One click in 136,201 conversational search impressions

Across six web properties over 28 days, search queries of eight words or more drew 136,201 impressions and produced one click.

Not a low click rate. One click. The query was "how does perplexity choose which sources to cite," it appeared 52 times at an average position of 1.8, and somebody clicked it once.

Over the same window, queries of three words or fewer drew 70,693 impressions — roughly half the volume — and produced 175 clicks, while ranking at an average position of 13.33 against the long band's 6.46. The band that ranks better converts at roughly one two-hundred-and-sixty-fifth of the rate.

This is the query-side companion to what the Machine Relations Index measures on the source side. The Index records which domains answer engines cite. This measures what the person does after the machine has answered, and across this dataset the answer is that there is frequently no person.

What was measured #

The dataset is a Google Search Console query-and-page export covering six properties: authoritytech.io, machinerelations.ai, jaxonparrott.com, paralax.ai, paralabs.ai and christianlehman.com. The window runs 2026-08-18 through 2026-09-15, a complete 28 days, exported 2026-09-18 with Search Console's customary three-day reporting lag observed. It contains 16,535 query-page rows carrying 282,078 impressions and 212 clicks, an aggregate click-through rate of 0.075%.

Every row is a query-and-page pair with its impressions, clicks and average position as Search Console reports them. Rows were bucketed by the word count of the query string. No sampling, no weighting, no exclusions.

Two limits belong here rather than in a footnote, because they bound every number that follows.

First, Search Console withholds queries issued by very few users, so the query dimension of any export is a subset of a property's real demand. This measurement therefore describes the named-query layer of these six properties, not the whole of their search demand. The withheld queries are by definition the rarest ones, which is where genuinely novel conversational phrasings would sit — so the conversational share reported here is more likely understated than overstated.

Second, word count is a proxy for conversational shape, not a detector of machines. A long query can be a person typing a full sentence. The subset that is provably machine-issued is counted separately below, and it is small.

The bands #

Three things in that table matter more than the headline.

Ranking is not the constraint, and the relationship runs backwards. The 8–14 word band holds the best average position of any band, 6.46, and places 34.5% of its impressions inside the top three results — more than double the top-three rate of the short-query bands that take almost all the clicks. Of the 136,201 impressions on queries of eight words or more, 41,415 were in Google's top three and 16,004 were at position 1.5 or better. They returned one click. A remedy aimed at rank or at title copy has nothing to fix here; the ranking is already there.

The volume is not marginal. Queries of eight words or more are 48.3% of all impressions in the window. This is not a long-tail curiosity at the edge of the report. It is approximately half of everything these properties were shown for.

The drop is not gradual. Between the 4–7 word band and the 8–14 word band, impressions rise, average position improves by more than five places, and clicks fall from 36 to 1.

The counterfactual #

A click rate near zero is easy to dismiss when the absolute click count is small. The test is what the volume should have produced.

At the 4–7 word band's own click rate of 0.048%, the 136,201 impressions on longer queries would have returned approximately 65 clicks. At the aggregate rate of all queries under eight words, 0.145%, they would have returned approximately 197.

They returned one.

Question shape gives the same answer #

Word count is a crude instrument, so the same dataset was cut a second way: rows whose query opens with an interrogative (who, what, when, where, why, how, which, can, does, do, is, are, should, will) or contains a question mark, against everything else.

Question-shaped demand is essentially half of these properties' impressions. It converts at roughly one tenth the rate of everything else, from an average position more than four places better.

Where the clicks actually went #

Of the 212 clicks in the window, 124 — 58.5% — came from queries containing a brand name belonging to one of these six properties. Those brand queries account for 3,253 impressions, 1.2% of the total.

So the click behaviour of this dataset, stated plainly: the majority of clicks come from people who already knew the name and were navigating to it. The informational demand, which is the overwhelming majority of the impressions, produces almost nothing.

The page-level view says the same thing. 555 distinct pages received at least one impression from a query of eight words or more. Of the 40 pages that received 500 or more such impressions, 39 received zero clicks from them.

These are the seven pages that took the most impressions from queries of eight words or more, in order, with no rows omitted between them.

A page answering five hundred distinct questions, several thousand times, to no visitor at all, is a different object from a page that ranks badly. It is being read by something that does not click.

The fifth row is the one that rules out an easy explanation. It took 6,542 impressions from three queries, not from hundreds — so this is not a page thinly spread across a wide tail of rare phrasings. A handful of long questions were asked of it repeatedly, it was shown, and nobody arrived. Breadth of demand and concentration of demand produce the same outcome here.

The machine-issued slice, in proportion #

Part of this traffic is provably not human, and the mechanism is documented by Google rather than inferred. Google's guidance on AI features and your website states that AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources — to develop a response." Those issued searches are searches. They generate impressions. Nothing in them ever clicks.

A smaller and more literal artefact appears in the same export. 207 rows carry an unstripped prompt wrapper, in the form:

context: location: united states (not for language). do not include location references in your response. question: what is a good ai visibility score and how is it measured

Nobody types that. It is scaffolding addressed to a language model, sent to a search engine by software that did not remove its own wrapper before querying. Widen the filter to include instruction syntax aimed at a model — "you must," "act as a," "be neutral," "mandatory rules," "include source urls" — and the set reaches 262 rows.

