What this page establishes
- A responsive search ad holds up to 15 headlines of 30 characters and 4 descriptions of 90 characters, and Google counts characters including spaces and punctuation, not words.
- Google's Misrepresentation and Unreliable Claims policies prohibit unsubstantiated superlatives like "best" or "guaranteed" without qualifying evidence, and AI models produce these phrases by default unless explicitly told not to.
- Negative keyword lists benefit from AI's pattern-matching on a real search terms report, but the report has to be pasted in — a model has no visibility into which queries actually triggered your ads.
- Ad strength ratings in Google Ads reward headline diversity, not headline volume, so 15 near-identical headlines score worse than 8 genuinely different angles.
- The FTC's guidance on endorsements and testimonials, most recently updated in 2023, applies to ad copy referencing customer results — a fabricated statistic in ad copy carries the same regulatory exposure as one in an article.
- Pinning a headline to a fixed position in an RSA disables Google's automatic optimization for that slot, and Google's own documentation recommends pinning only when a legal or brand requirement demands a specific headline always appear.
- A/B testing ad copy variants requires waiting for statistical significance, not eyeballing a short-run click-through-rate difference — Google Ads itself will not report a winning combination with confidence until enough impressions have accumulated.
The two limits every Google Ads prompt has to state
Google's responsive search ad format allows up to 15 headlines, each capped at 30 characters, and up to 4 descriptions, each capped at 90 characters. These limits are enforced by character count including spaces and punctuation, not by word count, which means a model asked for a "short headline" will routinely produce something 34 or 38 characters long — close enough to look right and long enough to be rejected at upload.
Stating the limit as a number rather than an adjective removes the ambiguity: "30 characters maximum, including spaces" is checkable in a character counter; "short and punchy" is not. The same applies to descriptions at 90 characters, and to the newer expanded formats Google has rolled out for specific campaign types, which carry their own separate limits documented in Google Ads Help and worth checking against the current policy page before a large batch goes out, since these limits have changed more than once.
Display, Performance Max and other formats carry their own limits
Two related formats carry different limits worth knowing before a prompt goes wrong. Display ads within Google Ads support up to 5 headlines at 30 characters, 5 long headlines at 90 characters, and 5 descriptions at 90 characters — enough overlap with RSA limits to cause confusion if a prompt is copied across formats without adjustment. Performance Max asset groups add a headline limit of 4 short headlines minimum and a business name field capped separately at 25 characters, which a generic "write ad headlines" prompt will not know to respect unless told explicitly which format it is drafting for.
A related detail worth stating in the prompt itself: Google's character counter treats certain characters, including some emoji and special symbols, inconsistently across ad formats, and copy that passes a plain character count can still trigger a length warning at upload if it contains characters outside the standard set. The safest instruction for a model is to stick to standard letters, numbers, and basic punctuation unless a specific symbol is a genuine part of the brand name.
A prompt for a full headline set
The failure mode to design against is redundancy — 15 headlines that all say the same thing in slightly different words score worse for Ad Strength than 8 headlines covering genuinely different angles, because Google's system combines headlines and descriptions dynamically and needs real variety to test combinations against each other.
Asking the model to self-report the character count does not replace checking it — a model's arithmetic on character counting is unreliable — but it does surface headlines that are obviously over length before they reach a character counter, saving a review pass.
Why "best" and "guaranteed" are a compliance problem, not just a style choice
Google's advertising policies on Misrepresentation and Unreliable Claims prohibit ads that make unsubstantiated claims a reasonable consumer would rely on, and superlatives without qualification — "best," "#1," "guaranteed" — sit squarely in that category unless the advertiser can substantiate them with real, checkable evidence such as an independent ranking or a specific, honored guarantee policy.
A model asked to write "compelling, persuasive ad copy" will reach for these words by default, because they are common in the training data for advertising text and the instruction rewards persuasiveness without constraining its source. The fix is an explicit forbidden list in the prompt, paired with a substitution rule: instead of "best-in-class support," specify the actual thing that makes support good — "24-hour response time" or "average 4.8-star rating from 2,000 reviews" — because a specific, checkable claim survives both Google's policy review and a customer's later scrutiny in a way a vague superlative does not.
Negative keywords: AI reads the search terms report, it does not generate the data
A search terms report — the actual queries that triggered your ads — is data a model has no access to unless you paste it in. What AI does well, given that report, is spot patterns across hundreds of rows faster than a manual scan: irrelevant queries clustering around a shared word, queries indicating the wrong buyer intent, or queries for a product you do not sell.
