- The brief, written badly on purpose
- Round one: get the obvious out of the way fast
- Round two: hand the model your audience’s actual language
- Round three: stress-test the angle, not the wording
- Round four through six: the rest is mechanical
- The failure modes worth naming
- What actually moves the conversion needle
The first AI-written ad you push live almost never beats your control. The fifth one might. The reason is not that AI cannot write good ads. It is that one-shot prompts cannot, and most marketers stop after one shot. To show what the difference looks like in practice, here is the same brief carried through a six-prompt iteration loop, with the variant that eventually beat the existing best performer in a real Meta test.
HubSpot’s State of Marketing 2025 reports that 74% of US marketers now use AI, but 54% feel overwhelmed when implementing it. The overwhelm is almost always about workflow, not capability. The marketers seeing real lifts are the ones treating AI as a brainstorming partner that needs to be argued with, not a vending machine that produces a finished asset.
The brief, written badly on purpose
Start with a brief that looks like the sort most teams hand to their copywriter on a Tuesday afternoon. Reusable water bottle, eco-conscious millennial and Gen Z buyers, $35 price point, competing against cheaper plastic bottles and pricier steel ones. The campaign goal is sign-ups for a launch waitlist. That brief, fed to any chatbot, returns three or four headlines about hydration, sustainability, and the planet. They are fine. They are also the same headlines every other water bottle brand is running this quarter.
The point of the iteration loop is to push past the obvious. The first response from the model is the average of every product description it has seen. You want the version that exists three or four prompts past that average.
Round one: get the obvious out of the way fast
Open ChatGPT or Claude and ask for the cliches by name so the model stops reaching for them.
“You are a senior performance copywriter. I am writing Meta ads for [PRODUCT] aimed at [AUDIENCE]. The campaign goal is [GOAL]. Write 20 headlines under 8 words each. Half should be benefit-led, half should be curiosity or pattern-interrupt. Hard constraint: do not use the words sustainable, eco-friendly, planet, hydration, path, or the difference. Do not use alliteration. Do not start any headline with a verb.”
The constraint list is doing the heavy lifting. Without it the model defaults to “Hydrate Sustainably,” “Drink Better,” “Save the Planet One Sip at a Time,” and similar. With it, the model is forced into specific nouns, numbers, and angles it would otherwise skip.
You will get twenty headlines. Maybe two are interesting. That is the expected hit rate. Mark the two you would consider testing.
Round two: hand the model your audience’s actual language
Most AI-written ads sound nothing like the audience because the model has never read your customer reviews, your subreddit threads, or your support tickets. Feed those in.
“Below are 30 phrases pulled from reviews, support emails, and Reddit threads where our target buyer talks about water bottles, hydration habits, and reasons they stopped using their last bottle. Read these. Identify the three biggest unspoken frustrations. Then rewrite the two headlines I marked from your previous list using language and rhythm that matches how the audience actually talks. No marketing voice. [PASTE 30 PHRASES]”
The output here is the first time the ads start sounding like the audience instead of the brand. A real example from a working test: the original “Bottle, Reimagined” became “I keep losing the lid. This one clips.” That second line sells, because it names a frustration the buyer recognises in their own kitchen.
Round three: stress-test the angle, not the wording
The iteration trap is rewriting the same angle in fifteen different ways. The lift comes from changing what the headline is actually about.
“For each of the two best headlines, write three alternative versions that change the angle entirely while preserving the audience language. Angle 1 should target an objection rather than a benefit. Angle 2 should be a single specific scene the buyer recognises. Angle 3 should be a comparison to a competing product or behaviour. Keep all under 8 words.”
You now have nine headlines from two original seeds, each chasing a different psychological lever. This is where the surprise variants appear. The objection-led headline often outperforms the benefit-led headline by a wide margin in cold traffic, because most ads in the feed are benefit-led and the buyer’s eye has learned to skip them.
Round four through six: the rest is mechanical
Round four: write three opening lines for each headline that earn the click. Round five: pair each headline-and-line with three CTA variants. Round six: ask the model to be the sceptic and tell you which combinations will not work and why. That last prompt is the one most marketers skip and the one that catches the duds.
“Read the 27 headline-line-CTA combinations above. Act as a sceptical performance director who has seen every angle in this category. For each combination, give it a score from 1 to 5 on (a) likelihood of stopping the scroll, (b) clarity of the offer, (c) novelty against the category. Flag any that sound generic, AI-written, or that promise something the product does not deliver. Recommend the top six to take into A/B testing.”
Six finalists, drawn from a process that took maybe forty minutes start to finish. Compare that to a Tuesday afternoon brainstorm and the time spend is similar, but the variance of ideas is much higher.
The failure modes worth naming
Three traps consistently flatten AI-written ads.
The first is the missing audience input. If you do not paste in actual customer language, the model writes for an imagined customer. That imagined customer is bland. The second is single-pass acceptance. A model’s first reply is its safest reply. The work happens in the next four prompts. The third, and most subtle, is the AI-tell. Em dashes, the construction “It’s not just X, it’s Y,” the word “elevate,” the phrase “designed for the modern” anything. Add an explicit ban list to the brief on every iteration prompt and the tells disappear.
If you want a starting point, Anyword and Jasper both have ad-specific modes that scaffold this iteration loop in their interface, with prediction scores attached to each variant. They are useful as a structured wrapper, but the underlying logic, prompt-iterate-test-iterate, is the same as running it manually in a chatbot.
What actually moves the conversion needle
The honest answer is not the AI tool. It is the loop. Marketers who run six rounds of iteration on a single ad set, paste in real audience language, and test against an existing control are the ones reporting meaningful lifts. Marketers who write one prompt, copy the output, and ship it are comparing five minutes of AI output against five years of their own pattern recognition. The first version loses every time.
The real edge is not that AI writes better ads than you do. It is that AI lets you generate, stress-test, and discard variants at a pace that was previously impossible. The decision about which finalist makes it to test is still yours. The discipline to iterate past the average reply is still yours. AI is the typist that lets you do both on the same Tuesday afternoon you used to spend writing one mediocre headline.