ChatGPT ads: write them like search copy or like a feed ad?

I'm trying to work out how to think about ChatGPT ads before putting real money behind them. My working guess is that they're closer to search than to Meta. The user has just typed a question and the ad sits next to an answer they're already reading, so a generic "Best CRM, 20% off today" will probably feel like an interruption. I'd lean toward copy that names the specific problem behind the question, with a landing page that picks up the same thread.

What I'm less sure about is how much control we get over what triggers an ad. If matching is loose, a keyword-style list might not mean much, and the first few weeks could mostly just tell us which questions show up. Has anyone run these yet and seen whether the intent holds, or is it still more of an awareness play at this stage?

Elena, I'd split this into two questions: does the intent hold up, and can you even see it? Early on, the second is the harder one. The first numbers you get will probably be clicks and cost, not a read on whether those users were further along than a search visitor.

So before spending, decide what counts as intent for you. Pick something a few steps deep, like a pricing page view or a demo start, and put a dedicated UTM on the destination URL. Don't lean on referrer data from a chat interface, since it can be missing or inconsistent. Then compare those micro-conversion rates against your search traffic landing on the same page, not against the site average. And if you only look at the first week, you'll mostly be measuring volume, which is too thin to say anything about intent.

Agreed on the UTM, but I'm not sure the baseline holds. Search visitors on a page usually come in on a short keyword, so they're often early in their research. Someone who reaches a ChatGPT ad has typed a full question, something like "which CRM works for a 12-person agency that bills hourly." That person is more specific on average, so beating search on the same page doesn't prove much, and losing to it doesn't either. A fairer comparison is search traffic that came in on similarly long queries, and you can roughly isolate that in Search Console by filtering for longer phrases.

The other thing that tends to break these numbers is the form. Plenty of landing pages drop the UTM by the time someone submits a demo request, so the lead lands in the CRM with no source at all and the chat traffic quietly disappears from your reporting. Does your CRM keep the first-touch source, or only what the form sends?

Search Console only shows queries Google decides to report, and anonymized long-tail queries get dropped. So your long-query slice will look smaller than the real traffic. Treat it as a rough baseline, and say that when you compare numbers.

The anonymization is a privacy threshold, so it drops the rarest queries, and those are the long ones you care about. If you're already running a paid search campaign, the search terms report shows the actual query strings with cost and conversions, which makes it a better baseline than Search Console for this comparison.

Fair, the search terms report is the better baseline if you already run search. What I'd do is export it, bucket queries by word count, and compare only the 7-plus word group. That gives you a like-for-like pool with real cost and conversions, instead of an average dragged down by short terms like "crm software."

One catch is that Google hides very low-volume queries in that report, so the longest tail will look thinner than it really is. Put that in your notes next to the numbers.