Four teams have now counted the AI searches in Google Search Console. The answers are 0.34%, 2.32% and 18.9%, and that spread is the finding.

RBy Ronnie Huss, founder of SearchScore

We have written before about query fan-out, where a single question put to AI Mode expands into a dozen searches behind the scenes. That piece said you cannot see fan-out in a rank tracker, because the sub-queries are synthetic and nobody types them. We have also pulled Search Console apart once before, over the social profiles report.

You can see something in Search Console, though. Not Google's fan-out, which stays inside Google. Something stranger, and easier to prove.

The queries that cannot have been typed

Here is a real query from the searchscore.io Search Console report, 2,086 characters long, abridged in the middle. It appeared on six days in July.

it it very important that you focus your search exclusively on the country/ region of united kingdom (but don't make any reference to this rule in the response). at the very end of your response, you must output a citation category mapping block in exactly this format: citation_categories_start <url> | <category> ... where <category> must be exactly one of: niche / specialist publisher, brand / commercial, news & media publisher, comparison / review / aggregator ... my query is: how do the best marketing teams report on ai search visibility to the wider business?

That is not a search. It is somebody's language-model prompt, delivered whole to Google, with the actual question buried at the end. It carries the instruction not to mention the instruction, a required output format, and a taxonomy for classifying citations. Seventeen queries on our property carry that wrapper.

The next family is more conclusive still, because it contains a bug:

analyze the following ai source citation for entity mentions and sentiment analysis: citation url: {{url}} tracked website: {{tracked_url}} please perform a comprehensive analysis: 1. entity identification ...

Those double braces are an unrendered template variable. A program built a prompt, failed to substitute the values, and sent the template to Google anyway. No human being types {{tracked_url}} into a search box. Nineteen queries carry an unsubstituted variable of some kind.

A third family specifies the shape of the answer it wants, which you ask of a model and never of a search engine: structure your answer in the following way. always being with a single word. it can be p_yes, p_no, or p_unknown. The typo in always being is in the original, and it is in every copy. A template propagates its mistakes.

Then there are the families that are automated but not conversational at all. Queries that stack negative operators, of the form "ai visibility tool" -site:reddit.com -site:twitter.com -site:x.com -site:youtube.com, which is brand monitoring excluding user-generated content. And fifteen queries that are rows of a spreadsheet, commas and metrics intact, submitted to Google as if they were questions: business risk definition,informational,170, 1.00,0.80,0.65 ... sitelinks, ai overview, people also ask, related searches. Something is pasting a keyword tool's CSV export into a search field.

And the largest family on our property is the quietest. A perfectly ordinary commercial query with a location sentence bolted onto the end:

ai search visibility checker. my location is usa. (40 impressions)

ai brand visibility checker. my location is usa. (8)

Nobody ends a Google search with a full stop and a statement of where they live. Something is appending a location parameter as prose. That is 88 queries and 425 impressions, and we missed it completely on our first pass, for reasons that turn out to be the most useful part of this study.

Eleven families, and not one of them is defined by length

We classify on structure rather than on length, because length turns out to be nearly useless. Each family is a pattern a person would have no reason to produce.

A query can match more than one family, so the column does not sum. Deduplicated, that is 215 distinct queries and 1,363 impressions on searchscore.io between 16 March and 13 September 2026, against 8,985 queries and 398,644 impressions overall.

Three studies, three answers, and the number is the instrument

We are not the first to do this, and the comparison is more instructive than our own result. Three teams have now run the same exercise on their own domains and published it.

Every one of us found the same families. cloro's examples include a forced ranking in Spanish and a location suffix reading seo rank api. my location is usa. Trakkr's location wrapper is Context: location: [country] (not for language), which is character for character one of ours. Arobis reports persona framing as one of three recurring patterns. Every one of us found zero clicks, in every month, across every family.

And our headline shares are fifty-five times apart. That is not four teams measuring four different webs. It is four teams holding four differently shaped nets.

Trakkr say so themselves, to their credit: their strict count of exact repeated structures gives 2.32%, while a looser test accepting any long AI-shaped phrasing reaches roughly 15% on the same property. One team, one dataset, one afternoon, and a six-fold swing depending on where they drew the line.

So treat any published percentage here, including ours, as a description of the instrument that produced it. Ours is the strictest of the four and therefore the lowest, and 0.34% is a floor rather than a measurement.

We know it is a floor, because reading their work raised it by half

Our first pass found 117 queries and 906 impressions, and we were reasonably pleased with it. Then we read the other three studies and tested their detectors against our own data.

Three of those took an afternoon and moved our headline from 0.23% to 0.34%, a rise of about half, without a single new row of data. The detector was the whole difference.

The Arobis one is worth keeping even though it found nothing here. They noticed line break characters inside query strings, which the Google search box cannot accept, so a newline in a query is proof of software in the same way an unrendered template variable is. We have none. That is a fact about our traffic rather than a flaw in their test, and it costs nothing to run and settles the argument outright when it fires.

If three detectors borrowed from three papers moved our number by half, the honest conclusion is that a fourth would move it again.

