AI Search Content AuditWhich facets of the question does your page answer?

Classic query fan-out is when an AI expands one question into several text search queries. A March 2026 paper from Google Research and the University of Illinois (Jiang et al., arXiv:2603.06397) skips that text step: a small diffusion model turns the question's embedding straight into several fanout vectors and retrieves content that sits near them in meaning. No readable sub-query is produced. The paper tested fashion outfits and music playlists, and it does not say this runs in Google Search or AI Overviews.

Enter a query and the URL of the page you want to check. We split the query into 4 to 12 fanout directions, depending on how broad it is, and show which facets the page answers and which are thin. Each direction comes with an ideal chunk, so you can see what to add to raise the score.

No. The paper's model is a small diffusion model, and it was not tested on web search. This tool splits a question into 4 to 12 distinct facets, writes an ideal chunk for each, embeds those chunks, and uses the vectors as stand-ins for fanout vectors. Treat the result as a guide to how retrievable the page is by vector search, not as a live ranking.

We split the page's text into chunks of about 700 characters and embed each one. For each fanout vector we take the cosine similarity of the closest chunk, map it onto 0-100, and average across all directions. A similarity of 0.6 or more counts as covered, 0.45 or more as partly covered, and anything lower as a gap.

The Gemini and ChatGPT query fan-out tools extract the text queries the AI actually ran. This tool skips queries entirely: it shows which chunks of your page sit closest to each fanout vector when matching happens directly between meanings, and which directions are left empty. It scores semantic closeness, not keyword overlap.

It shows how far apart the fanout vectors point from each other (1 minus their average cosine similarity). Higher means the query breaks down into more distinct facets, which also makes it harder for one page to cover every direction.

Real AI search scales its fan-out to the question, so this check does too. A narrow question like "How tall is Tokyo Tower?" gets 4 or 5 directions, a comparison or buying question gets 6 to 8, and a broad question that mixes budget, region and timing gets up to 12. Coverage is the average across all directions, so scores for questions with different numbers of directions don't compare one to one. To compare before and after an edit, use "Rerun with the same directions" above the result, which measures the page against the same directions as last time.

Each axis is a numbered fanout direction. There are 4 to 12 of them, depending on how broad the question is. Each point is the page's closest passage for that direction. Hover a point to see the similarity and that passage. Further out means closer in meaning. Reaching the green band means covered, the amber band means partly covered, and inside that is a gap. Directions that dip toward the center are where new content can raise the score.

The tool reads the first HTML your server returns and doesn't run JavaScript. It can't read pages that block bots, pages that draw their text with JavaScript, or files that aren't HTML. Many AI crawlers don't run JavaScript either, so if the tool can't read your page, AI search may not see its content.

This tool checks the one page you enter. If you want to know where competitor pages beat yours for the same query, or how to raise your score far enough to get cited by AI search, contact us from the result screen and an umoren.ai consultant will put together a comparison and a plan.

Use the ideal chunks under "Where the score can grow" as a brief and add paragraphs to your page that answer the same need. Specific figures, names and conditions pull the vector closer. If covering every direction on one page would feel forced, create a related page and link to it.

If you're logged in, the result is saved to your account history so you can come back to it. We also log the query, the URL and a summary of the result to improve the service and prevent abuse. We fetch the page once for the check and don't store its full text.