The Queries No Human Typed: AI Fan-Out Queries in Search Console Data of Five UK Aesthetics Websites

Here’s a search query. It appeared in the Search Console account of a UK aesthetics clinic 669 times over 16 months. The clinic’s page ranked, on average, at position 67 for it. It was clicked zero times.

We don’t think a person typed that. Six hundred and sixty-nine appearances that far down the results with no click is odd on its own, and if you were looking into cellulite treatment for yourself, you wouldn’t say “in a clinic setting”. That phrase comes from someone asking on behalf of a clinic, or something asking on behalf of someone.

We spent the last month pulling long queries out of five aesthetics websites, and we found 565 like it.

What fan-out is

When someone asks Google’s AI Overviews or AI Mode a question, the system doesn’t run that question once. It breaks it into lots of sub-questions, searches for each of them separately, reads the pages it finds, and writes an answer from what it read. Google calls this query fan-out, and its own documentation says both AI Overviews and AI Mode may use it, issuing multiple related searches across subtopics to build a response. The model writes those sub-questions.

The same page confirms that appearances in AI features are counted inside the ordinary Web search type in Search Console. What Google’s documentation doesn’t say is exactly what gets recorded when a page is retrieved for a sub-question and then not used in the answer. Our working hypothesis, and the reason this study was run, is that at least some of those sub-questions are showing up in ordinary query data: a query nobody searched, at a position nobody scrolled to, with a click nobody actually made. We can’t confirm that from the documentation. What we can do is look for the queries and see whether they behave the way we think.

What’s already known

We’re not the first to notice these. Metehan Yesilyurt was filtering Search Console for unusually long queries in June 2025, and Barry Schwartz wrote it up with the caveat that they might just be very long human searches. Chris Long released a 10-word regex for surfacing them in February 2026. In August, Suganthan Mohanadasan published “Yes, Go On”, a 16-month analysis of his own site that found 1,127 of these queries and 20,300 impressions, sorted them into seven kinds (conversational follow-ups, reply fragments, pasted strings, agent probes and so on), and found that most of the pool is hidden anyway: in a 59-day BigQuery export, 57.7% of his web impressions carried no query string at all. The same week, John Mueller confirmed that AI Overviews and AI Mode data sit in the general performance report.

So the fact that AI-shaped queries appear in Search Console is established. What isn’t established is what they look like across more than one site, how much of the long-query population they account for, whether a human reader agrees with the filter, and what actually happens when you run them. That’s what this study does: five sites in one vertical, a comparison group, hand-labelling, and live verification of every query we tested.

Our own earlier work sits on the other side of this process. In the 314-clinic AI visibility audit, we asked ChatGPT, Google’s AI Overviews, and Perplexity which UK clinics they recommend for six treatments across nine cities, and found the three platforms agreed on only 36% of their recommendations. That study measured what the AI says. This one looks at what it asks first.

What we did

We took five aesthetics websites and pulled every query of seven words or more from the 16 months Search Console held on 5 September 2026 (24 April 2025 to 3 September 2026). Two are patient-facing clinics (Clinic A and Clinic B), two are practitioner-facing device companies (a distributor and a manufacturer), and one is a hair removal salon. We’ve anonymised all five and haven’t quoted any query that identifies a site by brand, product, person, or place.

We then applied one fingerprint to every property, identically. A query is flagged when it has all of the following:

• seven or more words
• ten or more impressions
• zero clicks
• at least two of five structural signals: a question-word opening (what, which, how, can), a conditional or comparison clause (if, rather than, without needing), specification language (recommended, suitable, criteria, professional, clinic), complete grammar (comma clauses, a terminal question mark), or a whitespace anomaly (double spaces, line breaks)

Against the above, we built two comparison groups. The first is every seven-plus word query on the same properties that did get clicked. The second is the one that matters more: every seven-plus word query with zero clicks and ten or more impressions that didn’t match the fingerprint. Those 1,396 queries passed the same impression and click thresholds as the flagged group and failed only the structural test, so they’re the right control for the question “does zero clicks alone produce this pattern, or does the wording?”

Two things to clear up first. Some differences between these groups are built in by how we defined them. The flagged group has zero clicks because we needed it, and the clicked group has clicks because we needed that too, so we can’t present the click rate as a discovery. The same goes for question marks: they count towards the fingerprint, so of course the flagged group has more of them. The comparisons worth paying attention to are the ones we didn’t select for, like how long the queries are beyond the seven-word minimum, and which words they use that the fingerprint never checks.