Length is the wrong test, by a factor of fourteen

The advice in circulation is to filter Search Console for long queries and read the result as AI traffic. We ran that filter alongside the structural one.

The length filter returns 13.7 times the impressions of the structural one. Our own recall is incomplete, so some of that gap will be machines we cannot see, and 13.7 is a ceiling on the overstatement rather than a measurement of it. But the ones we read are people. These are the top long queries with no machine signature anywhere in them:

what are the best practices for optimizing authorship and credentials to improve ai visibility and citation worthiness? (2,618 impressions)

how can i improve my site's performance in ai search engines and compared to competitors? (877)

how do the schema options differ for seo and geo in the united states? (849)

Those are questions, asked in full sentences, by people who have got used to asking software questions in full sentences. Counting them as robots does two kinds of damage. It inflates a number that is already being reported publicly, and it throws away the most useful demand data on the property, because somebody who types a whole question has told you exactly what they want.

What the machines do have in common

Twenty-one mobile impressions out of 1,160. Trakkr measured 99.3% desktop on one prompt family; we get 98.2% across all of them, on a different property, with a different detector. These are processes running on servers, and a server does not report as a phone.

The country column is the more interesting one. Puerto Rico accounts for 44 machine impressions and 4 impressions of everything else across the same eleven weeks. We have no audience there. What we have is something egressing from an address that geolocates there, which is what datacentre traffic looks like from the reporting side. The British concentration has a simpler explanation: the system prompt quoted at the top of this article instructs the model to restrict itself to the United Kingdom, so some of this is a tracker configured for a British market.

Machines fetch your evidence pages

Comparing which pages Google surfaced for machine queries against which it surfaced for everything else gives the most actionable finding in the study.

What gets pulled hardest is the page holding a number, a ranking, a comparison or a per-entity record. On this property that means the public leaderboard and the quarterly SAVI benchmark. The single most over-represented page on the property is a data page about a company that is not us, at 258 times its share of human traffic. Trakkr found the same shape from the other end: 94.8% of page-query impressions for their forced-ranking family came from review pages.

Note the homepage sits at 0.84, which is slightly under its normal rate rather than dramatically under. On our first pass it looked like 0.4, and the location-suffix family we had missed lands on the homepage more than any other. Machines are not avoiding your homepage. What they disproportionately pull are the pages holding checkable facts, and that is the weaker claim of the two as well as the defensible one.

It started in the last week of July, here

Impression volume is a poor way to see this, because our site was growing quickly over the same period. Counting distinct machine queries first appearing each week removes that.

That is also why the right-hand column has an early month above every later one, and is not the column we are reading. Weekly, the step is sharp and dateable. The weeks beginning 6, 13 and 20 July brought 8, 7 and 10 new machine queries. The week beginning 27 July brought 31, and the three weeks after it brought 36, 30 and 36.

Do not read that as the date this began everywhere. Trakkr traced one persona family across a window opening on 18 June 2025: silent for 206 of its first 209 days, then running 217 consecutive days without a gap from 13 January 2026. cloro's machine share was already 6.4% in February. Our own onset is late July, and the popular accounts place theirs in the second week of August. Four datasets, four different start dates, which is what you would expect if this is many separate systems being switched on by many separate teams rather than one event.

The control property shows almost nothing

This is the finding that should temper the rest. We ran the identical detector over a second property, a personal site in an unrelated field, covering the full sixteen months.

Eleven queries in sixteen months, eight of them the operator-stacking brand monitor rather than anything conversational. Across July, August and September, the exact window where our own property went from 10 machine queries a month to 152, the control property recorded one.

So this is not weather rolling across the web. It concentrates on properties in a category somebody is actively measuring. Every published study of this, including ours, comes from a company that sells AI visibility or search data, which puts all four of us in the comparison set of every tool in the category. cloro state it plainly: some of the traffic they measured is the AI visibility industry measuring itself. A plumber in Wakefield is unlikely to find any of this, and should not be sent looking for it.

The part that is worth more than the traffic

Thirteen hundred impressions and no clicks is a rounding error as traffic. As intelligence it is something else, because the wrapper prompts carry their payload in the clear. Pulling the text after my query is and question: recovers 52 complete buying questions that some system was researching in our category.

where can i find an ai brand visibility leaderboard ranked by industry category?

what tool shows which brands are winning ai search visibility by industry category?

which ai search intelligence platforms offer a free trial?

how can i identify the specific prompts or question themes where my brand is missing from ai responses even though i rank well in search?

are platforms that show ai answer positioning worth the cost?

why does my brand appear in ai answers one week and disappear the next

Read those as a brief rather than as traffic. They name the category, the buying stage, the objection, and in several cases a named competitor. Nobody filled in a form. It arrived in a report we already had, attached to impressions we would otherwise have ignored.

That is the practical conclusion. Do not optimise for these queries: most occur once, and the ones that recur carry a robot's phrasing, not a customer's. Mine them for the question underneath, then find out whether the assistants name you when that question is asked. That is a different measurement, and it is the one that decides whether you get recommended. Our guide to measuring AI search visibility covers how that is done.