The instruction not to block converting queries matters because an overly broad negative keyword can silently cut off traffic that was actually working — a mistake that shows up as a slow decline in impressions weeks later rather than an obvious immediate error.
Match type choice changes the risk profile of each suggestion. A broad match negative keyword blocks the widest range of related queries and carries the highest risk of unintentionally excluding good traffic; an exact match negative blocks only that precise query string and is the safest default when uncertain. Asking the model to default to exact match unless a clear pattern justifies broader exclusion keeps the negative list conservative until performance data confirms a wider block is safe.
Ad extensions: sitelinks, callouts and structured snippets need the same specificity rule
Sitelink and callout extensions suffer from the same genericness problem as headlines when prompted loosely. "Learn More," "Shop Now," and "Contact Us" as sitelink titles waste an opportunity Google gives advertisers to show something specific — a named product line, a specific offer page, a real support number.
Restricting sitelinks to a supplied list of real pages, exactly as with internal linking prompts for SEO content, prevents the model inventing a plausible-sounding page that does not exist on the site.
Structured snippets follow a stricter format still — Google requires them to fit a predefined header category such as "Types," "Styles," or "Brands," with values listed rather than written as sentences. A prompt for structured snippets should name the specific header category and ask for single-word or short-phrase values only, since a model given no format guidance will often write full sentences that Google's structured snippet format rejects outright.
Pinning headlines: know when it helps and when it fights the system
Google Ads allows pinning specific headlines or descriptions to fixed positions within an RSA, which guarantees that text always appears but disables Google's automatic testing and optimization for that slot. Google's own Ads Help documentation recommends reserving pinning for cases with a genuine legal or brand requirement — a required disclaimer, a trademark that must appear verbatim — rather than using it by default because a particular headline feels important.
A useful prompt addition when generating headlines: ask the model to flag which, if any, of the generated headlines contain content that would need to be legally accurate every time it appears (a licensing number, a required disclosure, a regulated claim), since those are the genuine candidates for pinning, and everything else should be left to Google's testing.
Ad Strength is about combination diversity, not headline count
Google Ads rates each RSA's Ad Strength from Poor to Excellent based on how many distinct, high-quality headline-and-description combinations the system can assemble, which rewards genuine variety over volume. Fifteen headlines that are minor rewordings of the same sentence produce a lower Ad Strength than eight headlines covering distinctly different value propositions, because the underlying combinatorial diversity is what the score measures.
This audit step catches a mistake that is easy to make when generating headlines in a single batch: asking for 15 headlines produces 15, but does not guarantee 15 genuinely different ones.
Testing ad copy: wait for significance, not for a feeling
A common mistake after launching AI-generated ad variants is declaring a winner after a day or two based on a visibly different click-through rate, when the sample size is too small for the difference to be statistically meaningful. Google Ads itself will not confidently attribute performance differences between combinations until enough impressions have accumulated for each, and pulling a variant early based on a short-run gap risks discarding copy that would have performed equally well or better over a longer run.
A reasonable prompt-adjacent discipline: before asking a model to "write better versions of the winning headline," confirm the sample size and time window actually support calling it a winner — typically at least a couple of weeks and enough impressions per headline for Google's own reporting to show a clear divergence, rather than the first 48 hours of a new campaign.
The FTC angle: claims in ad copy carry the same exposure as claims in content
Advertising copy referencing customer outcomes, results, or reviews falls under the same regulatory framework as any other marketing claim. The FTC's guidance on endorsements and testimonials, most recently updated in 2023, requires that claims be truthful, substantiated, and not misleading — a standard that applies whether the copy was written by a person or generated by a model, and one that a fabricated statistic or an invented customer quote violates identically either way.
A prompt-level safeguard: never ask a model to generate a customer testimonial, a specific results claim ("increased sales by 340%"), or a review excerpt unless real, verifiable source material for that exact claim is attached. If no such source exists, the honest instruction is to omit the claim rather than have the model produce a plausible-sounding placeholder that later gets copied into a live ad.
Shopping and Performance Max: different inputs, same discipline
Shopping campaigns and Performance Max asset groups pull most of their creative from a product feed rather than from headline copy alone, which shifts where AI genuinely helps. A model cannot generate accurate product titles, prices, or availability — that data has to come from the actual feed — but it can help write the supporting short and long headlines, descriptions, and callouts that Performance Max assembles alongside feed data across Search, Display, YouTube, and Discover placements.