Two of us then hand-labelled a sample of 300 queries as “machine”, “human” or “unsure” without seeing the automatic result, and finally we ran 37 of the flagged queries live in clean browser sessions to see what Google does with them.

A word on what this can and can’t prove. Google doesn’t label these queries, and we can’t completely prove machine origin for any individual one. Some of our 565 are probably the conversational follow-ups and pasted strings in Suganthan’s taxonomy rather than fan-out sub-questions, and the fingerprint can’t tell those apart. What we can show is that a distinct population of queries exists in the data, that it doesn’t look like the queries people actually click on, and that it behaves the way fan-out would predict. Everything that follows should be read as evidence for a hypothesis, not a guarantee of how Google works.

What we found

565 queries, 18,260 impressions

Across the five sites, 565 property-query pairs matched the fingerprint (552 distinct query strings, since 13 appeared on two properties), accounting for 18,260 impressions over 16 months. Median position 5.9, and 88% sat at position 10 or better.

Flagged = matched the fingerprint. Zero-click, unflagged = seven-plus words, zero clicks, ten or more impressions, failed the structural test. Clicked = seven-plus words, one or more clicks.

On most of these measures, the 1,396 unflagged zero-click queries resemble the clicked ones rather than the flagged ones. They’re about the same length, they start with “how to” at the same rate, and they ask for before-and-after photos. (They ask for “near me” less often than the clicked group, which is what you’d expect: local intent is what gets clicked).

They’re ordinary long searches that happened not to get a click: “if i get botox in my forehead will it stop sweating”, “can you do red light therapy after laser hair removal”. The flagged group is four words longer on average, almost never asks “how to”, never asks for photos, and a quarter of it is written from the point of view of a clinic buying something. Whatever is producing that difference, it isn’t the lack of a click. The pattern holds inside a single site as well as across the five: on site A alone, the 279 flagged queries average 12.5 words with no “how to” and no before-and-after requests, against 8.3 words, 17% and 10% for its 1,175 unflagged zero-click queries, so the difference isn’t an artefact of mixing device and consumer sites.

(These features aren’t all direct selection criteria, but the fingerprint can still shape their distribution: “how” contributes to the question-opening signal, and longer queries have more room to score. The clinic-point-of-view row overlaps most; the before-and-after row least.)

That length difference matters more than it looks. Pew Research tracked 68,879 real Google searches in March 2025 and found an AI summary appeared on 8% of one- or two-word searches and 53% of searches of ten words or more, rising to 60% for question-shaped searches. Long questions are what trigger AI answers. Long questions are also what this population is made of.

Device sites had the highest share of flagged queries

Nearly a third of every long query hitting the device distributor’s site matches the fingerprint, and 5% of everything it’s seen for in Google falls into this group. At the other end, the salon, a consumer service business with local intent, got one. 

Five sites can’t tell us whether that holds across the sector, but the ordering is consistent with purchase complexity: the more considered the decision, the more sub-questions you’d expect an AI to generate, and the more often pages about it would be retrieved.

The clinics sit in between, and Clinic A’s 279 queries include something we hadn’t expected. Several are practitioner questions (“what is the cost of professional body contouring machines for clinics”), and they appeared in the patient-facing clinic’s data too.

The same question appears on competing sites

Thirteen flagged queries appeared, word for word, on two of our properties. The largest was this one:

994 combined impressions across the device distributor and the device manufacturer, zero clicks on either. Identical wording on two sites doesn’t prove a single retrieval event pulled both, but it’s the pattern you’d expect if one did. When we ran it live, the AI Overview cited one of them and not the other.

The decision tree

The clearest structural feature in the dataset is repetition with substitution. On the device properties, the same sentence stem appears dozens of times with one qualifier swapped:

That looks like a model running through all the branches of a purchase decision. The repeated syntax across different buying criteria, with American spelling on a British device site, is what makes this group worth investigating rather than dismissing as long-tail noise. The GLP-1 thread is interesting too: 31 of the 565 queries mention GLP-1 medication, weight loss, or the skin laxity that is caused by it.

On the clinic side, the recurring feature is the described patient. “Busy moms seeking a quick skin refresh.” “Skincare trend followers.” “Someone who prioritises comfort over intense procedures.” People don’t usually describe themselves in the third person in a search box. Something generating a question on behalf of a person would.

