AI Ad Writers Invent Customer Numbers and Star Ratings. Here Are the Exact Phrases
Ask any language model for ad copy and a proportion of what comes back will contain a number nobody counted. Not occasionally — reliably, in a small set of recognisable shapes.
We block 37 of those shapes on every variant we generate. This page lists the ones that actually appear, because knowing the phrasing is what lets you catch it in whatever tool you use.
Why models do this
An ad headline has a slot in it where social proof usually goes. Trained on millions of real ads, a model learns that the slot is normally filled with a number and a happy noun, so it fills it. It is not lying in any meaningful sense — it is completing a pattern, and the pattern says "put a number here".
Get 10 scroll-stopping ad headlines instantly — no signup.
Use the tool free →The result is copy that reads professionally and asserts something no one at your company has ever measured.
The shapes that actually appear
Counted customers. "Join 1,000+ happy customers." "Trusted by 10,000 stores." "Loved by 2,500+ shoppers." The most common by a wide margin.
The noun is not fixed, which is what makes this hard to catch by keyword. A real generation for a dog seat cover produced "Over 10,000 Happy Pups and Their Owners Trust Our Waterproof Car Seat Cover" — the noun was "Pups". A filter looking for "customers" or "shoppers" sails straight past it. It has to match the shape: a large number next to a trust-or-love verb.
Star ratings and review counts. "Rated 4.9 stars." "10,000+ five-star reviews." "98% of buyers came back."
Superlatives presented as fact. "Best-selling mug of 2026." "#1 rated." "Top-rated." "World's best."
Press mentions. "As seen in Forbes." "Featured in TechCrunch." Invented wholesale.
First-person testimonials. "I switched and saved $400 a month." "I used to go through three bottles a week." The model writes as a fictional customer. Sometimes it adds quotation marks, which makes it worse rather than better — now it is an attributed quote from a person who does not exist.
Uncited attribution. "One reviewer said…" "Customers say…" "Real talk from real customers."
Manufactured scarcity and momentum. "Sold out 5 times in 2 weeks." "Going viral on TikTok." "Limited to the first 100 buyers."
Dollar outcomes. "Saved me $400." "Replace your $2,000/month agency." "Made $10,000 in 30 days."
Why this is your problem, not the model's
If you run the ad, you made the claim. In the US the FTC expects advertisers to be able to substantiate objective claims, and most other markets have an equivalent. "The AI wrote it" is not a defence anyone has had success with.
The practical risks, in the order they are likely to bite:
- Platform rejection. Meta and Google both review claims, and unsupportable ones get disapproved. That is the cheap outcome — you find out immediately.
- Review-bombing and complaints. Customers who see "rated 4.9 stars" and then find your store has eleven reviews notice, and some of them say so publicly.
- Regulatory attention. Rare for a small store, expensive when it happens.
The insidious one is the first case where nothing happens at all. The ad runs, nobody complains, and you conclude the copy is fine — so the next fifty ads get generated the same way.
How to catch it
Read every generated variant once before it runs. Slow, and it is the only method that works completely.
Search for the shape, not the words. A number of three or more digits sitting near "happy", "trust", "loved", "joined", "switched", or a star rating, is worth a second look regardless of the noun.
Prompt against it, then verify anyway. Telling the model "do not invent customer counts, ratings or press mentions" reduces the rate substantially. It does not reach zero. We use both: instructions in the prompt as the primary defence, and pattern matching as a backstop, because each catches what the other misses.
Give the model real numbers to use. The most effective single fix. A model with nothing true to say reaches for something plausible; a model given "made in Sheffield since 2019, 400 sold last year" will use those instead, and they are yours to stand behind.
What honest ad copy uses instead
Specific product facts, which are more persuasive than invented numbers anyway. "42oz removable tank" beats "loved by thousands". "No magnets, no screws" beats "the #1 choice". A concrete detail is checkable, which is exactly why it is credible.
If you genuinely have the social proof, use it — a real review count is one of the strongest things you can put in an ad. The problem is only ever the invented version.
Frequently asked questions
Does this happen with every AI ad tool?
It happens with every underlying language model, so it happens with every tool that does not filter for it. Whether your tool filters is worth asking directly.
Can I just tell it not to make things up?
It helps a lot and it does not finish the job. Instructions reduce the rate; they do not eliminate it, which is why we run pattern checks over the output as well.
What if the claim is actually true?
Then use it, and put the real figure in the product description so the model works from your number rather than inventing one. That is the difference between "trusted by thousands" and "412 sold since March".
How do I know if my current ads have this problem?
Read them for any number or superlative and ask whether you could produce evidence for it today. If you cannot, remove it. It takes ten minutes and it is the highest-value ten minutes in this whole subject.
Does AdCreator block all of it?
We block 37 documented patterns and instruct against the rest, and we still tell merchants to read their copy. Anyone claiming a filter catches everything is making an unsubstantiated claim about unsubstantiated claims.