Analyze results
Prompt coverage
Which of your pages should own each prompt, which of the prompt’s fan-out questions that page actually answers, and when the gap is big enough to need a new page.
Definition
Prompt coverage
What it answers#
Engines cite pages that answer the sub-questions behind a prompt, not only the prompt itself (see query fan-out). Prompt coverage turns that into a to-do list:
- Which page should own this prompt? One URL on your site that is closest to the question and its sub-questions.
- Which sub-questions does it answer? Each fan-out query is scored against your pages.
- Refresh or new page? Missing questions that belong together are flagged as a new page candidate instead of sections to bolt on.
How it is calculated#
Scoring a sub-question#
Every fan-out query is compared with each candidate page of yours. A page is split into elements (title, headings, paragraphs), and the page’s score is the similarity of its best-matching element to the query (cosine similarity of embeddings, 0 to 1). The query’s state comes from the best page:
| State | Score |
|---|---|
| Covered | 0.75 or higher |
| Partial | 0.60 to 0.75 |
| Uncovered | below 0.60 |
The panel prints these bands (“covered ≥ 0.75 · partial ≥ 0.6”) and a count such as “5 of 12 fan-out queries need work”.
Proposing an owner#
The split is even on purpose. A page that answers the prompt but none of the sub-questions is a landing page, not an owner; a page that covers the sub-topics but not the question itself is a hub. Neither should beat a page that does both. A page scoring below the Partial band (0.60) is never proposed. If a prompt has no fan-out yet, the owner is scored on the prompt alone.
New page candidates#
When 3 or more uncovered questions cluster together by meaning, they are flagged under New page candidates (“N related fan-out queries with no page near them”). Fewer than three reads as a section to add to an existing page.
Worked example#
A prompt has 12 fan-out queries. Your best page scores 0.81 on 4 of them, 0.66 on 3 and below 0.60 on 5.
- Covered: 4, Partial: 3, Uncovered: 5, shown as “8 of 12 fan-out queries need work”.
- If 3 of the 5 uncovered queries are all about pricing and integrations, they appear as one new page candidate. The other two are sections to add to the owner page.
Which pages are candidates#
Coverage only compares pages that already have vectors, so it never crawls your whole site on its own:
- Your cited pages. Pages on your domain that engines have already cited for your prompts (up to 50).
- Organic keywords candidates. If Search Console or Bing Webmaster Tools is connected, up to 5 of your pages that get organic impressions for related queries in the last 90 days, shown as a grid of impressions and average position per page and query. Each is labelled as an exact, fan-out, semantic or related match. See Search Console and Bing.
If two candidate pages share organic queries, the panel shows a badge with the count and a link to the cannibalization report.
Pinning an owner#
The proposal is a suggestion. If you know which page should own the prompt, Pin a page: enter its URL and select Pin. Pinning crawls and analyzes that page (so it gets vectors if it had none) and recomputes coverage against it. The toast reports the fan-out queries recomputed and the credits charged. Clear pin lets the proposal take over again.
If you see No eligible pages to compare, none of your pages has vectors yet. The listed fan-out queries are planning inputs, not confirmed gaps. Pin a candidate to create one.
When it updates#
Coverage recomputes automatically after a prompt’s semantic map is rebuilt, and when you pin or clear an owner. A recompute where nothing changed (same prompt text, same fan-out set, same page vectors) costs nothing. The selected date range does not change the scores.
Coverage vs. the citation analysis heatmap
Coverage scores your pages. The heatmap in the fan-out table scores the cited pages. There, a fan-out query where no cited source reaches 0.55 similarity is a gap in the whole result set: nobody answers it well, which makes it the easiest question to win.
How to act on it#
- Start with not-ranked prompts that have an owner. Adding the Partial and Uncovered questions to an existing page is the cheapest fix.
- Write the section as a direct answer. One heading per sub-question, the answer in the first sentence, then detail.
- Build new page candidates as new pages. Use Generate article on the prompt detail page, or brief a writer from the cluster. See AI articles.
- Re-check after the next runs. Once the page is re-analyzed, the states update.