ChatGPT Ads: judge them on click cost or on what happens after the chat?

I'm still not sure how to read ChatGPT Ads, and I don't think the usual yardsticks fit. My guess is the clicks behave more like search than social, since the person has just asked a question and has some intent. That would make a CTR or CPC comparison against Meta pretty much beside the point. But last-click conversions won't tell the full story either, because someone can research in the chat, leave, and come back later through branded search or a direct visit. Take a dental clinic running a small test. They'd need to watch branded search volume and direct bookings in the weeks after the ads run, not just what the ad platform reports. Has anyone found a clean way to separate those effects yet, or is everyone just trusting the platform's numbers for now?

A geo holdout is the cleanest way to separate this. Run ads in a few cities, keep similar ones dark, then compare branded search and direct bookings against the pre-test baseline. It's crude, but it beats trusting last-click.

A geo holdout is the right idea, but for a single dental clinic the numbers will probably be too small to read. If each city books a handful of patients a week, one busy Monday can move the comparison more than the ads do. Before launch, write down the cost per booking you'd accept and the minimum bookings per city you need to call it. If the test can't hit that count within its window, treat the result as inconclusive and either extend it or shrink the scope. There are no ChatGPT benchmarks yet, so the only threshold you can trust is the one you set before you see the data.

Agreed, and I'd lock in a lag window before launch too. Dental patients often research for a week or two before they book, so if you only count bookings made during the flight, the early ads get undercounted and whatever happened before the test leaks in. Decide now whether you count bookings by the date they were made or by the date of the first visit, and write it down next to the cost per booking. Otherwise you'll end up picking whichever version looks better once the numbers are in.

Locking the window makes sense. I'd also be careful with chat history, since it can make the lag look worse than it is. A patient who asked about implants on Monday might come back a month later in a brand new session and never mention the ad. Count those as assisted at best, not as proof.

I'd decide scale or stop on cost per booking from a source the ad platform doesn't control, and for a clinic that means the booking form. Add a required "how did you hear about us" field with an option for "asked an AI assistant," and have the front desk log the same answer for phone calls. The answers will be messy because people misremember, but that's exactly the path Theo describes: someone asks in ChatGPT, leaves, and comes back through Google a month later. The platform only shows the click, which is the part you can trust least here. If the self-reported count and the platform count disagree, use the self-reported one for the decision and the platform one only to diagnose which ads are working.

Agreed on the form, but keep the self-reported answer separate from the tracking data. Store the UTM parameters and any click ID as hidden fields, so a booking that says "Google" but arrived with a Microsoft click ID still shows up as a mismatch. Those mismatches are useful, because they catch people who clicked an ad and then remembered a different name a week later.

For the self-reported side, put a short free-text line under the dropdown, something like "what did you ask?" A patient who writes "asked an AI about implant costs" tells you more than a checkbox does, and you can sort the phrasing later. Keep the dropdown itself short, though, or people will skip it.

The mismatch check only works if the parameters actually reach the form. A lot of clinics use embedded third-party booking widgets, and those often drop query strings on the way through. If the UTMs never arrive, every booking looks like a mismatch and you end up doubting patients who were telling the truth.

So before launch, click your own ad on a phone, go through the whole booking flow, and confirm the hidden fields carry the right values. And keep a separate column for each click ID type, since a gclid tells you nothing about a msclkid or ttclid.