Method notes

An AI answer is composed for one user and never published as a ranked page, so there is no results page to inspect. The only observable evidence is the answer itself: which brands it names, in what terms, and which sources it attributes.

Across the runs we operate, the sources that keep showing up share three practical traits. They are reachable without JavaScript, so a crawler that does not render pages still reads the content. They are unambiguous about what the page is and who publishes it, which is what structured data is for. And they agree with the rest of the web about the same entity: an address, a product name or a claim that contradicts every other source is a weak citation candidate.

That is why DelveDeep separates owned sources from third-party ones in every report. A brand can rewrite its own pages; it can only respond to what a review site or a directory says about it.

A single answer is anecdotal. The Lighthouse model turns a run into comparable numbers by scoring each query independently and then aggregating, so two runs taken weeks apart can be read side by side.

Per query we record whether the brand was mentioned at all, how the mention is framed, where it sits against competitors, and which sources the engine attributed. Queries are weighted: a question that names your brand is not the same signal as an open category question where you have to be recalled unprompted, and adversarial questions are scored separately so a negative mention never counts as presence.

The output is a small set of KPIs per run plus the underlying per-query evidence, because a score nobody can trace back to an answer is not usable in a review meeting.

Being cited requires being readable first. The domain audit runs 57 checks against a site and scores them on a 169-point scale, grouped by how much each one actually costs you: critical access problems, high-value structured data, then medium and low signals.

The critical block is about access: is the homepage reachable, is it HTTPS, does robots.txt allow the AI crawlers, and is the main content present in the served HTML rather than assembled by JavaScript. A site that fails here cannot be fixed by better copy.

The rest covers what makes a page quotable: Organization and FAQ structured data, a sitemap with real lastmod dates, llms.txt, canonical and language declarations, heading structure, image alt text, descriptive link text and outbound references. We run the same audit on this site, and publish what it finds rather than only what it passes.