Can I see AI traffic in Google Analytics at all?

Partially. When somebody clicks a link inside an answer you will usually see a referral from chatgpt.com, perplexity.ai or a similar host, and you can group those into a channel. What you cannot see is the far more common case where the assistant answered and nobody clicked, which is the part that matters.

My traffic is flat, so has anything really changed?

Flat traffic is consistent with nothing changing and also with losing a growing share of recommendations while an existing channel grows slightly to compensate. The two look identical in a line chart. Flat traffic is evidence about traffic, not evidence about whether you are being recommended.

The concept is not. Featured snippets, weather boxes, currency conversion and knowledge panels have been answering questions without a click for over a decade. What is new is the scope: assistants will now handle a comparison and a recommendation, which used to require visiting several sites.

What can I measure instead of clicks?

Whether you are named. Ask the assistants a fixed set of buying questions on a schedule, record which businesses appear and in what order, and watch the trend. It is a directional index rather than a count of real conversations, but it moves when your position changes, and traffic alone does not.

The zero-click problem: when the answer replaces the visit

Zero-click search is any question answered in full on the results page or inside an assistant, so the person never visits a website. For a business the consequence is that a recommendation made to a potential customer leaves no referrer, no session and no trace in your analytics. You cannot see a visit that did not happen, which makes flat traffic a poor source of reassurance.

What zero-click means concretely

A zero-click result is one where the person gets what they came for without leaving the page they are on. Somebody asks what time a shop closes, how many millilitres are in a cup, or what a term means, and the answer appears in front of them. No site is visited. No analytics anywhere records a person.

That much has been true for over a decade, and most businesses shrugged, reasonably enough, because the questions being answered were trivial. Nobody loses revenue to a weather box.

What changed is the size of the question an answer can swallow. "Which conveyancing solicitor should I use in Bristol, and what should I expect to pay?" is not a trivial question. It is the whole research phase of a purchase, and an assistant will now handle it in one reply: three firms named, a sentence of reasoning each, a rough price range, and a suggestion of what to ask them.

The person who asked that may well go on to contact one of those firms. They will probably find it by searching its name, or by clicking a link in the answer. For the two firms not named, nothing happens at all — and nothing happening is not an event any system can log.

Why your analytics cannot show it

Web analytics is built on a simple premise: a browser requests a page, and a script on that page reports the request. Everything you know about your audience descends from that one event. Sessions, sources, bounce rates, conversion paths — all of it requires somebody to arrive.

A recommendation that does not produce an arrival produces no event. There is no partial signal, no "almost visited" metric, no impression count the way there is for a search result. From your analytics' point of view, the conversation did not occur.

It is worth being precise about what this does and does not mean. Your analytics is not broken, and it is not lying. It is answering the question it was designed to answer, which is "who came to my site and what did they do". The question it cannot answer is "who was told to go somewhere else". That was always outside its remit; it simply did not matter much until recently.

The little that analytics can show

There is a narrow band of visibility, and it is worth claiming. When somebody does click a link inside an answer, the visit usually arrives with a referrer, and you can see it.

Referrals from assistant hosts. Traffic from chatgpt.com, perplexity.ai and similar hosts shows up in your referral report. Grouping them into a custom channel takes a few minutes and gives you a trend line worth watching.
Branded search lift. People who hear a name in an answer frequently search that name rather than clicking. A rise in branded impressions in Search Console, without a matching rise in anything else, is a weak but real hint.
Direct traffic with no obvious cause. Equally weak, equally worth noticing, and equally easy to misread — direct traffic is a bin for anything untracked.
Enquiries that mention it. The most honest signal available to a small business is a human one. Adding "how did you hear about us?" to your enquiry form costs nothing and occasionally returns the sentence you needed.

Each of these tells you something about the minority of cases that ended in a click. None of them tells you about the majority that did not. Treat them as a thermometer in one corner of a large room.

Why "traffic is flat, so nothing has changed" is false comfort

This is the most common reasoning error in the whole subject, and it is a genuinely understandable one. Traffic is the number everybody watches, it has not fallen, and so the conclusion seems to follow.

The problem is that flat traffic is consistent with several different underlying situations, and they are indistinguishable from the chart alone. Nothing has changed. Or: you are losing an increasing share of assistant-led research while returning customers and branded search grow enough to cover the gap. Or: a seasonal rise is masking a structural fall. The line looks the same in all three.

There is a second, sharper version of the error. If a competitor is being named in answers and you are not, the damage shows up in their numbers, not yours. Your traffic does not drop when somebody else is recommended to a person who was never going to find you anyway. You lose the opportunity without losing a measurable visit. By the time it reaches your revenue, it has been invisible for a long while.

No analytics configuration, and no product, can recover the missing data. Nobody can see inside real conversations between customers and assistants — those are private, and any vendor implying otherwise is overselling. The honest position is that the zero-click gap is not measurable directly, and the only sound response is to measure something else that moves in the same direction.

A worked example

Take an independent law firm in Austin with a steady flow of enquiries. Over a year its organic sessions rise slightly. Its rankings for the phrases it tracks are unchanged or marginally better. On every dashboard it owns, the year was fine.

Ask an assistant which firms in Austin handle a particular kind of case well, and it names three others. It names them because two of those firms are listed in a widely cited legal directory with detailed practice-area descriptions, and the third publishes a plain-English guide to exactly that kind of case. None of that is visible in the first firm's analytics, because none of it involves the first firm at all.

The firm's position did not get worse. The set of people who will ever encounter it got smaller, slowly, in a part of the market it cannot observe. That is the shape of the problem: not a drop, but a ceiling you cannot see.

What you can measure instead

If you cannot count the visits that did not happen, count the recommendations. Ask the assistants the questions your customers ask, and record what they say. Done consistently, that produces a signal with three useful properties.

Presence. Are you named at all, for the questions that precede a purchase? This is binary, blunt and the most valuable thing to know first.
Position within the answer. Being named third in a list of three is different from being the one recommended in the first sentence.
Who appears instead. The competitor set an assistant names is often not the competitor set you track, and the difference tells you where the corroborating sources point.

Three disciplines make this worth the effort. Ask the same questions each time, because changing the wording changes the answer and destroys the comparison. Ask repeatedly, because the models are probabilistic and one run is noise. And qualify your geography: Richmond, Perth, Boston and Hamilton all exist in several countries, and an unqualified town name produces answers about the wrong one.

How to think about what the number means

A visibility measurement of this kind is a directional index. It is not a count of real customers, not a transcript, and not a forecast of revenue. It behaves rather like rank tracking did before anyone had conversion data: imperfect, noisy, and still far better than guessing.

Used that way it answers the questions that matter. Am I in the conversation at all. Is that improving or sliding. Did the work I did three months ago change anything. Who keeps being named instead of me, and what do they have that I do not. The approach, and its limits, are set out in more detail in measuring AI visibility honestly, and the mechanical differences from search are covered in AI visibility versus SEO.

The uncomfortable part of zero-click is not that the data is bad. It is that the absence of data reads exactly like good news. A flat chart feels like stability, and for a while it can be the sound of a market moving on without you. If you want a starting point rather than a project, a free visibility check will tell you whether you are named today, and who is named in your place.