Ad copy from an AI: feed it the landing page or a short product brief?

How do you all handle source material when asking an AI to draft ads? I see two camps. One pastes the whole landing page and says "match this." The other writes a short brief: offer, price, who it's for, two claims they can actually back up, and words to avoid.

My view is the brief wins, mostly because a landing page is full of stuff that isn't ad-worthy: nav text, legal lines, three value props competing with each other. The model then tends to grab whatever sounds most polished, which is often a feature list instead of the thing a searcher actually cares about.

Where I'm less sure is volume. For a store with 200 SKUs, a brief per product is real work, so pasting the page may just be the pragmatic choice. Does anyone have data on whether the page-dump version holds up on click-through?

I don't have click-through data comparing page dumps with briefs, so I'd be wary of any clean number people quote. The failure you describe is the one I'd bet on, and it gets worse at 200 SKUs than at 20. The model doesn't pick the strongest claim, it picks the most polished sentence, and product pages are polished in all the wrong places.

The brief doesn't have to be written by hand, though. Paste each page once, ask the model to fill your template and list the claims it found along with the line each one came from, then cut anything that isn't a reason someone would search for the product. That turns a blank-page job into a quick review.

The step people skip is checking those claims against actual search terms. A feature can be true and still be a phrase nobody types.

Checking claims against search terms is right, but I'd pull the search term report for the category before writing the brief instead of after. It shows the exact words people use for the problem, like "waterproof running jacket" versus "rain jacket for running", and those phrases should drive the headline, not the product name on the page.

Jun, the search term report only shows queries that already matched your ads, so it mostly reflects your current targeting and match types, not everything people actually type. Google also hides very low-volume terms, which means some of the long-tail phrasing you're after may never show up.

Say a running shoe store bids only on "running shoes" in phrase match. The report will be full of "running shoes for flat feet" variants, but you won't see anything like "trail shoes that drain fast," because that never matched. For a category you aren't running yet, Keyword Planner can show volume for phrasings the report can't. I'd use both: the report for the words people already use with you, and Planner for the ones they'd type if you showed up.

The zero is the part I'd watch. Keyword Planner often shows nothing or a tiny range for the long-tail phrasings you most want, like "trail shoes that drain fast." That usually means there isn't enough data to measure, not that nobody searches for it. Planner volumes are also averages, and they can swing a lot for seasonal products.

So for a category you aren't running yet, test those phrases in exact or phrase match with a small daily budget, and look at impressions and cost per click over a couple of weeks before writing a phrase off. If the category runs in Shopping or Performance Max, the search terms view is thinner, so put the test in a campaign where you can actually see the queries.

Iris, when Planner comes back empty, I'd pull the claim from complaints instead of the product page. Reviews, support tickets and comments under competitor posts show how people describe the problem in their own words, and that's often where a phrase like "drains fast" comes from. Rule of thumb: if the same problem shows up in three or more of those places, it goes in the brief, even if Planner shows nothing for it.

Those people already had the problem and said so out loud, so the claim is tied to something real rather than whatever sounds most polished. It also gives you copy that reads like a person wrote it, which matters on short-video placements where nobody reads a headline twice.