Three ads. Three arguments. One page.

The formats are not a list somebody wrote down. They are derived from 20,851 Meta ads scraped from a swipe file, grouped by category, and ranked by how many real advertisers use each one — so "benefit-driven" leads with 13,881 examples behind it and "gamified" sits at the bottom with one.

When a product page is analysed, its category is matched against that library and the generator is shown how ads in that category are actually composed. A clothing brand and a supplement brand get different formats because their categories use different ones, not because the model felt like varying.

One limit worth stating plainly: the library carries no ad copy. Every record has a visual description and a format, and the text field is null across all 20,851. So it can teach the generator what winning ads look like and how they are built — never how they are written. The words come from your page.

Every generating call carries the same rule, and it is deliberately open-ended rather than a list of banned words: state no price, discount, saving, guarantee, return window, delivery promise, percentage, certification or comparative claim that the source page does not support.

Price is the strictest case. It is taken only from structured product data the page publishes about itself — never from a figure matched somewhere in the HTML, because a number on a page is as likely to be a crossed-out anchor, a shipping threshold or an unrelated product as it is to be the price.

This is not a content-safety feature. An ad that invents a discount costs more than it earns the moment a customer holds you to it, and the platform that served it is not the one who has to honour it.