours.
• Two variables in one test, or four new variants, are refused: one variable, one to three new variants.