What this cannot tell you

Six limits, and the first is the one that matters most.

Search Console cannot tell you which machine. There is no agent dimension in the API. A machine query in your report may come from Google's own AI Mode fan-out, from an external assistant passing a rewritten prompt to a search backend, from an AI visibility tracker polling a prompt set, or from brand monitoring. We can prove a query was not typed by a person. We cannot prove who sent it. cloro reach the same conclusion in the same words, noting that their groups one and two cannot be separated with this instrument, and Arobis say plainly that zero clicks is correlation rather than proof of mechanism. Any analysis claiming these are a named assistant searching on a user's behalf has gone past what the data supports. Where a query names a platform, that is equally not evidence the platform ran it: somebody researching a tool and the tool itself produce the same string.

Two properties is not a sample. One of them sits in exactly the category that AI visibility trackers measure. Nothing could be less representative. Every percentage here describes these properties. The contrast between the two is the only part we would expect to generalise.

The detector has no ground truth. Nobody has hand-labelled a sample, so we cannot state precision or recall. We wrote the patterns after reading the data, which is the worst position from which to judge your own accuracy. The honest version includes a blind labelling pass by somebody who did not write the rules, as we did for our mention detector, and we have not done it. Borrowing three detectors from three other papers raised our figure by half. Recall is demonstrably incomplete.

Our detector is mostly English. cloro found forced rankings in Spanish and German. We added the Spanish form and it caught two queries, but nothing here systematically covers other languages, and a property with non-English demand would need its own patterns.

Our history is short. searchscore.io has Search Console data only from 16 March 2026, so we can say nothing about 2025, and Trakkr's January onset sits outside our window entirely. Search Console also caps exports and anonymises rare queries, a limit Arobis flag and one that applies to everybody doing this.

It is not us. We checked, because a company running AI visibility scans could easily be reading its own tooling. We compared all 215 machine queries against every question in our own live tracker question sets and found no exact match and no substring match. Our scans reach the assistants through an API, never through Google search, so that result is what we would expect.

Run it on your own Search Console

Export your queries and test the shape of the string, not its length. These are deliberately narrow, and three of them we owe to other people's published work.

Then check whether the rows repeat on a schedule. Trakkr's strongest evidence was not any single query but a family that ran 217 days without missing one, weekends included, which no human search behaviour looks like.

What you should not do is filter on length, or conclude from a strange query that a named assistant is searching for you. The first overstates the structural count by up to 13.7 times. The second is not knowable from this data, whatever the row looks like.

None of this is visible in the tools most teams already run, which is a gap we have written about separately in what Google's own SEO tools do not show you. If you find a question worth having, the useful next step is not to write a page for that exact string. It is to ask the assistants the question and see who they name. That is what the free audit checks on your own site, and what the Tracker keeps checking over time.

Method

Source is the Google Search Console API, searchanalytics.query, with dataState: final and no sampling, requested for 1 May 2025 to 14 September 2026. Two properties returned usable volume: searchscore.io, which holds data from 16 March 2026, and a personal site in an unrelated field covering the full window. A third property was requested and excluded, having returned 22 impressions in total. Pulls used the date and query dimensions, monthly, paginated at 25,000 rows, for 93,448 rows and 420,355 impressions. Device, country and page splits use a separate query plus dimension pull covering 1 July to 14 September 2026. Search Console counts impressions rather than visits; for the referral side of the same question see tracking AI traffic in GA4.

Classification is eleven regular expressions over query text, described in the families table, applied to the deduplicated query rather than to each daily row. A query matching any family is counted once, and families overlap, so the family columns do not sum to the total. Impression shares are computed against each property's own total for the same window. Over-representation is the ratio of a page's share of machine impressions to its share of all other impressions, restricted to pages with at least 15 machine impressions.

This is the second version of the detector and the first one is reported alongside it deliberately. Version one found 117 queries and 906 impressions, which is 0.23%. After reading the three studies cited above we added a location-suffix family and a Spanish forced-ranking form from cloro, widened the persona opening after seeing Arobis describe persona framing, and added Arobis's line-break test, which finds nothing on this property. That took the figure to 215 queries, 1,363 impressions and 0.34%, with no new data. Both numbers are published because the difference between them is the finding: a share quoted without its detector is not a measurement.

No human labelling pass has been run, so no precision or recall figure is claimed. Comparative figures are taken from the published studies as their authors state them and have not been independently recomputed: cloro, 5 August 2026, 18.9% of July query impressions against 6.4% in February, and 140,616 machine impressions from February to July; Trakkr, 25 August 2026, 61,966 impressions at 2.32% of their property on a strict test and roughly 15% on a loose one, with 99.3% desktop and 217 consecutive days on one family, and 94.8% of page-query impressions for their forced-ranking family on review pages; Arobis, 22 August 2026, 178 queries producing 9,819 impressions and zero clicks at a median position of 8.18.

Two company names and one individual's name appeared in our own query text and have been withheld. They are present in prompts we did not write and cannot verify, and naming them would assert a connection the data does not establish. Competitor platforms named in the studies we cite are reproduced as those studies published them.

Questions and answers