Only what the vendor documents, dated October 8, 2026, then what our weekly checks see Gemini and AI Overviews actually read, and what a business does about it.
Google's own documentation is the most explicit of the four:
Sources: developers.google.com, "AI features and your website" (last updated 2025-12-10) and "Google's guide to optimizing for generative AI features" (last updated 2026-07-10), and the Google crawler list, read October 8, 2026. Google does not document how AI Overviews or Gemini choose among eligible pages.
For a business this is the plainest brief of the four: be indexed, be eligible for a snippet, and answer the sub-questions the fan-out will ask. A page that covers "best X in Y" as one essay competes with pages that answer the three or four narrower questions the fan-out issues; the narrower page is the one lifted. The Gemini app is a different surface from AI Overviews and behaves differently in our checks: it sometimes answers without searching at all, from what it already knows, and on those questions a brand has to be in the model's memory, which means the lists and publications it was trained on.
Rows from scans and checks run in September and October 2026, with the companies unnamed where they have not agreed to be named.
| Category | What happened |
|---|---|
| Architecture | Asked for the best custom-home architects in two regions, Gemini answered both without running a search: seven firms each, from memory, no page read, the client absent from both. On the city where it did search, it read a regional list and named the firm it found there first. |
| Resorts | Asked for the best honeymoon resorts in a region, Gemini named six from memory with no search; asked about new openings, it ran six searches, read eleven pages and named the client's opening resort third of seven, describing it with details that need checking because none of the pages read was the client's. |
| Leather goods | Asked for the best MagSafe wallet, Gemini ran a search for the client by name, read two of its pages (a blog post and a product page) and named the client's product fourth of six, the only category answer in that scan where the client's site was opened. |
| Supplements | Asked whether a brand was legit, Gemini read a competitor's blog and the competitor's surveillance test of a sister brand's gummies, and attached that test as a caveat on the client's capsules, then told the buyer to "stick to dry capsules or powder". |
| Leather goods, origin | Asked where a brand's goods are made, Gemini ran eight searches, three of them site-restricted to the brand's own domain, read seven of its pages, found no statement, and wrote "manufactured primarily in China" as fact. |
The 50 questions run on Gemini with Google Search grounding every week, and the archive keeps the searches it ran, so an answer given without a search is marked as such in the dashboard. The work follows Google's own brief: pages that answer the narrower questions the fan-out asks, indexed and eligible for a snippet, with the facts stated in text; the directory and list entries the searches return corrected where they are wrong; and, for the questions Gemini answers from memory, inclusion on the lists and publications it was trained on, which is slower work and is logged as such.
Google says no: no new machine-readable files, AI text files, markup or Markdown are needed, and no special schema. Ordinary structured data still helps rich results, so we keep it; we do not sell it as the lever.
That on that question the model answered from what it already knew, and a page published last week cannot change it. The dashboard marks those answers; the work for them is presence on the lists and publications the model learned from, and the weekly check shows when a search starts happening.
No. AI Overviews and AI Mode are features of Google Search; the Gemini app is a separate assistant that can use Google Search grounding. Rightcited checks the Gemini app weekly; the page-side work for AI Overviews is the same work, and Google says the requirements are the same as for Search.
Or start smaller: the AI Page Check runs one page through all four assistants for $2.99.