Audit Content Chunk Citability and Page Cannibalization

Being cited does not mean every section earns it: most pages carry passages that say nothing checkable, or that repeat the page next to them. Here you separate passages worth quoting from pages doing the same job.

The GEO problem it solves

Canonicalizing one of two URLs answering the same query only nominates a preferred address. It does not repair an architecture where both carry the same answer.

The citation record ties a URL to a monitored answer with its prompt, provider, model and date, never to the passage. Overlap is credible only when the same prompts repeatedly cite both pages and their answer-carrying sections make the same claim.

How the workflow breaks down

1. Start from what is cited. get_page_presence returns your first-party pages with citation counts and distinct prompts, labelled WORKING at three or more citations, NEEDS_ATTENTION at one or two, NOT_CITED at zero, and DISCOVERED when cited but unmonitored. No date filter, so the counts are all-time.

2. Build the dated baseline. get_answers over explicit dates, paginated, flags whether your domain was cited; get_answer_details returns the exact URLs with their position, prompt, provider, model and date.

3. Gate the passages yourself. None of these tools return passage text: it comes from an authorized fetch or a CMS export. Cut each page into heading-led passages and put each through four binary gates: it answers the question in its first two sentences; it names the brand rather than "we"; it stands alone; it carries one verifiable specific. What matters is how many it fails.

4. Compare pages on two things. They answer the same user question, and their passing passages make the same claim, both yes or no. Then count the baseline prompts citing both.

What a run actually returns

A real site of 263 pages had 13 cited at all. The top four carried 66, 44, 43 and 31 citations; the next tier carried 6, 5 and 5; everything below sat at 1 or 0.

Citation counts do not taper, they fall off a cliff, and that changes what this audit is for. With four pages carrying most of the evidence, the useful question is not which of the 250 uncited pages to rewrite; it is whether anything in your library competes with those four for the same question. A page splitting the evidence behind a top row costs more than a page nobody cites ever will.

What to do with the output

Keep the baseline with the plan: dates, prompt IDs, providers, models, rows and your edit date. Re-running that window is the only comparison available, so file it before publishing.

Then work top down, stopping when capacity runs out.

Add DISCOVERED pages to monitoring first: cited but unmonitored, they cannot be followed. Then rewrite the cited passage of any NEEDS_ATTENTION page failing two or more gates, the one case where record and text point at the same page. Then differentiate pairs that answer the same question with the same claim and share at least two baseline prompts; two pages cited for one prompt can be complementary, so that count is a gate, not a verdict. Consolidate only when one of the pair has zero citations in the window and fails two or more gates. Everything else waits, including NOT_CITED pages clearing all four.

The Complete Prompt

I want to audit section-level citability and page overlap. Keep apart what the record shows, what the page text says, and what you infer. Run list_brands and, if I monitor more than one brand, ask which to use.

1. Run get_page_presence and group my pages by label: WORKING at three or more citations, NEEDS_ATTENTION at one or two, NOT_CITED at zero, DISCOVERED for cited pages I do not monitor. It has no date filter: say the totals are all-time. If no pages come back, say there is no page-level citation data and stop.
2. Use the last 30 days unless I say otherwise. Run get_answers over that range, paginate fully, keep the responses flagged as citing my domain, then run get_answer_details on each: one row per answer-source pair with prompt, provider, model, date, URL and position. If zero rows come back, say there is no baseline, then widen the window or stop rather than scoring against nothing.
3. Ask me for the page text of those URLs, since none of these tools return it: an authorized fetch, a CMS export or pasted HTML. Strip navigation and footers.
4. Segment each page into heading-led passages and put each through four yes-or-no gates, quoting the deciding line: does it answer the question in its first two sentences, does it name the brand instead of "we" or "our platform", does it stand alone, does it give one verifiable specific such as a figure or a date? Report which gates fail, not a score; this is never a Qwairy metric. Then name the passage most likely reused behind each cited URL, as inference.
5. Compare pages pairwise on two things only, each yes or no: same user question, same claim in their passing passages, plus how many baseline prompts cited both. Then give me a decision map in this order: add DISCOVERED pages to monitoring; rewrite the cited passage of NEEDS_ATTENTION pages failing two or more gates; differentiate pairs matching on both comparisons and sharing at least two prompts; consolidate only where one has zero citations in the window and fails two or more gates. Leave the rest alone, name the passages to rewrite, and add a re-check protocol reusing the same prompts, providers, models and dates. Never call two pages competing on a shared keyword alone, and never say an edit caused a citation change.