Here is the proportion that matters. Those 262 rows carry 734 impressions: 0.26% of the window. The 207 rows carrying the literal wrapper account for 580 impressions on their own, 0.21%, with zero clicks, at an impression-weighted average position of 4.1 — a median of 3.3 and a spread from 1.0 to 15.0, so these rows rank well on average rather than uniformly at the top.

This artefact has been documented independently and repeatedly through 2026: on a Japanese property finding two such queries (kimagure-weblog.net), on a property finding 163 rows from a single GEO tool's template (echowi.ai), and on a property recording 17 impressions of the same United States wrapper string observed here (imisofts.com). Practitioners have published filtering recipes for finding it (practicalecommerce.com, noqode.fr), catalogues of assistant-conversation fragments appearing as query rows (suganthan.com), and analyses of what it means for click expectations (vidclean.net).

Each of those accounts measures one property and reports a count. Set against a denominator, the visible wrapper is one five-hundredth of the impressions in this dataset. It is the most legible symptom and the smallest one. The conversational half of the demand is where the clicks stopped, and most of it carries no wrapper at all.

What this does and does not establish #

It does not establish causation. Search Console reports an impression and a click; it does not report why a click did not happen. At least three mechanisms are consistent with this data and cannot be separated within it: an AI Overview or AI Mode response answered the question on the results page, as Google's own guidance on AI features anticipates; the impression was generated by query fan-out and had no human attached; or the result was seen and judged not worth a visit.

It does not generalise to all sites. These are six properties in one topic area — AI search, citation measurement and earned media — and that subject matter is unusually likely to be retrieved by AI systems. The direction is consistent with published third-party work on AI Overviews reducing click-through (Ahrefs, Semrush), but the magnitude here should be read as a measurement of this dataset, not an industry constant.

It is also not a performance claim about these pages. A property that converts 282,078 impressions into 212 clicks is not being held up as a success. The value is the shape of the failure, and the fact that the shape is measurable on any property with Search Console access.

What it does establish, within this dataset, is narrower and firmer: conversational query length and question shape are each associated with a collapse in click-through that ranking position does not explain and cannot repair, across approximately half of all impressions.

Running this on your own property #

This requires no tooling beyond Search Console and takes about ten minutes.

1. Open Performance, then Search results, and set the date range to the last 28 days. Note the total impressions and clicks.
2. Add a filter on Query, choose Custom (regex), and enter (\S+\s+){7,}\S+ to isolate queries of eight words or more. Record impressions, clicks and average position.
3. Invert the filter to "doesn't match regex" with the same expression, and record the same three numbers.
4. Compare the two average positions. If the long band ranks better and converts worse, the pattern in this study is present on your property.
5. Filter Query for context: and for question: to find literal prompt wrappers. Expect a small number, and do not mistake its size for the size of the phenomenon.
6. Filter Query for a question mark, and compare that segment's click rate against the rest.

The comparison in step 4 is the one that carries information. A low click rate on long queries is ambiguous on its own; a low click rate on long queries that rank better than your short ones is not a ranking problem.

Two further notes on interpretation. Google states that sites appearing in AI features are reported in the Performance report within the "Web" search type, and documents separately how AI Overviews and AI Mode are counted toward that data — so these rows sit in the same table as classic results and cannot be split out from it. And site owners who wish to change their participation in these experiences can do so through the Search generative AI control, which is a policy decision about inclusion rather than a measurement instrument.

Why this belongs in machine relations #

If roughly half of a property's search demand arrives in conversational form, ranks in the top three, and returns no visitor, then the visit has stopped being the unit that measures whether that demand was served.

What remains measurable is whether the machine that answered the question used your page to do it. That is the question the Machine Relations Index exists to answer, and it is why the Index measures sources rather than brands: it records which domains six answer engines cite, per category and question shape, across a daily release. A publisher can look up its own domain's standing in that release directly.

The conclusion this dataset supports is not that search traffic is ending. Short, navigational and brand-shaped queries in this window still click at 0.248%, and they are a quarter of the impressions. The conclusion is narrower and more useful: the conversational half of search demand, where the questions are richest and the ranking is best, is now being answered somewhere the click cannot be observed — and measuring it requires looking at source selection rather than at sessions.

Methodology #

Source: Google Search Console query-and-page export, six properties, window 2026-08-18 to 2026-09-15, exported 2026-09-18. 16,535 query-page rows, 282,078 impressions, 212 clicks. Rows bucketed by query word count on whitespace; question shape determined by leading interrogative or the presence of a question mark; brand queries determined by the presence of a property's own brand name. Average position is impression-weighted throughout. Counts of prompt wrappers use literal case-insensitive substring matching: the 207-row set is every row whose query contains the string "context: location:", and the 262-row set adds every row containing any of the five instruction phrases quoted above. Positions quoted for that set are impression-weighted, with the median and full range given alongside. No rows were excluded. Anonymised queries are absent from the export by Search Console's design and are not recoverable.