Why more variants is not the point
Every variant in a test takes a share of the delivery. With a fixed budget, more variants means less data per variant, a longer wait for a readable result and a higher chance that the apparent winner is noise. When variants differ in several ways at once, even a clear winner does not say why it won, so the next brief has nothing to build on. A test of many loosely different variants is less a test than a lottery.
The ad platforms add a second reason. Responsive formats already combine and rotate assets inside a single ad. Google documents that a responsive search ad is built from multiple headlines and descriptions, which the system assembles into different combinations, learning which perform best. Meta documents features that vary text, format and enhancements across the people who see an ad. Adding a large set of AI variants on top of that mixing makes it even harder to say what the audience responded to.
The value of AI in copy testing lies elsewhere. It lowers the cost of drafting, so the team can afford to write many candidates, compare them against the hypothesis and launch only the few that test it cleanly. It also makes it practical to write variants for angles that would never have been tried when every line cost a copywriter's time. The constraint moves from producing copy to choosing it.
| Approach | What it produces | What the team learns |
|---|---|---|
| Launch every AI variant | Many lines differing in several ways | Which line won this time, not why |
| Let the platform assemble freely | Delivery optimised to the platform's objective | Which assets served most, with limited control |
| One hypothesis, few purposeful variants | Lines differing only in the tested element | Whether the tested element matters, usable in the next brief |
| A sequence of focused tests | A record of answered questions | A body of messaging knowledge that compounds |
From hypothesis to brief
Start with a sentence that names the element, the expected effect and the reason. "Leading with the guarantee will raise enquiries from first-time customers more than leading with price, because first-time customers are worried about risk, not cost." The element is the opening angle; the expected effect is on enquiries, not clicks; the reason comes from research, reviews or sales conversations. The experimentation loop frames this as the start of a cycle that ends with a recorded learning, and the cycle only works if the hypothesis is written before the copy.
Then write the brief the model will draft from. It carries the hypothesis, the audience and their stage, the element that varies and the elements that must stay constant, the approved facts and claims the copy may use, the brand voice instructions, the format and character limits of the placement, and the number of final variants needed. The constant elements matter as much as the varying one: if the guarantee variant also mentions a discount, the test can no longer separate the two.
Tell the model what it may not do. It may not add facts, figures, awards, comparisons or promises that are not in the approved claims list. It may not change the offer. And it should mark, rather than invent, anything it thinks the copy needs but the brief does not supply.
Generating variants with a purpose
Ask the model for far more candidates than will run, grouped by the angle they test. For an opening-angle test, that might be a long list of guarantee-led openings and the same number of price-led openings, with the rest of the ad held identical. Then select. A person chooses the strongest candidate for each angle, checking that each is a fair representative of its angle and that the two differ only where the hypothesis says they should. A weak version of one angle against a strong version of the other tests the writing, not the idea.
Useful variation axes for copy include the opening angle, the benefit emphasised, the proof offered, the framing of the problem, the specificity of the promise, the call to action and the tone. Pick one per test. Where the platform format assembles several assets, such as the headlines of a Responsive Search AdA Google Ads search ad built from multiple headlines and descriptions combined at auction time., the hypothesis can be tested between two ads that each hold a consistent angle across all their assets, which is the structure the responsive search ads guide describes for comparing angles.
Models have habits that undermine a test. They converge on similar phrasing across supposedly different variants, reuse the same handful of power words, and drift towards superlatives. Read the selected variants side by side and rewrite anything that makes two angles sound alike. A brand voice pack, with its avoided-words list, keeps the drafts in the brand's register across every variant, so voice does not become an accidental second variable.
Claim and policy checks before launch
Advertising rules apply to what an ad says, however it was drafted. In the UK, the Advertising Standards Authority administers the advertising codes, which require claims to be substantiated and not to mislead; teams in other markets check the equivalent rules where they advertise. A model adds claims readily: "the fastest", "guaranteed results", a specific saving, a number of customers. Every factual, comparative or outcome claim in every variant is checked against the approved claims list and the evidence behind it before launch, and any line whose claim cannot be supported is rewritten or dropped. The AI claims and consumer protection guide covers the claims that need most care.