When it started

We have complete month-by-month data for four properties. Fingerprint impressions on the device manufacturer peaked in July 2025 and again in April 2026. For the device distributor, they went from zero in January 2026 to 1,099 in July 2026. Clinic B spiked in December 2025 and has been climbing since. The shapes differ by property. All four are higher at the end of the window than at the start which is consistent with more searches being answered through AI features.

What happens when you run them

We ran 37 of the flagged queries, the highest-impression non-identifying ones from each property, in clean, logged-out UK sessions on 5 September 2026, entered into standard Google Search with no further interaction.

All 37 produced an AI Overview on the results page without us doing anything. We’d expected most; we got every one. Whatever these queries are, they’re the kind of question Google now answers with AI.

Ten of the 37 cited the property whose Search Console they came from: five of ten for the device distributor, two of six for Clinic B, two of ten for Clinic A, one of ten for the device manufacturer, none for the salon.

Split by position: 10 of the 25 queries where the property averaged position 10 or better produced a citation, and 0 of the 12 where it averaged worse than that, including several at 50 to 67. So ranking well didn’t guarantee a citation (fifteen well-positioned queries got none), but in this sample, nothing that ranked badly got one.

Other studies have looked at the overlap between organic rankings and AI citations. Ahrefs found 76% of AI Overview citations came from top-10 pages in mid-2025, then 38% in a later, larger sample, and BrightEdge puts it at under 17%. Those figures measure live organic rank at the moment of citation; ours is a 16-month Search Console average compared with a citation observed on one day, so the two aren’t directly comparable. The test that would make them comparable is recording organic position and citation outcome in the same search session, which is on the list for the next round.

Our reading is that Google’s fan-out fetches deep and cites shallow. A page gets pulled as candidate material from well down the results, registers an impression, and loses the citation to a page that ranks higher for that sub-question. Thirty-seven queries can’t confirm it, but it’s the pattern you’d expect if it were true.

What the new Search Console report shows

Google launched generative AI performance reports in Search Console in June 2026 and rolled them out to every property worldwide on 31 August. They show how often links to your pages appeared inside AI Overviews, AI Mode, and AI features in Discover, by page, country, device, and date. They don’t show clicks, positions, or the queries involved.

So the new report counts the times a link to your page was shown inside an AI answer. It can’t tell you what question the AI was answering when it showed you, and it has no view of pages that were retrieved and not shown, if that is what our flagged queries represent. The fingerprint queries are a way of looking at the query side that the report doesn’t offer. Read the two together and you have both halves of a picture that neither shows on its own.

Why zero clicks isn’t zero value

Three things follow for anyone reading a Search Console report for an aesthetics business.

Your impression count includes queries that, on our reading, no patient made, and your CTR is deflated by them. Clinic A’s 8,082 fingerprint impressions are a rounding error against 3.4 million total, but on a device site running at 5% they move the average, and if you report CTR, this is now part of the denominator. It’s also adding on top of a broader fall we’ve written about before: organic click-through on queries with an AI Overview dropped by around 61%, while brands cited inside the Overview saw 35% more organic clicks than those that weren’t (our earlier analysis covers the numbers; Seer Interactive’s 2026 update across 53 brands shows a partial recovery since, with cited pages still well ahead).

The impressions are also a signal about what the AI is looking for on your pages. If the queries hitting your device page are about warranty support and staff training, those topics are worth checking against what the page actually says.

And the position column should be read differently for this group. Given what the live checks showed, a flagged query at position 4 is probably worth more than the zero in the clicks column suggests, and one with a poor average position is worth checking live before you decide what it’s worth.

What to change

These are recommendations that follow from the pattern in the queries. We haven’t tested them against citation outcomes yet; that’s the next study.

Write the sub-question, not just the head term. The queries in this group aren’t “skin tightening”. They’re “which skin-tightening device should a clinic buy for high patient acceptance”. Content that answers that specific question in its first sentence is at least eligible for it.

Cover the decision tree. If the device queries focus on training, warranty, room integration, staff turnover and post-GLP-1 protocols, that’s an outline of the questions being asked. Check how many of them your device page actually answers.

Own the persona questions. For clinics, the described patients are a content brief: no-downtime options for someone with a busy schedule, comfort-first treatments for the nervous, what to expect after weight loss. Written plainly, with the clinic’s own answer rather than a generic one.

