AI for Ad Copy: Variants With Guardrails
Let the Model Write the Variants, Not the Strategy or the Claims
Short answer
Use AI for ad copy by writing the brief yourself and letting the model produce variants inside it. Fix the angle, audience, proof, offer and banned claims before prompting, build in platform limits, and check every claim against evidence before launch. Advertisers need proof before an ad runs, whoever wrote it.
1. What is AI actually good for in ad copy?
Producing many distinct executions of one angle, fitting copy into character limits, and adapting a message to different placements or objections. It is not good at choosing what to say or at knowing what is true about your product.
| Task | AI is | Why |
|---|---|---|
| Choosing the angle and offer | Weak | Depends on your positioning, margins and what competitors already claim |
| Writing 15 headline variants of one angle | Strong | Fast, varied, and easy to review against a brief |
| Fitting copy into limits | Useful, but verify | Models are unreliable at counting characters themselves |
| Adapting one message per objection | Strong | Given the objections and the answers, it rephrases well |
| Stating product facts, prices, results | Risky | Will fill gaps with plausible numbers and claims |
| Writing customer quotes or testimonials | Never | Invented endorsements are deceptive |
The angle is the part that decides whether an ad works, and it should come from your positioning and a read of what competitors are already saying. The write ads from competitor research guide covers how to find an open angle. This guide starts once you have one.
2. What should I give the model before asking for ad copy?
A short brief with seven fields: the angle, the audience, the proof, the offer, claims you may make, claims you must not make, and the platform format. Without it, the model writes generic ads with invented specifics.
| Field | What it contains | Example |
|---|---|---|
| Angle | The one idea every variant expresses | Canceled slots refill themselves |
| Audience | Who sees it and what they are doing now | Owners of clinics and salons with one to five rooms or chairs, taking bookings by phone and a paper diary |
| Proof | Evidence you can show if asked | Automatic waitlist refill on every plan; two customer stories with permission |
| Offer | What the click gets them, exactly as on the landing page | A free guided setup call in the first week (as listed on the landing page) |
| Allowed claims | Specific statements you have verified | "Waitlist refill on every plan", "set up in an afternoon" (from onboarding data) |
| Banned claims | Things the model must not say | No "best", no "#1", no competitor names, no savings figures, no invented quotes |
| Format | Platform, placement and limits | Google responsive search ad: headlines up to 30 characters, descriptions up to 90 |
3. Which platform limits should the prompt build in?
For Google responsive search ads: up to 15 headlines of 30 characters and up to 4 descriptions of 90 characters, any of which may be combined. For Meta, check the Ads Guide for the placement, since recommended text lengths vary by format.
Google responsive search ads accept a minimum of 3 and a maximum of 15 headlines, each up to 30 characters, and 2 to 4 descriptions of up to 90 characters, and Google assembles combinations from them [s1]. That has a direct consequence for AI prompts: every headline must make sense next to any other, so variants should not depend on being read in order or repeat each other.
Pinning a headline to a position guarantees it shows, but Google notes that pinning reduces the number of combinations that can be matched and is not recommended for most advertisers [s1]. Keep pins for the one line that must always appear, such as a required disclaimer.
Meta publishes recommended text lengths per placement and format in its Ads Guide, and the recommendations are short [s2]. Build the current recommendation for your placement into the prompt, and put the key message in the first line, since longer text can be cut off.
4. Which prompts produce usable ad variants?
Four patterns cover most needs: a Google headline set built in categories, a description set, Meta primary text variants on one angle, and objection-led variants. Each pastes the brief and restricts the model to its claims.
