A yellow spring seesaw on a blue playground surface, tipped down hard to one side.

Creative testing

How do I A/B test ad angles in ChatGPT ads when one ad group gets almost all the impressions?

By The Tofu team ·

You can't force an even split. OpenAI documents no A/B test, experiment or ad rotation setting for ChatGPT ads, and delivery leans toward whatever it predicts will get clicks or conversions. For a fairer test, give each angle its own campaign with the same budget, targeting, context hints, bid and landing page, so one angle can't take the others' budget. Turn off creative enhancements for those campaigns, compare results by ad, and wait for several thousand impressions per angle before you pick a winner.

Ask Tofu next

Pick a question and Tofu Ads Agent answers it in chat.

As of November 2026.

Why does one ad group get almost all the impressions?

Because ChatGPT ads delivery isn't built to share traffic evenly. OpenAI says a Clicks campaign aims for "clicks from people likely to engage," and a Conversions campaign weighs ad quality, relevance, click likelihood and conversion likelihood. All ad groups in a campaign share its budget, so the one the system rates highest can take most of it. Hints may play a part too, since they describe the needs each ad group's ads fit (see how to write context hints).

Published tests show lopsided totals. In Workshop Digital's B2B test (August 26, 2026), lifestyle photos got 45,334 impressions, against 12,560 for general graphics and 11,736 for industry-specific images. Workshop didn't say how many ads used each image or how they were grouped, so not all of that gap is down to OpenAI. An uneven split isn't a fault, but an ad that gets little delivery never gets a fair hearing.

Does ChatGPT Ads have an A/B test or ad rotation setting?

Not in OpenAI's developer docs. The campaign, ad group and ad settings have no experiment, traffic split or rotation option, and reporting has no test report. OpenAI's docs do say to "keep related creative variations together where those settings are shared," and its help article on ad groups, as seen in search results, suggests several ads per ad group to try different approaches to one theme. That lets OpenAI pick a winner; it isn't a controlled test. OpenAI's help center couldn't be read to check for anything more in Ads Manager. (Meta, by contrast, has a test tool: see testing Meta ads fairly.)

How should I set up an angle test?

Pick the setup by how clean a comparison you need:

SetupEven delivery?Good for
Several ads in one ad groupNo. OpenAI chooses which ad showsLetting OpenAI find a winner when you don't need to know by how much
One ad group per angle, in one campaignNo. Ad groups share the campaign budgetQuick reads, accepting a lopsided split
One campaign per angleCloser. Each angle gets its own budgetA comparison you can act on

OpenAI's own guidance points the same way: "Use separate campaigns when you need separate budgets." For a campaign-per-angle test, keep everything the same except the angle:

  • Objective, budget amount and budget type
  • Locations, platforms and audiences
  • Context hints (next question)
  • Bid strategy and bid amount
  • Landing page and tracking template
  • Start time

Change one thing per test: the message in the title and body, or the image, not both. Two limits remain. Equal budgets don't guarantee equal impressions, because an angle OpenAI rates lower may not spend its budget, so compare rates, not totals. And OpenAI doesn't say how it handles two of your campaigns that fit the same conversation, so your angles may still compete with each other.

If you can't afford parallel campaigns, run angles one after another for equal full weeks. That's weaker, since anything else that changes between the weeks muddies the result.

Should each angle use the same context hints?

Yes. Hints describe the needs an ad group's ads fit, so different hints per angle turn a copy test into a hint test too. Paste the same list into each ad group and check that the saved lists match. For writing hints, and updating a list without losing any, see how to write context hints for ChatGPT ads.

How do I read the results by ad?

Use reports split by ad. OpenAI's Insights API returns impressions, clicks, spend, CTR, CPC and conversions for each ad when you set aggregation_level=ad. Three habits help:

  1. Tag each ad. Put utm_content={ad_id} in the tracking template so your analytics and CRM tie visits and leads to each ad.
  2. Compare click-driven results. Reports count conversions up to 30 days after a click and 1 day after a view by default. Set the view window to 0 so one ad isn't credited for leads it was only seen near.
  3. Wait a day. OpenAI says conversion data can take at least a day to appear.

Judge B2B angles on more than CTR. A cheaper click that never books a demo isn't a winner. There's no breakdown by prompt or hint; see what ChatGPT ads reporting shows.

How much data do I need before I pick a winner?

More than most B2B tests get in a week. OpenAI gives no number; it says to review "enough volume before making large changes to bids, budgets, or creative." Standard sample-size math (95% confidence, 80% power) gives a guide:

To tell apartImpressions per angle, roughly
1.0% CTR from 1.5%7,800
1.0% CTR from 1.2%43,000

These are our calculations, not OpenAI figures. Telling conversion rates apart takes far more traffic, which is why many B2B tests have to decide on CTR and landing page engagement first.

A real example: in Workshop Digital's test, industry-specific images got a 1.2% CTR on 11,736 impressions and general graphics 0.9% on 12,560. By our calculation (a two-sided test), that gap falls short of 95% confidence. Workshop didn't report significance either. Set your metric and sample size before you start, run whole weeks, and don't stop early because one angle leads.

Should I turn off creative enhancements during a test?

Yes, on the test campaigns. An OpenAI email published by Search Engine Roundtable on October 20, 2026 said creative enhancements would switch on for new and existing self-serve campaigns from October 30, 2026, unless you had opted out. They can write new headlines and descriptions, so the copy you're testing may not be what people see. The email says campaign settings take priority over the account default and can be changed in Ads Manager at any time. It doesn't say whether reports show which version ran. For everything enhancements can change, see ChatGPT ad specs.

Should I use a fixed bid or Maximize Results during a test?

Use the same strategy, and the same amount if you set one, for every angle. A shared fixed bid removes one more difference between angles; that's our view, not an OpenAI rule. Search Engine Roundtable reported on November 5, 2026 that Ads Manager's bid options for new campaigns change from November 9, so check what new campaigns offer first: see fixed bid or Maximize Results for ChatGPT ads.

Can I test images inside one multi-image ad?

Not as a test. Enabled accounts can put 4 to 6 square images in one ad, but OpenAI's reporting splits results by country, device, platform or product, not by image, so you can't see which image did the work. Test images as separate ads, or separate campaigns. For image specs, see ChatGPT ad specs.

Sources

Let an agent build your next campaign.

Tofu Ads Agent reads your CRM, plans the funnel, builds the audience and creative, and stages full campaigns across LinkedIn, Google and Meta for your approval.