What have authoritative sites measured lately?

On 2026-05-06, Semrush published “How to track ChatGPT traffic for you and for competitors.” Its value lies in splitting the problem in two: tracking your own AI referral traffic, and tracking a competitor’s. The first relies on your own site data; the second can only rely on third-party estimates. The two are nowhere near equal in accuracy.

They did not claim this was easy. The reason is down to earth: AI products do not redirect in a stable way, and some sessions get classified as direct traffic by analytics tools. Ahrefs made the same point in “AI Search ROI Is Messy” on 2026-10-02 — the conversion chain in AI search is broken. A user may see you in an answer and only search your brand name themselves a few days later, and in any analytics tool that path is recorded as organic search.

So the correct expectation for tracking AI traffic is not “one report that shows everything,” but “several complementary pieces of evidence pieced together into an approximation.” I used the same approach in how to track ChatGPT’s citation sources: record the original text and the source link first, then talk about traffic.

Why does AI traffic “disappear”?

First, the referrer is lost. When a user opens a link inside an AI app, that click may open in-app, or the user may copy it into a browser and open it there. In the second case the referrer information is broken, and the analytics tool can only count it as direct traffic.

Second, users don’t click the link. According to Ahrefs’ field test in the French market on 2026-08-27, click-through rate fell 23.1% when an AI Overview appeared. Many users are satisfied after reading the summary; they only remember your brand name and search for it themselves two days later. On that path, the data shows no trace of AI.

Third, the traffic is too small and drowns in the noise. If your site gets only a few dozen organic visits a day, the three or five visits AI brings show no pattern at all. At that point the priority is not buying a tool but growing the base — and doing the human confirmation well at the same time. You can sort out the measurement definitions first with breaking down GA4 organic traffic.

There is another often-overlooked source: links forwarded through email and instant messaging. A client copies the URL from an AI answer into WeChat or an email and forwards it to a colleague, who then opens it. On that path no analytics tool can see the source. B2B export decisions are often made by two or three people together, and the share of such forwards is not small. So when you notice direct visits to a content page rising and at the same time receive related inquiries, don’t rush to dismiss AI’s role — first ask the client how they found you.

Three layers of tracking: reports, logs, and people

Of the three layers, manual confirmation is the easiest to skip, yet it has the highest value. You don’t need to ask formally; one sentence is enough: “Would you mind telling me where you heard about us?” Clients are usually willing to answer. That single piece of information can directly verify the guesses made by the two reporting layers above.

The log layer is workable for export sites running a dedicated server or VPS; shared-hosting users often can’t get it. If you can’t get it, let it go — there is no need to switch hosts for this layer. Doing the reports and the human layer solidly already answers most questions.

How exactly do you filter in GA4?

Start with the referring domain

In the traffic acquisition report, filter by session source and focus on chatgpt.com, chat.openai.com, perplexity.ai, copilot.microsoft.com, and gemini.google.com. Don’t just look at whether there is volume — look at which page they land on. The landing-page distribution is often more informative than the numbers.

If the volume is zero, don’t jump to conclusions. It may genuinely be zero, or it may have been recorded as direct traffic. In that case, look at the landing pages of direct traffic: if they cluster on a few content pages, and those pages happen to be the ones you were cited by AI on, then you can reasonably infer the visitors came from AI. See how to read the AI-related query report for comparison.

Then look at landing pages and behavior

AI referral visits usually have two traits: the landing page is a content page rather than the homepage, and time on page skews short. That is because the user arrived with a specific question and leaves once it is answered. Don’t treat a short stay as poor quality outright — check whether they went on to a second page or clicked the inquiry button.

Finally, cross-validate

Put the GA4 trend line and the changes you recorded by hand in your AI citations into one table. Only when both move at the same time is it worth writing into a report. Rely on one side alone and you will be fooled by random fluctuation. For tool choices you can refer to the AI search monitoring tool comparison, but I suggest running the free approach for a full month before considering a purchase.

What do our own records show?

Let me give one of my own cold data points. This site has a 100% indexing rate, yet in 28 days it got only 12 impressions and 0 clicks. That means at this order of magnitude, talking about “the share of AI referral traffic” has no statistical meaning — the denominator is too small. What we did at this stage was fixed-question sampling and recording original text, not staring at traffic reports.

The data from our panel illustrates the difficulty of tracking even better. Across the T0 and T1 rounds, the total number of cited domains went from 322 to 245, with only 90 overlapping. The source set itself rotates violently, and it is very hard to map one traffic curve onto one particular citation. This is also why I insist on storing content-layer records and visit-layer records separately. The full method is in field research and method notes.

For most export sites my advice is consistent: until the base reaches a few hundred visits a day, make the content layer solid first and use that one sentence on the inquiry form to cover the business layer. Once the volume is there, add a dashboard. See how to set up AI visibility monitoring.

A low-cost implementation checklist

Three steps, no tools to buy, and you can start this week.

1. Build a source filter in GA4: list the main AI domains in a custom exploration report and check the landing pages once a week.
2. Add one source-confirmation line to your inquiries: add a dropdown to the form, ask by hand in chat and email, and log it in a shared spreadsheet.
3. Reconcile once a month: put the GA4 trend side by side with the citation changes you noted by hand, and write down one conclusion and one counterexample.

If you really want to turn this into a routine, see how to diagnose rising traffic with falling inquiries, so you don’t mistake low-intent AI visits for good news.

The questions the boss asks most often

Why can’t I see ChatGPT as a source in GA4?

A portion of those sessions get recorded as direct traffic or as an unidentified source, because a client-side redirect does not always carry complete referrer information. So GA4 is only one clue, never the sole piece of evidence.

Can I track AI referral traffic without buying a tool?

Yes. GA4 source filtering, server logs, and one line of human confirmation on the inquiry form add up to enough to make a judgment. Tools solve convenience, not feasibility.

How do I tell AI traffic apart from organic search traffic?

Look at the referring domain and the landing page. AI referrals usually land on content pages and the referring domain is a model site; organic search lands on the page that matches the keyword. When the two are mixed together, work backwards from the time of the inquiry.

AI referral traffic is small. Is it still worth tracking separately?

Yes. Small now does not mean small later, and this traffic usually carries clearer intent. Build the recording habit first, so that you have a baseline to compare against once the volume grows.