Write 15 headlines for a Google responsive search ad using only the brief below. Each headline must be 30 characters or fewer including spaces, make sense on its own, and make sense next to any other headline. Write them in groups: 4 that state the angle, 3 that name the audience, 3 that state a proof point from the allowed claims, 3 that state the offer, 2 that include the keyword [KEYWORD]. Do not repeat the same idea in two headlines. Use only the allowed claims. Do not use any banned claim. Output a numbered list, one headline per line. BRIEF: [PASTE BRIEF]
Write 4 descriptions for the same responsive search ad, each 90 characters or fewer including spaces. Description 1 expands the angle, description 2 gives a proof point, description 3 answers the objection "[MAIN OBJECTION]", description 4 states the offer with a clear next step. Each must stand alone. Use only the allowed claims and none of the banned ones. BRIEF: [PASTE BRIEF]
Write 5 versions of primary text for a Meta ad using the brief below. All five express the same angle, but each opens differently: a question, a specific situation the audience recognizes, a short customer-style observation (not a quote, not attributed to anyone), a direct statement of the benefit, and the offer first. Put the core message in the first sentence. Keep each version within [RECOMMENDED LENGTH FROM META ADS GUIDE]. Write in this voice: [PASTE 3 RULES FROM YOUR BRAND VOICE]. Use only the allowed claims. BRIEF: [PASTE BRIEF]
Here are the three objections our audience raises most often, with the answer we give to each and the evidence for it. For each objection, write one headline under 30 characters and one description under 90 characters that answer it without restating the objection negatively. Use only the answers and evidence provided. OBJECTIONS, ANSWERS AND EVIDENCE: [PASTE] BRIEF: [PASTE BRIEF]
Grouping headlines by category in the first prompt is what makes the set work. A model asked for "15 headlines" tends to produce fifteen phrasings of one line; asked for categories, it produces a set Google can actually combine into different ads.
5. What guardrails keep AI-drafted ads safe to run?
Check every claim against evidence before launch, match the offer and price to the landing page, never publish invented testimonials or numbers, and keep competitor trademarks out of ad text unless you have checked the policy and the law.
The legal responsibility for an ad sits with the advertiser, not with the tool that drafted it. In the US, the FTC's guidance for small businesses says advertising must be truthful and not misleading, and that advertisers must have evidence to back up their claims before the ad runs [s3]. Comparative claims are allowed when truthful, and need the same support [s3].
Below are ad variants and the brief they were written from. For every variant, list each factual claim it makes. For each claim, say whether it appears in the allowed claims, and quote the allowed claim it matches. Mark any claim that is not in the allowed list, any number, any superlative such as best or fastest, any competitor name, and any wording that could read as a customer quote. Do not rewrite the ads. ADS: [PASTE VARIANTS] BRIEF: [PASTE BRIEF]
Four rules for the human reviewer, applied to every variant:
- Every factual claim appears in the allowed claims, and you could show the evidence to a regulator or a customer.
- Prices, discounts and offer terms match the landing page and your product catalogue exactly, including expiry dates.
- No invented quotes, reviews, ratings or statistics, even in a style that looks illustrative.
- No competitor name or trademark in ad text. Google may restrict a competitor trademark in ad text after the owner complains, while bidding on it as a keyword is not restricted by that policy [s4].
6. How many AI variants should I actually test?
Fewer than the model can write. Generate widely, review hard, and launch a small set that differs on one thing at a time, so a result tells you something. Small budgets cannot separate many variants from each other.
The temptation with AI is to launch everything it produces. On a small budget that spreads spend so thin that no variant collects enough results to be judged. A better habit is to generate twenty, cut to the five that best express the angle, and test those against one another or against your current best ad.
Change one thing at a time between variants: the opening, the proof point or the offer. If two variants differ in everything, a winner tells you nothing about why it won. The creative testing for small budgets guide covers how to size a test.
Key Takeaways
You write the brief, the model writes the variants
Angle, audience, proof, offer, allowed claims, banned claims and format come first. Without them the model writes generic ads.
Build platform limits into the prompt
Google responsive search ads take up to 15 headlines of 30 characters and 4 descriptions of 90, combined in any order.
Ask for categories, not a count
Headlines grouped by angle, audience, proof and offer combine into real variety; "15 headlines" produces one line fifteen ways.
Proof before the ad runs
Every claim must be supported before launch, whoever wrote it. Run a claim check against the allowed list.
No invented quotes, numbers or competitor names
The riskiest lines in AI ad copy are the plausible specifics. Keep them out unless they are verified.
Generate many, launch few
Small budgets need a small set of variants that differ on one thing, so results are readable.
Frequently Asked Questions
Sources
- About responsive search ads Google Ads Help, September 2026.Headline and description counts and character limits, and Google guidance on pinning.
- Meta Ads Guide Meta, September 2026.Meta specifications and text recommendations by placement and ad format.
- Advertising FAQ's: A Guide for Small Business US Federal Trade Commission, September 2026.Truthful advertising, substantiation before an ad runs, and comparative claims.
- Trademarks, Advertising Policies Help Google, September 2026.How Google handles trademarks in ad text and keywords.
Explore More
Related analyses, benchmarks, and industry insights
Related Guides
Glossary Terms
Turn reading into a reaction
Oppira watches your competitors, keeps your playbook current, and drafts the response. See your market clearly by tomorrow.