AI Facebook ad copy: how to use ChatGPT and Claude for ad texts that bring leads

How long Facebook ad copy should be

Meta’s help center gives one rule of thumb — 125 characters of primary text, 40 for the headline, 25 for the description — and warns that text in ads may be further truncated across various placements and devices. (source) Its Ads Guide is more specific per placement:

Tell the assistant your shortest placement, not a generic “125”, and ask that the first sentence work on its own: on most placements the rest is folded or cut. With Advantage+ placements, that usually means writing for Reels first.

Use AI as a critic, not a ghostwriter

The most useful study on this measured ad quality by real clicks on social platforms. Its finding:

Using LLMs as sounding boards enhances the quality of the resultant ad copies for nonexperts

The other half matters as much: using the model as a ghostwriter did not provide significant benefits and is, in fact, detrimental to expert users, because it produces an anchoring effect, which leads to lower-quality ads. If you know your customers, write the draft yourself and ask the assistant to attack it — with your numbers in front of it.

Prompt: critique my draft against my results

Below: my 3 best and 3 worst ads from the last 90 days with spend, leads and cost per lead, and the landing page text.
Then my new draft.
1. Say what the best ads do that the worst don't (offer, first line, proof, call to action).
2. Judge my draft against that. Quote the weak phrases and say why.
3. Suggest at most 2 edits per weak phrase. Keep my wording where it already works.
Don't rewrite the whole text.

Anthropic’s guidance supports this setup: put the long inputs first — Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples. (source) OpenAI says the same about examples: show a diverse range of possible inputs with the desired outputs. (source) Our prompt guide has the base prompt for new copy in the voice of your best ads and a policy check.

Five angles, not five rewordings

AI-assisted writing tends to converge. In a study in Science Advances, stories written with AI ideas were rated more creative individually but were more similar to each other (source). For ads that matters, because Meta now groups look-alike creative together and leans on variety: its ranking system is built for advertisers who upload more and more creatives to support a diversification strategy (source).

Prompt: angles from different facts

From the landing page below, list the facts a customer could care about: price or payment plan, results with numbers,
who teaches or treats, how it works step by step, timing and next start date, guarantees, reviews (quote only real ones).
Write one primary text for each of 5 different facts. Each must lead with its fact in the first sentence,
fit in 44 characters for Reels and in 125 for the feed, and make sense without the others.
Then tell me which two are most different from my current ads.

Each angle becomes its own ad, so you learn which reason to buy works — not which synonym.

Which text won? Check before you believe it

The usual way to pick a winner — sort texts by cost per lead across a campaign — mixes up the text with the audience it ran to. When we built a text comparison into Variafy, one headline looked 43% cheaper across its campaign; compared with the other texts in the same ad sets, the difference was within chance. The audience was winning, not the words.

A simple rule: put the new texts as separate ads in the same ad set as the current one and wait until each has at least 30 leads. Even then a difference of a quarter can be chance (how much chance moves cost per lead), so only act on clear gaps. If your CRM is connected, compare by customers, not by leads.

Does AI-written copy perform better?

Nobody has published lead or sales results for AI copy against human copy. What exists:

• Meta’s own generator. In a 10-week test across about 35,000 advertisers, Meta’s improved text model raised click-through rate by 6.7% (from 3.1% to 3.3%) — compared with its previous AI model, not with advertisers’ own texts, and trained on clicks (source).
• Small vendor tests report double-digit click gains for lightly edited AI copy, on a handful of ads each.
• Figures like “AI copy lifts CTR 15–25%” that circulate in blog posts have no published study behind them.

So treat AI copy as a faster way to produce testable variants, and let your own results decide. That is also how Anthropic says it uses Claude for its ads: copy written using existing campaign and ad performance data, checked against character limits and uploaded (after undergoing manual review) (source).

Rules AI copy breaks most often

• Personal attributes. “Struggling with debt?” or “Do you have diabetes?” is the default hook of every model, and Meta rejects it: ads can’t assert or imply a person’s health, finances or other personal attributes (source). “You” is fine; “you” plus a condition isn’t.
• Invented proof. The FTC’s 2024 rule bans testimonials that misrepresent that they are by someone who does not exist, such as AI-generated fake reviews (source). Tell the assistant to quote reviews only verbatim from what you paste.
• Numbers and offers that aren’t on the page. Discounts, deadlines and results must match the landing page. Ask for every number to be traced to your data or the page. For course and coaching ads, the separate guide to Facebook ads for courses and coaches covers the policy scope and the proof behind a learning promise.

Meta doesn’t require a label on AI-written text; its “AI info” label is for images and video. More on rejections in why was my Facebook ad rejected.

Copy briefs from your real results: Variafy

Variafy connects ChatGPT, Claude and other assistants to your Meta ad accounts. Ask “Write new ad texts for this campaign” and its copy brief gives the assistant what this guide asks for:

• every primary text and headline of the campaign with spend, link CTR, results and cost per result;
• which texts beat the others beyond chance — compared inside the same ad sets, with a stricter bar the more texts there are — and which haven’t proven anything yet;
• with payments sent in from your CRM or a sheet, the money each text brought;
• the landing page text and the length limits.

The assistant writes; new ads are created paused and run only after you confirm.

How we checked

Questions

Related guides

• ChatGPT and Claude prompts for Facebook ads that use your own numbers
• Why was my Facebook ad rejected — and why is it stuck in review?
• ChatGPT vs Claude for Meta ads: which one fits your ad account