Search queries

When an AI model answers a prompt, it rarely answers from memory. It first runs its own web searches, reads what comes back, and writes the answer from that. Those searches decide which pages the model reads, and therefore which brands can appear in the answer at all.

The Search queries page collects them. Every search your project's prompts triggered, across every model and location you track, in one list.

Search queries and prompts are not the same thing#

A prompt is the question you configured. A search query is what the model typed into a search engine while working on it.

One prompt usually produces several search queries, and the model writes them itself. It rephrases, splits a broad question into narrower ones, and adds words you never used. This expansion is often called query fan-out. That gap is the point of the page: prompts tell you what you asked, search queries tell you what the model went looking for.

What each row shows#

What counts as one search#

Two searches share a row when they are the same words. Capitalisation, punctuation and spacing are ignored, and so is the order the words came in: "best running shoes 2026" and "best running shoes 2026" written with a hyphen, or with the year first, are all one row, named after the wording the models ran most.

Anything beyond that is a separate row. "running shoe reviews" and "running shoe tests" mean the same thing to you, but nothing in the wording says so, and merging them would mean guessing.

Grouping by topic or prompt#

The Group by control cuts the table into collapsible sections, by prompt topic or by individual prompt. Each section shows how many distinct searches it holds.

Sections split the runs, not the searches. Every run belongs to one prompt, so a search that two topics both triggered appears under both, with its runs divided between them. The section counts therefore add up to the period's total rather than double-counting it.

The two panels#

Below the table, two panels describe the whole period rather than the rows currently on screen.

Brands counts how often the models named one of your tracked brands inside a search. Only tracked brands appear: a name nobody configured has no identity to count under.

Common terms is the phrasebook. It ranks the word pairs and triples that recur across the searches, with years and bare numbers filtered out, because almost every search an engine writes carries a year and it tells you nothing. This is usually the panel a communications team acts on first.

Filters and export#

The date range, AI model, location and topic filters at the top of the page all apply. The search box narrows the whole list rather than only the rows currently on screen, so a search reaches queries further down than the page displays.

Each of the three views exports on its own, as CSV or Excel: the flat list of searches with their run counts, the brand counts, and the common terms. All three carry the filters and the search term the page currently has, so the files describe the same period you were reading.

How to use it#

• Read Common terms as the vocabulary your market is researched with, and check it against your own pages: the product page, the FAQ, the press area. Words that recur in the searches and appear nowhere on the page that should answer them are the first thing to fix.
• Carry the same wording into what you hand other people. Pitches, briefings, a guest article and an expert quote all become pages the models read, and the wording you bring is the part you control.
• Watch Brands for competitors the models name while researching your market, including in searches that were never about them.
• Watch for searches that are new in a period. They show a topic entering the model's research before it shows up anywhere else.
• Compare a search against the domains and URLs cited for the same period to see which pages that search surfaced.

A search query is a research clue, not a verified gap. It tells you what the model went looking for, not why it chose one source over another, and no rewrite can be promised to change a later answer. Treat it as the starting point for a change you then monitor.