How to track AI Overview traffic — Google mixes it into your ordinary numbers

The short version

• You cannot separate it. Google states that sites appearing in AI features “are included in the overall search traffic in Search Console”, inside the ordinary Web results, not as a category of their own.
• An AI Overview is recorded as one position, and every link in it gets that position. That is Google’s wording, and it means being cited in an AI answer can make your average position look better without anybody arriving.
• We tested it on ourselves. The Search Console API returned no search-appearance rows at all for our site, and 2,479 of our 2,492 impressions sat in one undifferentiated bucket.
• Our own 28 days show the shape everyone is worried about: impressions up 96%, average position improved from 12.0 to 9.3, and clicks flat at 11. Our click-through rate halved.
• We cannot prove why, and neither can anyone else — because of the first point. What we can do is show the arithmetic and the documentation, and let you check your own.

What Google actually documents

Two pages carry the whole answer. The first is Google’s Search Central documentation on AI features, which says: “sites appearing in AI features (such as AI Overviews and AI Mode) are included in the overall search traffic in Search Console”. It goes on to say they are reported in the Performance report inside the “Web” search type — which is where everything else already is.

The second is the Search Console help page defining impressions, position and clicks. On AI Overviews it says: “Clicking a link to an external page in the AI Overview counts as a click”, that standard impression rules apply, and — the sentence that matters most —

“An AI Overview occupies a single position in search results, and all links in the AI Overview are assigned that same position.”

Read that twice, because it is the mechanism. An AI Overview generally appears at the top of the page. If your page is one of the sources cited inside it, your impression is recorded at that position. Ten cited sources do not get positions one to ten; they all get the same one.

The same page covers AI Mode, which it describes as expanding “on AI Overviews to show a more interactive AI-powered response with links to web resources”. Clicks there count too, and there is a detail worth noticing: “If a user asks a follow-up question within AI Mode, they are essentially performing a new query. All impression, position, and click data in the new response are counted as coming from this new user query.” One person, one sitting, several rounds of questions — recorded as several separate searches.

So we tried to separate it, on our own site

Documentation is one thing; we wanted to see it. The Search Console API has a searchAppearance dimension, which is how you normally break results down by type — rich results, videos, and so on. We ran it against our own property for 10 August to 6 September 2026.

It returned nothing. Not zero rows for AI Overviews specifically — no rows at all. We then asked for each search type in turn, and found 2,479 impressions under web, 13 under image, and nothing anywhere else.

Being fair to the data: an empty search-appearance result partly reflects that we do not earn many special result types. It is not by itself proof that AI Overviews are unfilterable. But it matches what Google documents, and it means that for a small business site like ours there is no filter to reach for, because there is no separate figure behind it.

What our own numbers did

We publish roughly a post a day, so we have a reasonable amount of our own data to look at. We pinned two equal fourteen-day windows rather than using a rolling range, because rolling windows quietly move the goalposts.

Impressions nearly doubled. Average position moved from the middle of page two to the bottom of page one. And the number of people who actually arrived did not change by one.

The boring explanation, and why it does not hold

Before reaching for anything about AI, the obvious explanation deserves testing: we published more pages, so of course we got more impressions. That is a real effect and it would be lazy not to check it.

So we counted. In the first window, 57 pages earned at least one impression. In the second, 64 — a rise of about 12%. But impressions per page went from 14.7 to 25.6, a rise of about 74%. We were not mainly being shown more because we had more pages. Each page was being shown far more often.

The number of distinct queries we could see also doubled, from 24 to 48, while the share of impressions taken by our single biggest page stayed at 17% in both windows. So the growth was broad rather than one page getting lucky.

That kills the comfortable explanation without proving the uncomfortable one. Being shown for twice as many questions, at a better average position, to exactly the same number of visitors, is consistent with the mechanism Google documents. It is also consistent with several other things. We cannot tell you which, and the reason we cannot is the subject of this article.

The thing you are being sold, that Google says you do not need

While we were in the documentation, one line stood out, because it contradicts a good deal of what is currently being sold to small businesses: “You don’t need to create new machine readable files, AI text files, or markup to appear in these features.” And immediately after: “There’s also no special schema.org structured data that you need to add.”

We looked at one of those files in detail recently — almost nobody reads your llms.txt, and the things that do read it can run code. Google’s position here is a useful second opinion to have in writing before anyone invoices you for one.

What argues the other way

Several things, and they matter.

Our sample is one small site with 11 clicks in a fortnight. Eleven. At that volume the click number is close to noise, and a fortnight where two people happened to click differently would change the percentages a great deal. We are showing you a shape, not a finding about the internet.

An improving average position has ordinary explanations too. We publish daily on narrow topics, we have been deliberately targeting questions with little competition, and newer pages tend to enter at better positions than the site’s older average. That alone could move 12.0 towards 9.3 without an AI Overview being involved at all.

And we found at least one cause that has nothing to do with AI: on our biggest query by far, our page used a different word from the searchers. They typed “overdue”; our title said “unpaid”, and the word overdue appeared nowhere on the page. That is our mistake, not Google’s, and we have changed it. If a title fix moves those clicks, the AI story explains less than it appears to. We wrote about the wider version of this — what your website is for when Google answers instead of linking — and this is the same argument met in our own data rather than in a survey.

Finally, an impression inside an AI answer is not worthless. If someone asks a question, gets an answer, and sees you cited as the source of it, that is not nothing — it is just not a visit, and your analytics will never know it happened.

What to do about it

1. Stop treating impressions as a success measure. They now include appearances in answers that were designed to save the reader a click. The number can rise while your business gets quieter.
2. Compare pinned periods, not rolling ones. Two equal date ranges, set explicitly. A rolling 28 days hides exactly this kind of change because the baseline moves with you.
3. Watch clicks and outcomes. Enquiries, calls, replies, bookings. Those cannot be inflated by a summary, and they are what you are actually trying to grow.
4. Check your titles against the words people type. Search Console shows you their phrasing. If your page uses your word and they use theirs, you may be losing clicks for a reason that has nothing to do with AI.
5. Do not buy a file. Google says in its own documentation that you do not need one.