📗 This is the user manual for the Suite for ChatGPT Ads, the tool in the Ninja Scripts panel. Here you can search it, listen to it, highlight it and save it to favourites. It is updated with every release of the tool.
A ChatGPT Ads ad is a card: one headline (16 to 24 characters read), one line (32 to 48), a square image and a URL. With that little room, testing isn't optional. And testing here means creating variants and comparing them.
🧪 The lab, group by group
You get in from 🧩 Ad groups. At the top, the group's ads with their performance and a cautious verdict, for information only:
- An ad gets an opinion with at least 500 impressions and 20 clicks.
- «Best in group»: CTR 30 % above the group (or cost per conversion 30 % below).
- «Underperforming»: CTR 40 % below, or spend with no conversions when the group already converts.
It never pauses anything by itself.
Requesting variants
The AI starts from your ads and their CTR, the group's hints, the landing page, your business and — if you have the Suite for Google Ads — the terms that convert. Each variant comes with its angle and its hypothesis: what's being tested and why.
And something that makes the difference by the third or fourth round: the AI receives the map of angles already tested in that group, with their outcome. So it won't propose again what already lost, it iterates on what won and it prioritises new angles.
Only variants that fit the card and don't repeat an existing headline or line get through.
The image: four sources
Chosen per variant, before anything is created:
| When it makes sense | |
|---|---|
| The neighbouring ad's | When what you're testing is the copy: keeping the image isolates the variable |
| 📤 One you upload | PNG, JPG or WEBP up to 5 MB. The most controllable |
| 🎨 One your AI draws | With your Gemini or Ideogram key in «My accounts». Square, a single focus, no text or logos (the card already carries the headline) |
| 🔗 By URL | If you already have it published |
You see it in the proposal before the ad is created, you can ask for another or remove it, and it only travels to OpenAI when you press create.
The first ad in an empty group
A newly created group has no ads to start from. That used to be a dead end; now the lab drafts the first ones from the group's hints and the landing page you point it at. With that, an ad's image can be set before or after creating the campaign.
🧪 Tests: from variant to verdict
Every variant you create opens a test: challenger, the base it competes against, hypothesis, angle and metric. The test sits «ready» while the variant is paused and starts when you activate it (from the panel or from the Manager: the read detects it and takes the first day with impressions as the start).
Every morning it's judged, with a real statistical test on CTR — and on conversion rate when both arms have at least 5 conversions. Four possible answers:
| Verdict | What it means | What you're offered |
|---|---|---|
| Wins | A real and sufficient difference | Promote the winner (it becomes the group's base) and archive the old one |
| Loses | Same in reverse | Archive the challenger |
| No difference | Plenty of sample and no signal | The older one stays; you can free up the slot |
| Not enough data | After 28 days without reaching the floor | Closed with no verdict |
The minimums per arm: 500 impressions, 20 clicks and 3 complete days. And you can always choose Keep both: no verdict archives anything by itself.
⚠️ The honest limit, and it's written on the screen too: ChatGPT Ads doesn't let you control ad rotation, so OpenAI decides how impressions are split. This is an observational comparison, not a controlled experiment. That's why the minimums are stricter than they'd be in a test with a guaranteed split: when in doubt, no winner is declared.
The rhythm
One live test per group. With two at once the result can't be attributed: if there are, they're flagged as «overlapping» (the pairwise comparison still holds; the effect on the group doesn't). And if a group with traffic — over 2,000 impressions in 30 days — goes 21 days with no test, 💡 Recommendations tells you: that loop is what makes an account improve.