543 Page-One Queries. Zero Clicks. What Your Dashboard Isn't Telling You.
An Agent-Shaped Query Is a Search Written by Software, Not a Person
An agent-shaped query is a search written by software on a person's behalf, not typed by the person themselves — and it is now the fastest-growing thing in our Search Console.
Over the 28 days from 10 August to 6 September 2026, we captured 551 queries on which our pages ranked on page one of Google. At least 543 of them earned zero clicks. Together those 543 were shown 7,038 times (Indexable AI Search Console, 2026). The eight page-one queries that did earn a click — 31 clicks between them, on 448 impressions — were all brand or navigational: our company name, a term we coined, my own name (Indexable AI Search Console, 2026). Every non-branded page-one query on the domain returned zero clicks. Over the same window, with four noindexed video pages excluded, the site as a whole took roughly 51,615 impressions and 124 clicks (Indexable AI Search Console, 2026). Eleven of the zero-click queries were variants of a single nine-word scaffold with one noun swapped, 473 impressions between them (Indexable AI Search Console, 2026).
For context: in 2024, 58.5% of US Google searches ended without a click, and by 2026 fewer than one third still send a click at all (SparkToro, 2026). Gartner predicts 90% of B2B buying will be intermediated by AI agents by 2028, routing more than $15 trillion through agent exchanges (Gartner, 2025). A page ranking first in Google carries about a 58% AI-citation rate (AirOps, 2026).
If that were your brand, your monthly report would not flag it. It would show reach growing and traffic flat, and someone would call it a soft month.
It isn't a soft month. It's a surface your reporting cannot see.
Why Isn't This Just a Bad Headline Problem?
The obvious diagnosis is weak titles: rewrite the metadata, win the clicks back. We nearly spent a week doing that, and on most sites it would have been the right call — low click-through at a strong position usually is a snippet problem, and it is usually the cheapest thing to fix.
We scoped the work first. The only page-one queries that produced a click were the eight with our own name or a term we coined in them; the other 543 produced none at any position (Indexable AI Search Console, 2026). A title rewrite wins clicks from people who are choosing between results. A ceiling that low stopped looking like a title problem and started looking like a composition problem. Something about who was searching, not how we were described.
Then we read the queries.
evaluate the ai visibility products company agent analytics on model coverage
That query put us at position 5.0. It was shown 129 times. Nobody clicked (Indexable AI Search Console, 2026).
Read it aloud. No person types that. There is no verb a human would choose and no shape a thumb would produce on a phone. It reads like a task instruction — the kind of string assembled by software told to go compare vendors on a specific dimension.
It wasn't alone. It is one of eleven variants of the same nine-word stem, each swapping a single evaluation dimension — data accuracy, historical data, ease of use, prompt volume — all on page one, 473 impressions between them, zero clicks (Indexable AI Search Console, 2026). A person does not re-type the same stem eleven times to vary one noun. Widen the lens and the shape holds: in one slice of the scan, 401 of 495 zero-click page-one queries ran to eight words or more, or ended in a question mark (Indexable AI Search Console, 2026). Seventeen of them, carrying 531 impressions, were not in English — German and Spanish, on a site that has never published a page in either language. The largest, a German question about the best platforms for tracking AI visibility and competitive analysis, sat at position 3.0 and was shown 463 times (Indexable AI Search Console, 2026).
What Can't Be Proven Here?
Google Search Console has no field recording whether a person or a machine typed a query. Anyone claiming they can definitively separate the two is selling something.
So this is a pattern, not a proof. The eleven-variant scaffold and the 401 long or question-shaped queries share a signature — long, natural-language, filter-shaped, semantically exact — and it is the same signature we documented across 613 queries in a 90-day window earlier this year (Indexable, 2026).
One domain, 28 days. A field note, not a study. And the 551 is a floor, not a census: the Search Console API returns at most 500 rows per pull, so low-impression page-one queries between positions 5 and 10 are under-sampled. The honest phrasing is "at least 543".
Three alternative explanations are worth stating, because each would undercut the reading. The queries could be scraper or tooling noise rather than genuine buyer research — possible, though the phrasing is too purposeful and too commercially specific for a crawler. They could be humans pasting prompts they wrote for a chatbot into Google instead, which is really the same behaviour one step removed. Or Google could be synthesising query variants internally, in which case the impressions were never a person or an agent at all. We cannot rule any of these out from Search Console data alone, and we are not going to pretend otherwise. What we can say is that all three alternatives lead to the same operational conclusion: a growing share of the impressions in your report has no human attached to it, and no click will ever arrive to tell you so.
Why Might Zero Clicks Be the Correct Outcome?
If the reading holds, the zero is the mechanism working as designed.
Software doing comparison research doesn't need to visit your site. It needs the answer. Google surfaces your page, the tool extracts what it came for, and it moves on to assemble a recommendation for the person who asked. No click is generated because none was required.