Platform policies apply too. Both major platforms review ads against their published advertising policies, and the help centres set out the restricted categories, the editorial and style rules and the requirements for particular sectors. A variant that would be rejected wastes a test slot and can delay the whole ad set. Check the policy pages for the category before drafting, include the relevant restrictions in the brief, and treat a policy rejection as a signal to revise the brief, not to ask the model for workarounds.
The creative approval procedure is the gate every variant passes, with claims evidence attached. AI-drafted copy does not need a different gate; it needs the same gate applied without the assumption that the words were chosen deliberately.
Fitting platform formats
Each platform documents its own format rules, and the current help centre page wins over memory. For responsive search ads, Google documents the headline and description slots, their character limits, how assets are combined and how pinning restricts which asset appears in which position. The practical consequence for AI drafting is that every headline must make sense alone and beside any other, so the brief asks for interchangeable lines, and the review reads them in combination, not only one by one.
Meta documents the text fields for each ad format, how text may be truncated in different placements, and creative features that can suggest or vary text. If those features are on, the ad that serves may not be the copy the team wrote, which matters for a controlled test. The creative testing guide covers how to isolate the copy variable on Meta, including the platform's documented A/B test tool. The guide to writing Meta ad copy sets out what each field should carry, which is worth settling before any variant is generated.
Two format checks catch most AI drafting problems. The first is length: models are unreliable at counting characters, so measure every line against the limit rather than trusting the model's claim that it fits. The second is truncation: read each variant as it will appear in the smallest placement it may serve in, since a benefit cut off mid-phrase tests nothing.
Reading results and feeding them back
Read the result on the metric the hypothesis named, at the planned sample, with the decision rule set before launch. Click-through, which the CTR calculator helps compute, is the first number to move and the least likely to decide the question: a headline that attracts curious clicks can lower the conversion rate behind it. The A/B testing guide covers sample size, Statistical SignificanceA test result unlikely enough under 'no real difference' to pass a threshold chosen in advance. and how long to run; stopping early on an exciting difference is the commonest way to declare a winner that was never there.
Platform-assembled formats need a different reading. Google's asset reporting for responsive search ads rates individual assets and shows which combinations served, which is evidence about delivery rather than a controlled comparison; the platforms' documented experiment and A/B test tools are the place for controlled answers.
Then feed the result back. Record the hypothesis, the variants, the result, the confidence in it and what it means for messaging, in a log the next brief can draw on. A confirmed angle becomes part of the brief's constant elements; a rejected one is noted so it is not tested again by accident. Over time, as variants wear out through Creative FatigueThe decline in an ad's response as the same people see it again and again., the log tells the team which angles to refresh and which to retire, and the model can be given the log as context for the next round of drafts.
Testing a guarantee-led angle for a lettings agency
Fernquay Lettings, an invented lettings agency, wants more landlord enquiries from search. Its hypothesis is that a guarantee-led message ("Rent paid on the 1st, even if your tenant is late") will beat its current price-led message ("Fully managed from 9% of rent") among first-time landlords. A model drafts 40 headlines for each angle from a brief that fixes the descriptions and landing page. The account manager selects 8 per angle, and claim review removes 3 guarantee lines that promised "zero risk", which the agency's terms do not support. Two responsive search ads, one per angle, run as a controlled experiment for 6 weeks. The price-led ad has the higher click-through, 5.2% against 4.6%, but the guarantee-led ad produces 31 enquiries against 22 at a lower cost per enquiry. The team records the result, makes the guarantee the constant opening for the next test, and briefs a second test on proof: tenant vetting against landlord reviews.
Common mistakes
- Launching every variant the model wrote. Large sets split the data and teach little.
- Varying several things at once. A winner that differs in angle, offer and tone cannot explain itself.
- Writing the hypothesis after the results. A story fitted to a result is not a test.
- Trusting the model's character counts. Measure every line against the documented limit.
- Skipping claim review because the copy "only" came from a draft. The advertiser owns every claim.
- Judging on click-through. The metric the hypothesis named decides the test.
- Reading asset ratings as test results. Platform reporting on assembled ads shows delivery, not a controlled comparison.
- Not recording the learning. Without a log, the next brief starts from nothing and old tests are repeated.
