How do we know if ChatGPT actually sent us a customer?

Yes, you can know. You won't find it in the source field of your analytics tool, because AI engines usually strip the referrer and the buyer who arrives from one has often already made up their mind. You find it by joining three cheap signals: the first-touch referrer when one exists, a visitor ID that survives weeks of separate sessions, and one plain question on your booking form.

That join is the whole difference between a dashboard that reports "direct traffic" and a record that shows a buyer reading four pages across 24 days and booking a call five minutes after the pricing page. Same visitor either way. Only one version tells you whether the money you're putting into AI visibility is coming back.

Why doesn't our analytics tool just tell us?

Because an AI referral breaks the model analytics was built for. An engine reads your pages, summarizes them, and names you inside an answer. The buyer often never clicks the citation. They read the answer, remember the name, and type it into a browser a week or two later. Your tool logs a direct visit and credits the work that produced it with nothing.

We've written before about the gap between being read and being credited by AI search. Attribution inherits that same blind spot one layer further down the funnel. The engine reads far more than it links, and the buyer converts far later than the visit that convinced them.

The practical cost shows up in a budget meeting. Someone asks whether the content is working, and the most honest answer available is "traffic is up." No CFO funds a second year on that sentence. Measuring whether engines mention you at all is a separate job, and we covered the metric set for it in how to measure whether AI engines are recommending your company. What follows is the step after that one.

What does an AI-sourced lead actually look like?

Here's one of ours, from last week. A CEO at a healthcare AI company booked a call through our site. On the booking form, where we ask how someone first heard about us, he typed ChatGPT.

The record behind that booking: a first visit to the homepage on September 5, then nothing at all for three weeks. Two more visits on September 28. The pricing page on September 29, and a booked call five minutes and twelve seconds later.

Look at what that path breaks. Twenty-four days sat between first touch and booking. The buyer came back across four separate sessions, not one. And no engine passed a click on the session that actually converted. A referrer-based model scores that visit "direct" and hands the credit to nobody, which is how a channel that is working gets defunded.

One more thing fell out of it that had nothing to do with attribution. He'd read the price before he booked, so the call didn't open with an arm-wrestle about the number. It opened on whether we were a fit. Knowing which pages a buyer read before they raised their hand changes how you run the first fifteen minutes.

For scale, our own server log over a recent window held 529,453 hits and 1,190 humans sent by an AI engine. We walked through separating those groups in telling real buyers from bots. The humans are a rounding error next to the crawling, and they're the only line in that log that can turn into revenue.

How do we set this up without buying another platform?

Three moves, none of them expensive.

First, give every visitor a durable first-party ID and keep it across sessions. Without that, session four is a stranger to session one and a 24-day path collapses into a single anonymous hit.

Second, store the first-touch referrer and the first landing page against that ID, and never let a later visit overwrite them. Last-touch reporting is what turns a patient, AI-sourced buyer into "direct."

Third, ask on the booking form: how did you first hear about us? Leave it as free text. A dropdown hands people your categories and teaches you nothing you didn't already believe. Free text is where a buyer writes ChatGPT unprompted, which is the one signal that survives a stripped referrer.

Then join those three at the moment a call gets booked. That's the entire build. No new platform, no vendor call.

If you want the shape of the record, four fields carry almost all of the value: first seen, first referrer, first landing page, and the free-text source the buyer typed. Hang the booking on the same visitor ID and you can read any deal backwards without opening an analytics tool. Everything else you might be tempted to collect is interesting once and never again.

Two honest caveats. Self-report is noisy, because people misremember where they first saw something and some will name the last thing they touched. Treat it as corroboration for the log rather than as proof on its own. And the engines don't agree with each other, which we dug into in whether a recommendation from one engine means the others follow. One buyer naming ChatGPT tells you about that buyer, not about your standing everywhere.

When the log and the self-report disagree, the log usually has the sequence right and the buyer usually has the reason right. Somebody who first landed from a search engine and then writes ChatGPT in the box is often telling you the truth about what convinced them rather than about what first brought them in. Keep both answers on the record and stop trying to collapse them into one source.

What changes once we can see it?

The shape of the data sorts you into one of three situations, and each one points at a different fix.

If there are no AI-sent humans at all, you're not in the answer yet. That's a visibility problem, and more publishing is a reasonable response.

If AI is sending humans and none of them book, you have a harder problem, and it isn't the one most founders go looking for.

If both numbers are moving, you have a channel that works, and the budget question turns into arithmetic instead of argument. We put numbers around that decision in how much of our marketing budget should go to AI search.

What if AI sends people and they still don't book?

This is the uncomfortable case, and it's the common one. The engine did its job. It named you, the buyer arrived, read the pricing page, and left. Nothing in that sequence is a visibility failure.

It's what a page does when it explains what a company does without answering why the buyer should change, why they should change now, and why they should change with you. That last question is where most homepages go quiet. Being named by an engine gets a stranger to your door with no reason to walk through it.

The usual culprit is Solution-Centric Marketing: listing capabilities and trusting the buyer to assemble the argument for change on your behalf. Buyers won't do that work, and an AI engine summarizing your pages can only repeat the material you gave it. Volume doesn't fix it. Perspective does, and perspective lives in your narrative identity rather than in your logo or your palette.

If you'd rather see the outside read before you take our word for it, the Brand Signal Score scores your homepage on narrative clarity, trust, AI-readiness, and conversion, and it's free.

Building the attribution join itself is a small piece of engineering wrapped around a bigger change of habit. If you want to see what AI engines can read on your own homepage before you build anything, start with the free Brand Signal Score.

What this means for you

Build the join because it makes your next argument an honest one. Once a booked call carries its own history, you can tell the difference between a channel that isn't reaching anyone and a message that isn't convincing anyone. Those two problems look identical on a traffic chart and they need opposite responses.

Our own answer came back as 24 days, four sessions, and a buyer who typed ChatGPT into a text box. Before we could see that, every one of those bookings looked like luck.

“An engine can hand you the buyer. It can't hand them a reason to change.”