AI ad copy has a tell. Three adjectives where one would do, a rhetorical question in the first line, the word “elevate,” and a close that promises to “take your business to the next level.” Readers skim past it. So do the auction’s quality signals, because nobody clicks.
The problem is almost never the model. It is the brief. Give a model three lines of context and it fills the gap with the average of everything it has ever read — which is exactly what generic sounds like. Give it your actual search terms, your actual objections and your actual prices, and it writes like someone who works at your company.
Here is the workflow we use to get ad copy out of AI that a human would have been proud to write.
1. Feed it evidence, not adjectives
Before you open a chat window, collect four things:
- Your last 90 days of search terms. The literal phrases people typed. Not your keyword list — the search terms report. If you have never pulled one, start with the search terms report and where your budget actually goes.
- Five real customer quotes. Review text, sales call notes, a support ticket. Unedited.
- Your three most common objections. Price, timeline, “do you serve my area,” whatever it really is.
- One number you can prove. A price, a turnaround time, a warranty length, a count of customers.
That last one matters more than it looks. A specific number is the single fastest way to make a line sound like it came from a business rather than a language model. “Installed in 4 days” beats “fast, reliable installation” every time.
2. Write the brief in the customer’s words
A brief that produces usable copy looks roughly like this:
“We are a [what you do] in [where]. People find us by searching [three real search terms]. They usually worry about [objection]. What we can prove: [number]. Customers describe us as [two quotes]. Write 15 Google Ads headlines, max 30 characters, no exclamation marks, no words from this banned list: elevate, unlock, seamless, game-changing, journey, revolutionise. Every headline must contain either a search term or a number.”
The banned list does most of the work. Those words are the ones models reach for when they have nothing concrete to say, so blocking them forces the model back to your evidence.
3. Generate wide, then cut hard
Ask for 15 headlines and 6 descriptions, expect to keep 4 and 2. Generating in bulk is cheap; the value is in having enough variance to spot the two lines that sound like a person.
Cut anything that survives the “any competitor could run this” test. If a headline would work unchanged for the business down the road, it is not copy, it is filler. That single test removes most AI output.
4. Do the human editing pass
Four edits, in this order:
- Shorten the first four words. Models front-load throat-clearing. The point is usually the second clause.
- Swap one abstraction for one noun. “Better results” becomes “37 more booked calls.”
- Break one rule of grammar. A sentence fragment. It reads human because it is how people talk.
- Read it out loud. If you would not say it to a customer standing in front of you, rewrite it.
5. Match the landing page, or none of it matters
Ad copy that promises a 4-day install and lands on a page whose headline says “Your Trusted Partner in Excellence” will underperform no matter how good the ad is. The headline on the page should repeat the promise in the ad, in the same words. Our landing page checklist for paid ads covers the rest of the match.
6. Let the data pick the winner, not your taste
Run four headlines per ad group, give each one enough impressions to be meaningful, and judge on conversions rather than CTR. A high-CTR headline that attracts the wrong searcher costs you twice — once in clicks, once in the sales time spent disqualifying them. If your conversion reporting is thin, the GA4 reports that actually matter for a small business is the shortest route to fixing that.
What AI is genuinely good at here
Three things, reliably:
- Volume under constraints. Thirty headlines under 30 characters, each containing a specific phrase, in nine seconds. No human wants that job.
- Translation between registers. Taking one good line and producing a formal version, a blunt version and a friendly version so you can test tone.
- Finding the objection you missed. Paste 40 reviews and ask what customers hesitate about. The answer is often not what you assumed.
What it is not good at is knowing which of its own outputs is true. Every number, claim and guarantee needs a human to verify it before it goes live — a fabricated “trusted by 5,000 businesses” is a compliance problem, not a copy problem.
The ten-minute version
- Paste your search terms report, five reviews and one provable number.
- Ask for 15 headlines with a banned-words list.
- Delete anything a competitor could run.
- Edit four lines by hand.
- Make the landing page headline repeat the ad’s promise.
- Judge on conversions, not clicks.
AI did not make copywriting easier. It made the briefing the whole job. If you want a wider view of where else that shift is showing up, read what has actually changed in AI and digital marketing in 2026, and the ten AI tools worth a small business’s time.