Rank for the sub-question. Google’s own guidance says there are no additional requirements or special optimisations for AI features, and our sample is consistent with that in the least comforting way: in 37 live checks, no query outside the top 10 produced a citation, and position was the only variable we measured. Entity signals, author credentials and evidence on the page are our standing recommendations for the pages that do rank; they aren’t tested here.

Try it on your own site

You don’t need an API pull to see this. In Search Console, open Performance, add a Query filter, choose Custom (regex), and paste:

^(\S+\s+){6,}\S+$

That’s a looser version of Chris Long’s 10-word filter, set at seven words so it catches the shorter device-style questions in our data at the cost of more human long-tail to sift through. It returns every query of seven words or more. To get from that list to something like our flagged group, export it and do three more things:

1. Keep rows with 10 or more impressions and 0 clicks.
2. Keep rows with at least two of: starts with a question word; contains a conditional (if, rather than, without, instead of); contains specification vocabulary (recommended, suitable, criteria, professional, clinic, practitioner); ends in a question mark or contains a comma clause; has a double space or line break in it.
3. Read what’s left. The ones written from a role (“our clinic”, “for practitioners”) or about a described person (“someone who…”) are the ones to take seriously.

If you run it and find a pile of these, the question that matters is the one the filter can’t answer: when Google answers them, is it citing you or someone else? That’s what the next piece of this research is about, and it’s the check we run first on any clinic we work with.

What we’re watching

Google has said it will add metrics to the generative AI report over time. If a query dimension arrives, we’ll re-run this study against the official numbers and report how far the fingerprint was off, in either direction. Until then, treat the counts here as an estimate with error in both directions: Search Console suppresses rare queries before we see them (Suganthan’s export suggests that’s most of the pool), the fingerprint deliberately skips queries with one signal or under ten impressions, and our hand-labelling shows it also lets through some that a human reader wouldn’t flag.

This study only sees Google. ChatGPT, Perplexity and Claude fetch pages through their own crawlers and leave nothing in Search Console; the Medical Aesthetics AI Visibility Index and our own 314-clinic ChatGPT audit cover that side for aesthetics.

Fingerprint and method notes

A query is flagged when it has seven or more words, ten or more impressions, zero clicks, and at least two of these five signals, applied as case-insensitive regular expressions:

All patterns are case-insensitive.

Queries beginning “site:” are excluded. Data covers 24 April 2025 to 3 September 2026. Counts are property-query pairs: 565 pairs, 552 distinct strings. Two raters labelled 300 queries without seeing the automatic result: of 200 flagged, 167 machine, 5 human, 28 unsure; of 100 near-misses, 24 machine. The fingerprint infers likely machine origin from structure; it can’t prove it, can’t separate fan-out sub-questions from conversational follow-ups or pasted text, and Google’s documentation doesn’t say what’s recorded when a page is retrieved and not shown. Average position in Search Console isn’t a page number. Live checks are one day’s observation and AI answers vary between sessions.

Frequently Asked Questions

What is query fan-out?

When Google’s AI Overviews or AI Mode answer a question, they don’t run it once. They split it into several sub-questions, search for each of those, read what comes back, and write the answer from that. Google calls it query fan-out and says both features may use it. A model writes the sub-questions.

How do I find these queries in my own Search Console?

Open Performance, add a Query filter, choose Custom (regex), and paste ^(\S+\s+){6,}\S+$. You’ll get every query of seven words or more. Keep the ones with ten or more impressions and no clicks, then read them. Anything written as if a clinic is asking, or about a described type of patient, is worth a closer look. The full fingerprint is in the notes at the end of the article.

Does the new Search Console AI report show these?

It doesn’t. The generative AI report Google launched in June 2026 tells you how often links to your pages appeared inside AI Overviews and AI Mode, broken down by page, country, device and date. There’s no query, click or position data in it, so if you want to see the questions behind those impressions you still have to go through the ordinary Performance report.

Are zero-click impressions on these queries worth anything?

Sometimes. All 37 of the flagged queries we tested live produced an AI Overview, and the only ones that cited the site were the ones where it ranked in the top 10. So a flagged query at a good position is probably being read inside an AI answer, even though nobody clicks through. One at a poor position could be anything, and it’s worth running live before you decide.

Can you prove these queries were written by an AI?

No. Google doesn’t label them and we say as much in the article. What we can show is that they form a distinct group that doesn’t look like the long queries people click on, that they behave the way you’d expect fan-out queries to behave, and that every one we tested triggered an AI answer. We think that’s a strong case. It isn’t proof.