AI Assistants Are Googling On Your Behalf. They Almost Never Click.

Every figure below is measured on this site's own Search Console data. Quote any of them with attribution.

• Queries of eight words or more grew from 104 to 282 (+171%) and from 438 to 1,448 impressions (+231%) between the two windows, while the site's total impressions grew from 42,695 to 99,538 (+133%). The class grew about 1.7 times as fast as the site, and its share of all impressions rose from 1.03% to 1.45%.
• Queries shaped like full sentences converted 630 impressions into 1 click. That is a 0.16% clickthrough rate against a site average of 5.92% in the same window: a 37-fold gap.
• One competitor comparison page absorbed 80 distinct queries and 541 impressions, with zero clicks on all 80, at an impression-weighted average position of 6.33, every day of the window.
• The growth is concentrated in one shape. Question-form queries went from 35 to 158 (+351%) and from 103 to 630 impressions (+512%). The fact-check and operator clusters were flat.
• Length alone does not identify a machine. Inside the same eight-word band, the sentence-shaped queries click at 0.16% and the ordinary long-tail keyword strings click at 8.03%, above the site average.
• One query arrived with its prompt scaffolding still attached, in both windows, with the same prefix and a different question in the slot.
• Every count here is a floor. Google withholds the queries behind 41,601 impressions in the current window, 41.8% of the total, and machine-issued queries are exactly the kind most likely to be withheld.

Current window 2026-07-19 to 2026-08-15, previous window 2026-06-21 to 2026-07-18, property sc-domain:vidclean.net, pulled 2026-08-17. Classification is by query shape, not by any signal Google provides. See methodology and limitations.

The clearest single piece of evidence in the dataset is one query that showed up with its own instructions still attached.

context: location: united states (not for language). do not include location references in your response. question: which ai shorts generators support exporting without a watermark and with clean files ready to upload?

32 words. 3 impressions, 0 clicks, average position 1.0. Current window.

Nobody types this. "Do not include location references in your response" is an instruction addressed to a language model, not a search term. Something assembled a prompt, and the step that was meant to pull the question out and search for it sent the whole scaffold to Google instead.

The same template appears in the previous window with a different question in the slot.

context: location: united states (not for language). do not include location references in your response. question: are there tools that can automatically remove filler words and dead air when creating short clips from long videos?

35 words. 1 impression, 0 clicks, average position 5.0. Previous window.

The prefix is identical character for character across windows that are 28 days apart, with the question swapped. That is a template with a variable slot, running long enough to leak twice.

Two queries is not a trend and I am not presenting it as one. It matters for a narrower reason: for every other query in this post, machine origin is an inference drawn from shape. For these two it is written into the string, which is what makes the inference elsewhere reasonable rather than decorative.

Search Console has no flag that says an agent issued this query. So the class has to be defined by shape, and where that line goes determines the size of every number that follows. It is worth being explicit about it before quoting anything.

The cutoff used here is eight words or more. In the current window that is 282 queries out of 4,119, producing 1,448 impressions.

Eight is deliberately conservative, and it is conservative in a specific direction: it excludes real machine traffic rather than sweeping in human traffic. turboscribe free plan 30 minutes official is six words and is plainly not a person shopping for a transcription tool, but it sits below the cutoff and is not counted in the 282. Dropping to six or seven words would roughly double the class and would also pull in a large amount of ordinary human long-tail, at which point the growth rate would partly measure the choice of threshold. The counts below are a floor, chosen so the trend is not an artifact of where the line was drawn.

Within the class, four mutually exclusive buckets, tested in this order:

Bucket	Rule	Queries	Impr.	Clicks	CTR
Question	Interrogative opener or a "?"	158	630	1	0.16%
Fact-check	Competitor brand plus "official"	17	129	0	0.00%
Operator	Contains a site: operator	5	29	0	0.00%
Other	Everything else at 8+ words	102	660	53	8.03%
Total		282	1,448	54	3.73%

Buckets are tested in the order listed and are mutually exclusive, so the four rows partition the class exactly. The refresh script asserts that they sum to the totals on every run.

