Measuring AI referrals: can GA4 show ChatGPT traffic?

The short answer

Yes, GA4 can show you that ChatGPT, Perplexity or Gemini is sending visitors to your site. And no, what you see there is not the real number. It is the floor.

Part of those visitors arrive without a referrer and therefore count as direct traffic. That is not a misconfiguration in your analytics, it is how those visitors turn up. Copy a link out of a conversation, paste it into a new tab, and no origin travels with it.

That does not make measuring pointless. It only changes which question you ask. Not "how many visitors do I get from AI", because that figure is never right. Instead: is this channel starting to move, and what do these people do once they are here. You can answer both today, with the data you already have.

And the answer to the second question is why this matters. In March 2026, AI traffic converted 42% better than non-AI traffic, a record. A year earlier that same traffic converted 38% worse. Within twelve months it flipped from your worst channel to your best.

Why the number in your dashboard is wrong

A referral exists because one page calls another and passes on where the visitor came from. In a chat conversation that often does not happen.

Three things strip that origin away. A link someone copies and pastes into a new tab comes in as direct traffic. The built-in browser of an app frequently does the same. And some assistants route you through an intermediate page, so something does come along, but not always the name you expect.

What you see in GA4 is the share that passes its origin along properly. The share that does not sits hidden in your direct traffic, together with people typing your address from memory and with everything your email client fails to hand over.

So treat the figure as a thermometer, not as a receipt. The direction is usable. The absolute number is not.

How do you make it visible in GA4?

By default these visits disappear among the rest of your referral traffic. You have to set them apart to see them.

Build an exploration with session source as the dimension and sessions, engagement and conversions as the metrics. Filter session source on these hostnames:

Then put the same set into a channel group of its own, so it does not stay manual work every month. From that point on a line grows in your reporting that you can compare month over month.

Do this even if you expect zero for now. How assistants arrive at their shortlist is covered in our piece on agentic search and your webshop traffic. For nearly everyone the channel starts at a few sessions a month, and you want to see that first movement at the moment it happens, not six months later.

What you see when you do this on a real site

At a hospitality client of ours, ChatGPT sat at twenty-six sessions in a month this spring. Small number, and still the most interesting number of that month. Because twenty-four of those twenty-six visitors landed on the homepage.

That is no coincidence. Their menu was on the site as fifteen images, without text and without prices. An assistant asked "what does it cost there" cannot read that menu, so it points to the only thing it does understand: the front page.

That is the lesson this kind of measurement gives you. Not "we get too little AI traffic", but: which page do they enter on, and why not the page that holds the answer. In this case the job was not to make more content, it was to put a menu into text.

Adobe measures the same pattern at larger scale. Across the retail sites they checked, product pages scored an average of 66% on machine readability, against 75% for homepages and 74% for category pages. So the page where you earn your money is the least readable for the system that is supposed to recommend you. Which elements do make a category page citable is written out in the seven elements of a GEO-ready category page.

The measurement nobody does: just ask

There is a second measuring point that needs no configuration and that catches exactly the visitors you lose in GA4.

Put one open field on your quote form: how did you find us. Or ask the question at the end of your first reply to an enquiry. Two sentences of work.

Last week a brand came to us over WhatsApp. Asked that question, the answer was: online, through an AI assistant. In none of our dashboards did that enquiry show up as anything other than a message out of nowhere. The only reason we know where it came from is that we asked.

That is not an anecdote, it is a method. On this channel self-reporting is more reliable than your analytics, because the visitor does know where they came from and your software does not. Record the answers in your CRM, count them per quarter, and you have a second line next to your GA4 line. If you want to know how you get into that shortlist in the first place, start with your own product data.

What do those visitors mean once they are there?

The volume is small. The behaviour is not.

Over the first three months of 2026 Adobe measured 393% year-on-year growth in AI traffic to US retail sites. More important than the growth is what those visitors do: they stay 48% longer on the site and view 13% more pages than visitors from other sources, and engagement runs 12% higher.

That fits how such a visit comes about. Someone has already put their question to an assistant, has been given a shortlist, and only clicks through when they want to check or buy something. You do not get a visitor who is orienting, you get someone who is three steps further along.

It also explains why the same traffic converted worse a year ago. Back then assistants sent people to arbitrary pages. Now they send people to the answer, provided your page carries that answer in readable form. The same applies across channels: what survives of your brand on bol or Amazon is described in product content across channels.

The trap: drawing conclusions too early from small numbers

At thirty sessions a month, one order more or less already shifts your conversion rate by tens of percent. Anyone steering at that level is steering on noise.

So for the first six months, look at patterns rather than percentages. Is the number of sessions growing month over month. Do they enter on more than one page. Do you see the same questions in your forms that people put to an assistant. Those are signals that do carry meaning at small numbers.

There is one more reason to be careful with conversion figures on this channel: these visitors are pre-selected. They convert better because they come in later in their decision, not because your site suddenly works better. So do not credit yourself for a channel that is mainly good at skipping the orientation phase.

What you can do tomorrow

Four steps, in this order. The first two take half an hour, the third and fourth are real work.

Switch the measurement on. A channel group in GA4 with the hostnames above. Do this first, because nothing after it becomes visible until the measurement is running.

Put the question on your form. One open field. Do not make the field required, that lowers your conversion. Whoever fills it in hands you the signal.

Look at the landing pages, not at the total. If most of them land on your homepage while your question is answered somewhere else, that is where your work is.

Put your facts in text. Contents, size, price, stock, delivery time and reviews, as readable text on the page itself and not only in an image or behind a collapsed block. That is what an assistant can quote, and it is exactly where product pages score worst.

What you do not need for this is a new file on your server. Google states in its own documentation on AI and Search that you do not have to create separate machine-readable files to appear in Google. The gain sits in the page itself. How to set up a product page and a bol listing for that, we worked out in Webshop and bol listing ready for AI search results.

Why we track this

We build brands and webshops, and this year a channel has been added that behaves differently from all the others: small in volume, high in intent, and poorly measurable with the tools everyone uses.

The brands doing something about it now are not doing it because it already pays off. They are doing it because the measuring point is cheap and the learning curve is long. Start measuring only once the channel is big, and you miss the year in which you could have learned which pages get recommended and which do not.

If you want to know where your site stands: we look at the landing pages of your AI traffic, at the readability of your product pages, and at whether your answers appear as text on the page at all. That is an afternoon of work and it produces a list you can carry on with yourself.