A Prompt-to-Page Map for AI Search Measurement
For every prompt in your measurement set, write down the one page on your site that's meant to answer it. Then record which of your URLs, if any, assistants actually cite for that prompt across repeated runs. Each prompt lands in one of five states: no page exists; the page exists and is never cited; the intended page is cited; a different page of yours is cited instead; or your brand is mentioned but none of your URLs is cited. Each state points at different work, from writing a page to deciding which page should carry the answer. The map is the bridge between a visibility number and a content backlog, and it's the step most programmes skip.
A visibility rate tells you how often you appear. It doesn't tell you what to do on Monday. The prompt-to-page map does, because it turns 'we're cited in 6% of answers' into a list of prompts, each with a page, a state and an owner.
The five states
| State | What the runs show | What it usually points to |
|---|---|---|
| No page | Nothing on the site addresses the prompt | A decision: write one, or accept the gap |
| Page, never cited | Intended page exists; zero citations across runs | Check it's indexed, reachable and answers directly |
| Intended page cited | The planned page appears as a source | Leave it alone and keep sampling |
| Other page cited | A different URL of yours is cited | Decide which page should own the answer |
| Mentioned, not cited | Brand named; none of your URLs cited | Look at which sources were cited instead |
The mapping sheet
prompt_id,prompt_version,intended_url,surfaces,runs,cited_intended,cited_other_url,cited_other_count,mentioned_not_cited,state,owner,next_check
P07,v3,/guides/contractor-invoicing,S-01;S-03,12,0,/pricing,3,2,other page cited,content,2026-10-13
(illustrative row)An illustrative first pass
Twenty-five prompts, three runs each on four surfaces. Four prompts have no page. Nine have a page that was never cited. Five are cited on the intended page. Three are cited on a different page of yours. Four get a mention with no citation. The three 'other page' rows are often the quickest to act on, because a page of yours is already being used; the question is whether it's the right one. The nine 'never cited' rows need the most patience, since twelve runs is a thin basis for calling any single prompt.
Cross-checking with first-party page data
Two reports now show page-level AI data for your own site. Search Console's generative AI report groups impressions by the final URL a Google AI feature linked to, after redirects. Bing Webmaster Tools' AI Performance report shows citation counts per URL across Microsoft's AI surfaces. Neither ties a page to your prompt. Both can tell you whether a page you've marked 'never cited' is appearing in Google's or Microsoft's AI features for questions you aren't sampling. If it is, write that in the row. It's a different question from yours, and a useful one.
One intended page per prompt
If two of your pages are meant to answer the same prompt, you don't have a mapping problem, you have an overlap problem, and the 'other page cited' state will keep flickering between them. That's the same situation as two of your pages competing for one query in search. Resolve it by consolidating or differentiating the pages, then map the prompt to the survivor.
FAQ
What if one page is meant to answer several prompts? — That's normal for a strong guide. Map each prompt to it separately; states are per prompt, so the page can be cited for one question and ignored for another.
Does 'never cited' mean the page is bad? — No. It means it wasn't cited in the runs you sampled. Check indexing, reachability and whether it answers the prompt directly, and look at how many runs you have before concluding anything.
Should branded prompts go on the map? — Keep them on a separate map. A prompt naming your brand tests whether assistants describe you correctly, which is a different question from whether they find you.
Put this to work on your own website.
Fenn finds what your customers ask, drafts the articles and site fixes, and measures what ChatGPT, Claude, Gemini, Perplexity and Grok say about you — with every change waiting for your approval.