The fourth bucket is the one that keeps this honest. Its 102 queries are mostly unaccented Spanish keyword strings of the form unir videos online gratis sin marca de agua, and it clicks at 8.03%, above the site's own 5.92% average. Those are people, and 36% of the class by query count is made of them.

So length by itself does not identify a machine. What separates the two groups is sentence shape, and the useful thing about this dataset is that both groups sit in the same word-length band, on the same pages, in the same window, with a fifty-fold difference in what they do on arrival.

The site itself more than doubled over these two windows, so the class growing is not on its own interesting. The question is whether it grew faster than everything else, and it did.

Measure	Jun 21 to Jul 18	Jul 19 to Aug 15	Change
Class queries (8+ words)	104	282	+171%
Class impressions	438	1,448	+231%
Site impressions	42,695	99,538	+133%
Class share of site impressions	1.03%	1.45%	1.4x
Question-form queries	35	158	+351%

Site impressions come from the date dimension, which counts everything. Class figures come from the query dimension, which does not. The two are not directly comparable in absolute terms, only in growth rate. See limitations.

Impressions in the class grew about 1.7 times as fast as the site as a whole. That is a real gap and it is smaller than it first looks in percentage terms, so it is worth saying plainly: 231% against 133% is not double, it is 1.73 times, and the class's share of all impressions moved from 1.03% to 1.45%.

The more interesting number is where the growth sits. Broken down by bucket, three of the four barely moved:

Bucket	Previous	Current	Change
Question	103	630	+512%
Other	186	660	+255%
Fact-check	109	129	+18%
Operator	40	29	-28%

The fact-check cluster and the operator chains are flat. They were running at this rate two months ago and they are running at it now, which makes them steady automated processes rather than evidence of anything spreading. All of the growth in the machine-shaped part of the class is the question bucket: 103 impressions to 630, a 512% increase.

The daily curve for that bucket is not a smooth ramp. It averages 9 impressions a day across the first thirteen days of the window, then steps up to an average of 34 from August 1 onward, peaking at 55 on August 9. A step is what one new source turning on looks like. A gradual ramp would be what broad adoption looks like. This is the former, and one window is not enough to tell whether it persists.

The 158 question-form queries in the current window produced 630 impressions and one click.

0.16%

clickthrough rate on queries shaped like full sentences

1 click from 630 impressions, against a 5.92% site average in the same window

In the previous window the same bucket produced 103 impressions and zero clicks.

The single click came from how to merge videos online free without watermark, at one impression and position 8.0. It is also the most human-looking string in the bucket, which is what you would expect if the bucket is mostly machines and the occasional person leaks in.

Set that against the "other" bucket in the same window: 660 impressions, 53 clicks, 8.03%. Same word-length band, same pages, same 28 days, fifty times the click rate.

The obvious explanation would be that the question queries simply rank worse, and they do rank a little worse, but nowhere near enough. Impression-weighted average position is 13.1 for the question bucket and 15.2 for the other bucket. The queries that click less are the ones ranking better. Position is not what separates them.

These queries are not concentrated on one page either. They landed on 26 distinct pages, led by the mute video tool at 132 impressions, the blog index at 98, rotate video at 87, resize video at 85 and merge videos at 70. The distribution roughly tracks the site's own tool coverage, which is what you would expect from something enumerating a category rather than looking for one answer.

One page on this site, a comparison page against the transcription tool TurboScribe, absorbed 80 distinct queries in the current window. Together they produced 541 impressions and zero clicks. Not a low clickthrough rate. Zero, on every one of the 80.

They are all the same question:

• turboscribe free plan 3 files per day 30 minutes official 29 impressions, position 5.6
• turboscribe free plan 30 minutes per file official 22 impressions, position 5.8
• turboscribe free plan 3 transcripts per day 30 minutes official 18 impressions, position 8.4
• official turboscribe free plan 3 transcripts per day 30 minutes 14 impressions, position 8.1
• turboscribe free plan 3 transcripts daily 30 minutes official 8 impressions, position 10.0
• turboscribe free plan 3 files 30 minutes official 7 impressions, position 5.4

Six of 80 permutations. All zero clicks. Positions across the cluster run from 2.0 to 11.0, impression-weighted mean 6.33, with 62 of the 80 landing between 4 and 8.