That person may never see your URL. They see a shortlist. Your content either shaped it or it didn't.
Under that model, "position five, zero clicks" means your content was read. It was in the room when the shortlist was written. That is a materially different outcome from being ignored, and your analytics assigns them the same value: nothing.
The uncomfortable corollary is that the value shows up somewhere else, later, and misattributed. The buyer who received that shortlist may arrive weeks afterwards as direct traffic, or as a branded search, or as an inbound enquiry that appears to come from nowhere. The page that actually did the work never gets credited, so it never gets funded, and in the next planning cycle it looks like a candidate for pruning.
What Are the Two Readings of the Same Report?
This is the measurement problem, and it is the reason the number is dangerous rather than merely disappointing:
| What your report shows | Reading A — you're losing | Reading B — you're winning invisibly |
|---|---|---|
| Impressions up, clicks flat | Nobody found the page compelling | Something read it and left satisfied |
| CTR falling | Weak titles and snippets | Denominator now includes machines that never click |
| Position strong, traffic weak | Ranking for the wrong terms | Ranking for research queries with no human behind them |
| Direct traffic rising | Unattributed brand demand | Buyers arriving with a shortlist already formed elsewhere |
Both columns produce identical rows in the same report. One means you're losing. The other means you're winning in a way nobody in your organisation is being credited or funded for.
What Should a Marketing Leader Actually Do About It?
Five things, in order, none of which require new budget:
- Start by reading your longest queries. In Search Console, filter to non-branded and sort by length. The agent-shaped strings surface immediately. You can do this in about an hour, and it costs nothing.
- Then, split the two populations before you measure anything. Human-signature and agent-signature queries need separate scorecards. Blended, they produce an average that describes neither. Apply the split before any reporting cycle you actually act on.
- You should stop treating click-through rate as a quality signal on agent-shaped queries. Rewriting a title to win a click from software is optimising for an audience with no hands. In our window the only page-one queries that clicked were the eight brand and navigational ones, 31 clicks in total; on the other 543 a rewritten title had nothing to win (Indexable AI Search Console, 2026).
- Next, implement one new reporting line: citation, not just traffic. Ask how often AI engines name your brand on a fixed set of buyer questions, and schedule the pull weekly. It is the only metric that captures Reading B.
- Use this data as competitive intelligence. "Evaluate on model coverage" tells you model coverage is now a purchase criterion in your category — in the exact language of the brief. You can apply the same read to every competitor term that appears.
Frequently Asked Questions
Does this mean SEO no longer matters?
No — the opposite. Ranking is often what makes a page eligible to be retrieved and cited in the first place. A page at position 1 in Google has roughly a 58% AI-citation rate (AirOps, 2026). The ranking still matters; the click is what stopped being the measure of it.
How do I know if this is happening to my brand?
Sort your non-branded Search Console queries by length and read the longest twenty. If they read like briefs rather than searches — full questions, filter syntax, multiple qualifiers — you have the same pattern. It costs an hour to check.
Isn't zero-click just a normal SEO problem?
Zero-click traditionally meant a human got their answer from the SERP. This is different: there may be no human in the session at all. The traditional fix (better snippets) doesn't apply, because the reader has no hands.
What should I report to my board about this?
Report both denominators. Show human traffic as you do today, and add share of citation on a fixed prompt set. A board that only sees click-denominated metrics will conclude the channel is declining, when it may be relocating.
Who Owns This Inside Your Company?
If software is doing the shortlisting, the funnel no longer starts at your site. It starts with whatever the software hands back. You are not competing for a click — you are competing to be in the answer. The scoreboard for that does not live in Search Console.
Which raises the operational question this series exists to answer: who in your organisation owns that surface?
It is not the SEO lead's job as currently written — that role was scoped for ranking pages. It is not analytics, because the surface isn't instrumented. In most companies the honest answer is nobody, and the instinct is to fix it by opening a headcount req.
That instinct is what the next episode is about.
Key takeaway: your traffic report now measures one of two audiences, and it does not tell you which. The next step is not a new tool — it is a second denominator. Report human traffic as you do today, then add share of citation on a fixed prompt set, and you can finally see both columns of the table above. In summary: rankings still matter, clicks stopped being the proof of them, and the surface has no owner on most org charts.
Data: Indexable AI Search Console, 10 August – 6 September 2026, web search, all countries, excluding four noindexed /videos/ pages. Single domain, 28-day window. Page-one figures are a floor: the API caps each pull at 500 rows, so "at least 543" is the honest count. Position is an impression-weighted average, not a rank-tracker rank. Query attribution is inferred from linguistic pattern, not measured — Search Console exposes no such field. First published 14 August 2026 on a 10 July – 6 August window; figures refreshed 25 September 2026.