Paste a product URL and you get a brief: the name, a summary, the selling points, the price and the photography. It is the fastest route from a link to a lot you can actually render, because the photography comes with it.
It reads. It does not write.
No model is called.
The extractor reads structured data the page publishes about itself, schema.org markup, OpenGraph tags, meta tags. That is the whole mechanism, and it is the important property: a page carrying nothing structured returns a thin brief rather than invented copy.
A language model pointed at the same page would always produce a full, confident, well-written brief, including for a page that said almost nothing, where every specific in it would be a guess wearing marketing voice. You would then build an ad on it, render the ad, and find out at the review stage that the product does not do the thing your hook promised.
So a thin result here is information, not a failure. The brief names its own sources, which is how you tell “this page was uninformative” from “this reader was”. When it comes back thin, write the brief with the person who knows the product and pass it in directly. That takes five minutes and is correct.
Two failures worth recognising rather than retrying
| What happened | What to do |
|---|---|
| The site refused the reader. Many large storefronts block automated requests | Do not retry and do not work around it. Write the brief by hand and attach the images yourself |
| The logo would not download. Named in the result, and the brand kit is still made | Nothing. This is not a failed call, add the mark later if you want one |
Save it, so the next ad does not start from blank
An extracted brief is a reading of a page at a moment. Saving it as a product turns it into something the workspace owns.
The part that matters most is the photography. Attached images are downloaded once and then reused by reference, so an ad still builds after the source page has been redesigned, moved or taken down. That is the whole reason to attach them rather than pass URLs each time you build something.
Image URLs that fail are named individually rather than swallowed, so a product where three of five photos landed tells you which two did not. Read that list, a product saved with no usable image is the thing a vertical will refuse a lot for, and finding it here costs nothing.
Why the photography is the point
A prompt that says “the bottle” renders a different bottle in every shot. Nothing about writing the prompt more carefully fixes that, because the model has never seen your bottle.
A saved product’s photography is attached to each shot as a reference pack, and that is what holds one product across four clips. It is the difference between an ad about a category and an ad about your SKU.
Check the reference count on a shot rather than assuming it attached. An empty error list is not proof that anything landed, and a render that silently carried no references succeeds, looks fine, and does not contain your product.
The same reading, for a brand
Pointing the same mechanism at a brand’s site gives you the mark, the brand name and the one colour the site states about itself. It is the fastest way to a brand kit, and it has exactly the same honesty property: it reports what the site published, not what a model thinks a brand like that would look like.