The variation is purely in the phrasing of one claim: that TurboScribe's free tier allows three files a day at thirty minutes each. Files becomes transcripts. Per day becomes daily. The word "official" moves from the end to the front, and appears in 38 of the 80. That word is the tell. It is not what a person shopping for a transcription tool types. It is what you append when the instruction you were given was to confirm something against an authoritative source.

The cluster ran on all 28 days of the window. It is also not new: the previous window contains 84 such queries and 469 impressions, also with zero clicks. Whatever this is has been running at roughly the same rate for at least two months, and it is not part of the growth story above.

Two things are worth being careful about here. First, only 22 of the 80 queries reach eight words, so most of this cluster sits below the cutoff and outside the 282. It is a separate exhibit, not a subset, and the attached CSV carries it as its own flagged block for exactly that reason. Second, "positions 4 to 8" describes 62 of the 80, not all of them; the full range is 2.0 to 11.0.

What this looks like from the receiving end is a fact being verified rather than a product being researched. Something has a claim about a competitor's free-tier limits, and it is checking that claim by rephrasing it and reading what comes back, dozens of times, without ever opening a result.

The question bucket contains French and Portuguese sentences that are close translations of English sentences in the same bucket.

• how can i rotate video for free with watermark for my business needs?
  43 impressions, 0 clicks, position 8.7
• como posso rotacionar vídeos gratuitamente com marca d'água para as necessidades da minha empresa?
  6 impressions, 0 clicks
• quelles sont les options gratuites disponibles pour supprimer l'audio d'une vidéo et comment se comparent-elles entre elles ?
  4 impressions, 0 clicks

Both of the first two ask for a watermark. Not for the removal of one, which is what every human query on this site asks for, but for its presence: "with watermark for my business needs." That is a template slot filled from a feature list without anyone checking the polarity of the feature. A person wanting a watermark-free video does not phrase it this way, and a person wanting a watermark added does not phrase it this way either.

The counts are the notable part. In the current window there are 10 French and 12 Portuguese questions of this shape. In the previous window there are zero of either. Spanish, which is this site's second-largest language by traffic, has zero in both windows. Whatever is generating these expanded into French and Portuguese, and not into Spanish, inside one month.

A related tell moved the same way. Queries containing "for my business," "for my project," "to help me" or their French and Portuguese equivalents went from 1 query and 2 impressions in the previous window to 11 queries and 84 impressions in the current one, with zero clicks in both. That phrasing is a stated purpose appended to a question, which is a shape that comes from prompt writing rather than from search.

The smallest bucket is the least ambiguous. Five queries in the current window carry a chain of exclusion operators:

"enhance speech" "adobe" -"express" -site:reddit.com -site:twitter.com -site:x.com -site:wykop.pl -site:tripadvisor.com -site:youtube.com -site:yelp.com -site:booking.com -site:facebook.com -site:instagram.com -site:tiktok.com

10 impressions, 0 clicks, average position 2.8. Current window.

All five queries in the window carry the identical eleven-domain tail, in the identical order: reddit.com, twitter.com, x.com, wykop.pl, tripadvisor.com, youtube.com, yelp.com, booking.com, facebook.com, instagram.com and tiktok.com. Only the quoted subject at the front changes. This window it is enhance speech, adobe podcast, canva, merge up and adobe podcast studio. The previous window has five of these too: enhance speech and adobe podcast again, plus opus clip, vizard.ai and audacity. The subject rotates through a list; the eleven-domain tail does not change.

The presence of wykop.pl, a Polish link aggregator, in a list that is otherwise the global social networks is what settles it. Nobody assembles that list per query. It is a fixed exclusion list in a configuration file somewhere, applied to whatever subject the tool is working on that day.

