What queries does Google show a small AI-search site for? 83 of the 265 it discloses are written for a chatbot, not a search box: median 21 words, 366 impressions, 0 clicks

The read

We pulled every query Search Console will disclose for this property for 2026-07-26 through 2026-08-22, web search only, with the API's `dataState: all` flag so the most recent partial days are included. Three calls: the query dimension alone, query by page, and the property total with no dimension. The property total is 2572 impressions and 12 clicks. The 265 disclosed queries account for 1049 of those impressions (40.8%) and none of the clicks. The other 1523 impressions, the ones Google anonymises because the query is rare or private, hold all 12.

That last sentence is the one we did not expect. On this site, whatever people typed before they clicked is exactly the set of queries Google will not show us.

How each query was classified

We classified by shape, not by guessed intent, with a fixed rule order where the first match wins. A query is a `context block` if it contains the literal string `<context>`. It is an `operator` query if it carries `-site:` exclusions or a leading quote operator. It is `role or first person` if it opens with a role or a first-person frame ("as a COO", "I'm a brand manager", "our law firm wants", "we're an industrial company"). It is a `long question` if it has eight or more words and either ends in a question mark or opens with a question word. `Date-stamped` queries contain 2026, July or August. `Vendor-named short` queries name one of 30 third-party vendors third-party tools (Scrunch, Profound, AthenaHQ, Otterly, Peec, Suede and the rest of the list in the artifact) and matched nothing above. Everything else is a `short keyword`. "Prompt-shaped" in the headline means the first three classes together.

65 of the 265 queries are twelve words or longer; 43 are twenty or longer. The median disclosed query is 6 words, which is already long for a search box. The longest is 77 words.

What the prompt-shaped queries look like

The ten context-block queries all carry the same structure. One reads, in full: "between getmint and profound, for us - geo - transactionnel, who has superior customer support? who is the best? <context> company url: https://www.getmint.ai market: united states </context>". That is 27 words including an angle-bracketed block that states a company URL and a market. Nobody types an XML-style context tag into Google. Something assembled that string and sent it, and Google recorded an impression for our Profound comparison page against it at position 76.6.

The role-prefixed queries read like a briefing. "as the coo of a policy nonprofit, is there a real difference between ai visibility vendors, or should we just pick the cheapest among suede, athenahq, and scrunch?" (22 impressions, position 91.4). "i'm a procurement lead. what should i ask ai visibility vendors like suede, profound, and athenahq during an rfp to compare them fairly?" (11 impressions). "i'm the enrollment marketing director. suede vs athenahq for tracking prospective-student questions: which is stronger for a university?". Twenty-two of the 43 role-prefixed queries name Suede, a vendor we have never written a page about; the impressions land on our other vendor comparison pages anyway.

Three queries of 73, 75 and 77 words begin identically, "I'm a marketing manager at a large enterprise responsible for brand visibility...", and differ only in the final clause: "citations & sources analytics", "enterprise grade features", "global coverage". Each earned exactly one impression. That is a template with a variable slot, run three times.

A different cluster is short rather than long. 33 disclosed queries begin with "otterly" ("otterly.ai ai visibility reports dashboard", "otterly.ai ai visibility reports dashboards", "otterly.ai ai visibility reports features", "otterly.ai ai visibility reports features review") and together earned 35 impressions, about one each, at an average position of 79.9. Thirty-three near-duplicate permutations of one brand name, each fired once, is a grid, not a person.

Where they land

By query-and-page rows (a query that surfaced two of our pages counts once per page), the prompt-shaped impressions fall mostly on the pages built to compare vendors and explain the mechanism: `/library` (58), `/guide/why-ai-recommends-competitors` (43), `/compare/scrunch-ai-alternative` (35), `/compare/best-profound-alternatives` (35), `/compare/athenahq-alternative` (31), `/compare/peec-alternative` (28). The short-keyword impressions go somewhere else entirely: 227 of 316 land on `/guide`, 155 of them the three phrasings of "get recommended by ChatGPT" or "by AI" on one page at positions 76 to 78. The vendor-named short queries go to `/compare` (308 of 336).

What we can and cannot say about where these come from

Google's own methodology note, which we covered on 2026-08-07, says AI Mode activity is reported inside the general performance report and that every follow-up turn in an AI Mode conversation counts as a new query. So a 40-word first-person brief can enter Search Console from a human typing into AI Mode. It can also enter from an AI-visibility monitoring tool that runs a prompt panel through Google on a schedule; the `<context>` block with a vendor's own URL in it and the 33-permutation Otterly grid look far more like that. We cannot tell the two apart from our side of the report, and we are not going to pretend otherwise. What we can measure is the shape, and the shape is a third of everything Google will disclose.

We also hold the position numbers loosely. An impression at position 73 for a 60-word prompt does not mean a person paged to result 73. Google's documentation does not say how position is assigned for an AI Mode citation, and we have not found a primary source that does. The number is reported as read.

The claim that gets tested

Pre-committed read on 2026-09-06, same three API calls, same rules, same artifact script. If prompt-shaped queries are still at least 25% of disclosed queries and have still earned 0 clicks, the claim in the headline stands: on this site, chatbot-shaped queries produce impressions and nothing else. If prompt-shaped queries have earned 3 or more clicks by then, the claim is wrong and this post will say so at the top. If the share falls under 15%, the August reading was a burst, and that gets written down too. We will also report whether the 12 property clicks are still entirely inside the undisclosed set.

Limits

• The classification is by surface shape with a fixed rule order; a human who writes long questions is counted as prompt-shaped, and a tool that fires three-word queries is counted as short. The rules are in the artifact so anyone can re-run them.
• Search Console disclosed 40.8% of impressions as queries. The anonymised 59% is where every click lives, and we can say nothing about its shape.
• One property, 71 days old, 2572 impressions. These are our numbers. A site with real search traffic will have a different mix, and we would like to see it.
• `dataState: all` includes the last two days of partial data. The 2026-09-06 re-read uses the same flag so the two reads are like for like.