AI writes competent ad copy in seconds. It also inserts a legal problem into roughly half of what it produces, and does so confidently enough that the problem is easy to miss. Not through carelessness — the specific patterns advertising regulators restrict are exactly what high-converting ad copy looks like in the training data. Here is how to brief it, and what to delete before anything runs.
The brief that produces usable output
Most bad AI copy is the result of a bad brief. Four things change the output more than any prompt technique:
Who the customer is
Not "small businesses" — "salon owners who do most of their bookings on WhatsApp and have no website". Specificity here produces specificity in the output.
Why they choose you
The real reason. Evening appointments, you answer the phone, the quote includes parts. This is the thing only you know and the thing that makes an ad work.
What the offer actually is
A price, a guarantee, a first-visit discount, a free consultation. Vague offers produce vague copy, and there is no way around that.
The destination
Copy for a WhatsApp chat is different from copy for a product page. Tell it where the click lands.
💡 Two honest sentences about who your customer is and why they pick you will improve the output more than any amount of switching between tools. Most people write something generic, get generic copy back, and conclude AI cannot write.
The five things to delete every time
1. Invented statistics
"Trusted by 5,000+ customers." "Saves an average of 40%." "Rated 4.9." None of these came from your records — they were generated to fit the sentence rhythm. Every one is a claim you must be able to evidence, and several are straightforwardly misleading advertising if you cannot.
2. Superlatives
"The best", "the leading", "number one". These are objective claims about market position, not figures of speech. The fix is not to soften them into vagueness but to replace them with something checkable — "Gas Safe registered, same-day callout" rather than "the best plumber in Leeds".
3. Copy that describes the reader
This one gets you rejected before anyone regulatory sees it. Meta prohibits implying you know something personal about the viewer, and AI addresses the reader directly by default.
Rejected
"Are you struggling with debt?" · "Tired of back pain?" · "Overweight?"
Accepted
"Our programme helps people manage debt." · "Treatment for back pain." · "Weight management support."
4. Outcome guarantees
"Guaranteed results", "permanent", "risk-free", "double your revenue". Heavily restricted in health, finance and anything making earnings claims, and rejected automatically by both major platforms at a threshold neither publishes.
5. Testimonial framing
Copy written as though a customer is speaking, when no such customer exists. This extends to AI video: an avatar recounting a personal result reads as a testimonial to any reasonable viewer. Never publish a review, rating, testimonial or customer name that is not real.
The rule that handles all of it
💡 No claim, figure or superlative goes into an ad unless you could produce the evidence within an hour. Applied consistently that single habit removes most of the regulatory risk AI copy introduces — and costs you nothing, because specific verifiable claims outperform superlatives anyway. Vagueness is what underperforms, not honesty.
Translation is not localisation
If you advertise in more than one language, this is worth more than everything above.
Copy translated from English reads as translated, because English ad writing leans on compression and wordplay that does not survive the trip. "Book now, pay later" translates cleanly. "Stop guessing. Start selling." becomes two flat imperatives with none of the rhythm.
There is also a register problem. Formal register in many languages reads as official communication rather than as a business talking to a customer, and a literal translator will reliably choose the formal option. Give the model the offer and the audience and have it write in the target language from scratch — a tool offering only a "translate" button is doing the weaker of the two jobs.
The headline compression trap
Search ad headlines have tight character limits, and compression pushes AI straight toward the words that cause problems — "best", "cheapest", "#1", "guaranteed" all fit neatly where a specific claim does not.
Worth watching for specifically, because the same model that writes carefully at 125 characters will reach for a superlative at 30.
What AI is genuinely better at than you
Volume and variation. A person writes two headlines and gets attached to them. A model writes twenty in a minute, and you pick the three worth testing. That is a real advantage and the main reason to use it.
It is also better at rewriting to a constraint — "say this in 30 characters", "make this sound less formal", "rewrite without describing the reader". Those are mechanical transformations and it does them instantly.
It is not better than you at knowing what to say. That part has not moved.
Copy written for the market, not translated into it
Adyft writes ad copy from your offer and audience in the language of the market, and shows you every headline before it publishes. 14-day free trial, no card.
Start free — no card needed