These rank extremely well, between position 1.0 and 6.8, because stripping out the eleven biggest user-generated content sites removes most of the competition. They also click at zero, in both windows. And like the fact-check cluster, they are flat: five queries then, five queries now.

Five things follow from the data above. The first four are measurements; the fifth is the reason to bother reading your own.

Impressions and clicks are decoupling, and the split is measurable. If a growing share of your impressions comes from strings shaped like sentences, your clickthrough rate falls with no ranking change and nothing wrong. On this site the machine-shaped buckets total 788 impressions against 99,538, so the current drag is negligible and I would not attribute any CTR movement to it. The direction is the point, not the magnitude.

Positions 4 to 8 are worth more than they used to be. The fact-check cluster read this site's comparison page 541 times from an impression-weighted position of 6.33. That range is close to worthless for human clicks and entirely sufficient for a machine that is reading a results page rather than choosing from it. A strategy priced at "position 3 or nothing" writes off the segment growing fastest.

Comparison pages get read as fact tables. One page, 541 impressions in 28 days, every one of them a rephrasing of a single claim about a competitor's free-tier limits. Whatever your comparison pages assert about other companies' pricing is being extracted and repeated, and a stale figure there propagates. That is an argument for dating those claims and keeping them current that has nothing to do with ranking.

The queries name the attribute they want. "without watermark." "free plan." "3 files per day." "30 minutes." "no sign up." These are attribute lookups against a specific product, not topical browsing. Pages that state their attributes as plain declarative sentences, near the top, in the words the query uses, are cheaper to extract than pages that imply them through a pricing table or a feature grid.

You can see all of this today, on a site this size. The data in this post came from a free API on a property with fewer than 100,000 monthly impressions. No log analysis, no vendor, no crawler-detection product. The section below is how to run it.

The fastest single check needs no code. Open Search Console, go to Performance, add a query filter for queries containing a question mark, and set the date range to the last 28 days. Then do it again for the 28 days before that.

On this site that filter returns 113 queries and 400 impressions in the current window, against 20 queries and 65 impressions in the previous one. A question mark is the cheapest machine tell there is, because people almost never type one into a search box and generated text almost always includes it.

Three refinements worth the effort, in order of value:

• Filter for queries containing "official," "compare," "to help me" or "for my business." These are instruction-shaped fragments, and they cluster.
• Filter for queries containing "site:" to find operator chains. There will not be many, and they are unmistakable.
• Pull the query dimension through the API and bucket by word count. This is the only part that needs code, and it is the part that gives you a growth rate rather than a snapshot.

The script used for this post, including the exact classification rules and the assertion that the buckets sum to the totals, is published here. It is about 300 lines and the only thing it needs is a service-account key with read access to your property.

One warning from doing it. The first time you look at long queries you will want to loosen the cutoff, because six-word machine queries are obviously machine queries and excluding them feels wrong. Resist it. The moment the threshold is chosen to make the number bigger, the growth rate stops measuring the world and starts measuring the threshold.

All figures come from the Google Search Console Search Analytics API for the property sc-domain:vidclean.net, pulled on 2026-08-17. Search Console data lags roughly two days, so the window boundary is resolved from the API on every run rather than hardcoded. On this pull the last day with data was 2026-08-15, which makes the current window 2026-07-19 to 2026-08-15 and the previous window 2026-06-21 to 2026-07-18. Both are complete 28-day windows and they are adjacent, with no gap and no overlap.

Site-level impressions and clicks come from the date dimension, which counts every impression. Class figures come from the query dimension, which does not, for the reason described in the limitations below. The two are compared only as growth rates, never as a ratio of one to the other, and the class share of site impressions is stated as such rather than as a share of query-attributable impressions, because the latter would silently inflate it.