The business name constraint matters because Performance Max requires an exact match to the advertiser's verified business identity — a model given no constraint will happily write a more marketing-friendly variant that fails verification at upload.
A finished prompt that pulls the pieces together
Every field in this brief maps to a specific failure mode covered above — the character limits prevent disapproval on length, the forbidden list prevents policy disapproval on unsubstantiated claims, and the angle requirement prevents a low Ad Strength score from redundant headlines.
What to check before any AI-drafted ad set goes live
Four checks, in order. First, count characters in an actual character counter for every headline and description — do not trust the model's self-reported counts. Second, search-check each superlative or claim against Google's current advertising policies page, since policy specifics are updated periodically and a claim acceptable last year may not clear review today.
Third, confirm any statistic or customer reference traces to real, documented data, applying the same standard the FTC's endorsement guidance sets for any advertising claim. Fourth, after upload, watch Google Ads' own policy status for each ad — a disapproval notice identifies the specific violated policy, which is faster and more reliable feedback than guessing in advance which of dozens of policy provisions might apply.
A fifth check applies specifically to batches: when generating headlines or descriptions for several campaigns at once, spot-check a random sample rather than only the first few outputs. A model's early outputs in a long generation often follow the prompt more closely than later ones, where drift toward a default pattern can creep back in across a large batch — a five-minute sample check catches this before an entire campaign's worth of ads inherits the same drifted phrasing.
When AI-assisted ad copy is the wrong tool
Regulated industries — financial services, healthcare, legal, insurance — carry advertising rules well beyond Google's own policies, often at the state or national regulator level, and ad copy in these categories should go through the same compliance review a human-written ad would require, regardless of which tool drafted it. A model has no visibility into your specific regulatory obligations unless you supply them explicitly, and even then the review belongs with a qualified compliance function, not with the drafting prompt.
Similarly, any ad making a specific medical, financial, or safety claim needs the underlying substantiation to exist and be reviewed before the copy is written, not discovered after the fact when a regulator or a customer asks for it.
A related boundary: AI can help draft the marketing angle for a regulated product, but the required disclosures — interest rates, APR, risk warnings, licensing numbers — should be supplied to the model as fixed, non-negotiable text to include verbatim rather than left to the model's own phrasing, since a paraphrased disclosure can drift from the exact legally required wording without anyone noticing until an audit catches it.
Quick answers
How many headlines and descriptions does a Google RSA need?
Up to 15 headlines at 30 characters each and up to 4 descriptions at 90 characters each, counted by character including spaces and punctuation, not by word. Google recommends providing at least 8-10 headlines and all 4 descriptions to give its optimization system enough material to test combinations effectively.
Why do my AI-generated headlines keep getting disapproved?
Usually one of two reasons: the character count exceeds 30 despite looking short, or the copy contains an unsubstantiated claim — "best," "guaranteed," "#1" — that violates Google's Misrepresentation and Unreliable Claims policies. Explicitly forbid these words in the prompt and verify character counts in an actual counter rather than trusting the model's own count.
Can AI generate negative keywords without a search terms report?
Not reliably. A model has no visibility into which actual queries triggered your ads, so negative keyword suggestions without a real search terms report pasted in are guesses. Export the report from Google Ads, paste the actual query and performance data, and ask the model to find patterns in queries with clicks but no conversions.
Should I let AI write customer testimonials for my ads?
No, unless real, verifiable source material for that exact testimonial exists and is attached. The FTC's endorsement guidance treats a fabricated testimonial the same as any other false advertising claim, and a model asked to generate one will produce something plausible-sounding and fictional. Omit the claim entirely if no real source exists.
Does pinning headlines improve ad performance?
Usually not, and it often hurts it, because pinning disables Google's automatic testing and optimization for that position. Google's own guidance recommends pinning only for content with a genuine legal or brand requirement to always appear, such as a required disclaimer — not for a headline you simply believe performs best.
How long should I wait before judging which ad variant is winning?
Long enough for statistically meaningful impression volume to accumulate — commonly a couple of weeks at minimum, longer for lower-traffic campaigns. Google Ads will not confidently report a winning combination until enough data exists; declaring a winner from a 48-hour click-through-rate difference risks discarding a variant that would have performed equally well over a longer window.
Frequently asked questions
Does Google know or care if my ad copy was written by AI?
