Cited and retrieved sources

Every run records the URLs the assistant pulled and marks which of them shaped the answer. The distinction is the whole value of the screen, and optimizing toward the wrong one wastes a quarter.

Retrieved against cited

Retrieved means the assistant fetched the page. It was a candidate.

Cited means the answer drew on it: quoted, paraphrased, or linked as the basis for a claim.

A page can be retrieved every single time and never cited, which tells you it is findable and not useful. That is a different problem from never being retrieved at all, and it has a different fix.

Reading the source list

Sort by how often a domain is cited across your prompts. The top of that list is the set of pages currently answering the questions in your market. The deduped, role-labelled version of the same thing is source opportunities, which is the better screen for deciding what to work on. Three patterns show up:

• Aggregators and listicles. Someone else's comparison page is the answer. Getting onto it is often faster than outranking it.
• Community and forum threads. Where the market talks. Hard to influence directly and very influential.
• First-party pages, yours or a rival's. The ones you can act on this week.

Your own domain in the list

Your pages appear here with the same retrieved and cited labels. A page of yours that is retrieved and never cited is the single clearest signal in the product: the assistant found it, read it, and did not find a passage worth using. That goes straight to fixing a page, and the trend version of it is the cited against retrieved chart in coverage and page types.

Why a URL can look odd

Assistants fetch URLs with tracking parameters, redirects and occasionally a cached form. Sources are normalised so the same page is one row rather than five, which is why the URL shown may be shorter than the one the assistant requested.