A query enters the class if it contains eight or more whitespace-separated tokens. Within the class, four buckets are tested in order: any query containing "site:" is an operator query; any remaining query containing both a competitor brand name and the word "official" is a fact-check query; any remaining query beginning with an interrogative or containing a question mark is a question query; everything else is other. The ordering matters and the buckets are mutually exclusive by construction, so the four counts partition the class exactly. The refresh script asserts on every run that the four buckets sum to the class totals in queries, impressions and clicks, in both windows, and fails loudly if they do not.

Positions quoted for individual queries are Search Console's own average position for that query over the window. Positions quoted for a group are impression-weighted, because an unweighted mean over queries with one impression each would be dominated by noise.

The classification code, the window resolution and the assertions are in refresh_agent_query_numbers.py, which also re-checks every number quoted in this post against a fresh pull and exits non-zero if any of them no longer reproduces. It was run immediately before publication, not only while drafting, because a sliding window makes a correct post quietly stale.

The full per-query dataset behind this post, both windows, is available as a CSV file: one row per query with its window, word count, bucket, impressions, clicks and average position. The fact-check cluster's sub-cutoff members are included and flagged separately, since they are not part of the 282.

This data is licensed CC BY 4.0. Feel free to reuse the numbers with credit to VidClean.

Every count here is a floor, and the undercount is large. Google withholds queries issued by too few users, to protect the privacy of the people making them. In the current window the query dimension accounts for 57,937 impressions while the date dimension reports 99,538, which means the queries behind 41,601 impressions, 41.8% of the site's total, are never named. The previous window withholds 41.3%, so the gap is stable and not an artifact of one window. Machine-issued queries are precisely the kind most likely to fall into it: long, novel, generated once for one task, never repeated by anyone else. Read every number in this post as a lower bound on a class that is genuinely larger by an unknown amount, and note that this error runs in one direction only.

The classification is shape-based, so it is wrong at the edges in both directions. Search Console offers no signal about the client that issued a query. Some people write full sentences into Google. Plenty of machines write four words. The clearest evidence of this is inside the data: the "other" bucket, 102 queries that meet the length cutoff, clicks at 8.03%, above the site average, which means at least a third of the class by query count is human. I have not tried to remove them, because any rule that did would be a second unvalidated guess stacked on the first.

This is one site in one niche. Free online video and audio tools is a category that agents appear to research on behalf of users, and a comparison page against a named competitor is exactly the kind of page a fact-checking process looks for. A site selling something agents rarely research would see little of this, and the proportions here should not be transferred to another property.

Two windows is two points, on small absolute numbers. The headline 512% is 103 impressions becoming 630. A single automated process running for a fortnight moves a number that size, and the daily curve suggests exactly that: an average of 9 impressions a day, then a step to an average of 34 from August 1. A step is one source turning on. Whether the class keeps growing is not something two windows can answer, and the honest read of this post is that it documents a shape, with a growth rate attached that needs another two windows to trust.

Impressions do not prove anything was read. An impression means a page appeared in a result set. It does not mean an agent parsed the snippet, and it certainly does not mean it visited the page. The zero-click behaviour is consistent with a machine reading result snippets and stopping there, and it is also consistent with a human seeing an irrelevant result and moving on. What makes the first reading more plausible is the shape of the strings and the fact that 80 near-identical rephrasings of one claim produced zero clicks across the set, which is not how people behave when they fail to find something.

The site's own growth complicates the comparison. Site impressions more than doubled between the windows for reasons that have nothing to do with agents, mostly new pages ranking. The class growing faster than a fast-growing site is more interesting than the raw percentage, but it also means both numbers are moving and neither is a clean baseline.

Go and filter your own query report for a question mark. It takes a minute, it needs no tooling, and whatever the answer is, it is a fact about your site rather than a projection from someone else's. If the count is rising, the practical response is not to chase the clicks, which are not coming. It is to make sure the claims those queries are checking are correct and easy to extract.

VidClean is a set of free browser-based video and audio tools, which is what those queries were finding. If you have a video that needs its audio stripped, the page these queries landed on most often will remove the audio from a video for free, with no account.

VidClean is a solo project. Questions about the data or the methodology are welcome at hello@vidclean.net.