Google's ad review process evaluates the copy itself against its advertising policies — accuracy, substantiation of claims, prohibited content categories — not the tool used to produce it. There is no policy penalty specifically for AI-generated ad copy, and no disclosure requirement analogous to some content-labeling proposals discussed for organic search results. The practical implication is that the review bar is identical either way: an AI-drafted headline claiming "guaranteed results" fails the same Misrepresentation and Unreliable Claims policy that a human-written one would fail, and a well-substantiated, specific AI-drafted claim clears review the same way a well-substantiated human-written one would. The tool is invisible to the policy; only the claim matters.
What is the biggest mistake people make prompting AI for Google Ads copy?
Asking for "compelling" or "persuasive" copy without constraints, which reliably produces exactly the unsubstantiated superlatives — "best," "top-rated," "guaranteed" — that trigger policy disapproval. The second most common mistake is not specifying the character limit as a number, which produces headlines that read as appropriately short but measure 33 or 36 characters once counted, failing at upload. Both mistakes share a root cause: treating "write good ad copy" as sufficient instruction when Google Ads is a genuinely constrained format with hard character limits and an actively enforced policy framework. The fix in both cases is the same six-part brief structure used elsewhere — state the limit as a number, name the forbidden claims explicitly, and supply the real offer rather than asking the model to imagine one.
Can AI help me figure out which keywords to bid on?
AI can help you think through keyword categories and generate a broad candidate list to start from, but it cannot supply real search volume, competition level, or cost-per-click estimates — those require Google's own Keyword Planner or a third-party tool with live bid data. A reasonable workflow uses AI to brainstorm keyword variations and group them by likely intent, then feeds the resulting list into Keyword Planner to get real volume and cost estimates before committing budget, rather than trusting a model's guess at what a keyword costs or how often it is searched.
Is it safe to let AI write ad copy for a regulated industry like insurance or finance?
Only with much heavier human review than a typical consumer product ad requires. Regulated industries carry compliance obligations — specific disclosures, licensing statements, restrictions on performance claims — that sit on top of Google's own advertising policies and vary by jurisdiction and product type. A model has no built-in awareness of your specific regulatory requirements unless you supply them explicitly in the prompt, and even a well-constrained prompt does not substitute for a compliance reviewer's sign-off before a regulated-industry ad goes live. Treat AI-drafted copy in these categories as a faster first draft for the compliance reviewer to work from, not as review-ready output.
How do I stop AI from writing the same headline idea 15 different ways?
Name the distinct angles you want covered explicitly in the prompt — price, a specific feature, urgency, social proof, direct call to action — and cap how many headlines can cover any one angle, such as "no more than 3 headlines per angle." Without this constraint, a model asked for 15 headlines about one product will often produce 15 close paraphrases of the same central selling point, because that point is genuinely the strongest one available and the model has no instruction pushing it toward variety. After generating, run a follow-up prompt asking the model to group its own output by underlying message and flag over-represented groups, which catches redundancy the original generation pass missed.
Should I run AI-generated ad copy past someone before it goes live?
Yes, for every batch, and the review does not need to take long once the checklist is fixed. A five-minute pass covers four things: actual character counts in a real counter rather than the model's self-reported figures, a scan for any of the forbidden superlatives that may have slipped through, confirmation that any statistic or customer reference traces to real data rather than a plausible invention, and a check that sitelinks or callouts reference pages that genuinely exist. Teams that skip this step tend to discover the gaps only after a disapproval notice or, worse, after an ad has run for weeks making a claim nobody can substantiate. The five minutes is cheap compared to either outcome, and it is the same review discipline that applies to any AI-assisted business writing, not something unique to advertising.
Sources
- About Responsive Search Ads — Google Ads Help, 2025.
- Misrepresentation Policy — Google Ads Policy Help, 2025.
- About Ad Strength — Google Ads Help, 2025.
- Pin Assets in Responsive Search Ads — Google Ads Help, 2025.
- Guides Concerning the Use of Endorsements and Testimonials in Advertising — US Federal Trade Commission, 2023.
Every figure on this page names its source and year in the sentence that uses it. Where no methodology was published, the claim is stated qualitatively instead of dressed up as data.
Part 04 covers advertising in full
The A-Z AI Prompt Encyclopedia’s advertising and conversion part spans PPC strategy, Google search ads, Shopping, display, Facebook, Instagram, TikTok, LinkedIn, YouTube and landing pages — 300 prompts, 10 chapters. Ebook $12.99, paperback $38.99.
Get the bookPublished 2026-09-19 · Last reviewed 2026-09-19 by Mark W. Lamplugh Jr., author of the A-Z AI Prompt Encyclopedia.