What AI Overviews and AI Mode read before they answer

Fan-out queries do not only go to the web index. They go through QRS, the Query Result Service, which spreads them across specialised providers. The Google engineer quoted by Natzir Turrado describes leading QRS “on Search, AI Overviews, Discover”. The schemas count 138 result types under that name, delivered by close to a hundred providers.

Take this question asked to AI Mode:

a weekend in Lisbon in May: where to stay, what to see, how to get there?

1. Hotelsrooms and their rates, partners’ pricesRoomRateResult · HotelPartnerResult
2. Seasonthe best time to goLodgingWhenToVisitDeepDiveResult
3. Flightsflights and their pricesFlightsResult
4. Placesthe ranking of places, wait timesPlacesRankingResult · WaitTimeResult
5. EntitiesKnowledge Graph facts on every landmarkEnrichEntityResult · GeneralEntityDataResult
6. Web and forumsweb pages, the page opened, discussionsWebResultSnippet · BrowseResult · UgcDeepDiveResult

Every provider delivers its own result type, and all of them land in the same list as the web pages: the data attached to every result (ResultExtensions) holds a web result’s snippet and scores (WebResultSnippet, WebResultScores) just as it holds each provider’s output. A sentence taken from the Knowledge Graph therefore sits between two web pages. Natzir Turrado draws the rule from his tests on the Gemini app: “What matters is the position, because the sentence goes in the same list as the web pages.” When several queries go out at once, each brings back fewer results, and a sentence ranked low drops out.

When a searcher gives a thumbs up or down under an AI Overview or an AI Mode answer, Google records every source shown with its type (MagiSourceType). The link to a page is only one of them. The other eight come from Google itself, from commerce and from contributions by users and merchants, which Google sums up in its overview of AI Mode: “facts from our Knowledge Graph, info about the real-world, and shopping data for over 50 billion products”.

• LINKED_SOURCEa link to a web pagethe only link
• GOOGLE_KG_GENERATED_SOURCEcontent produced from the Knowledge Graph
• GOOGLE_GENERATED_SOURCEcontent generated by Google
• GOOGLE_USER_REVIEW_SOURCEa review left on Google
• SHOPPING_GRAPH_DATA_SOURCEShopping Graph data, Google’s product database
• SHOPPING_LICENSED_SOURCElicensed shopping data
• SHOPPING_GENERATED_SOURCEgenerated shopping data
• UGC_STRUCTURED_QUESTIONS_SOURCEcontributors’ answers to standard questions
• UGC_MERCHANT_JUSTIFICATION_SOURCEa merchant’s explanation

In the Knowledge Graph, every source of a fact carries its AI usage rules, a field of cumulative restrictions (data_access_restrictions). The six values draw clear lines between products: grounding in Search, that is AI Overviews and AI Mode, and grounding outside Search, that is the Gemini app and the Vertex AI API, are two separate rights, as are training Search models and training the others. A fact can therefore stay in a knowledge panel while being kept out of generated answers (LLM_RESTRICTION_SEARCH_NO_GENAI).

A publisher’s three settings, and what they reach

• Search Console · “Search generative AI” settingAI Overviews · AI Mode · AI in Discover

  Open to every site since August 31, 2026; by default, the site is included. With the option “Exclude my site’s links and content from Search generative AI features”, the site’s content is no longer used to generate answers in Search’s AI features; it stays in classic search, and Google says it does not use the setting as a ranking signal outside these features. In return, the site gets no more traffic or impressions from those features.

• nosnippet · max-snippetAI Overviews · AI Mode

  Both tags also apply to the inputs of AI answers: nosnippet keeps a page from being a direct input, max-snippet limits how much is used. max-snippet does not restrict the page’s structured data or licensed content.

• Google-ExtendedGemini app · training

  The robots.txt token governs the training of Gemini models, including those behind Search’s AI features, and the grounding of answers in the Gemini app and Vertex AI. It does not remove a site from AI Overviews or AI Mode.

The Gemini app searches the web through the pipeline of the Google Assistant, which it took over from. It picks among fifteen search routes (WorldFactSource): full Google Search, through the same QRS as AI Mode, but also AI Mode itself, consulted as a source, and results already prepared for an AI Overview. Every result gets a status (WorldFactStatus): it is used for the answer, or discarded for one of 29 named reasons. This is the pipeline in which Google-Extended takes a result out of an answer.

AI Overviews and AI Mode start from Search results, but they also read the data Google holds itself. Seven consequences for a site.

1. Check what the Knowledge Graph says about your entity, and fix it at the source

   AI Mode receives a person’s or a business’s biography as the graph writes it, with the “Relevant webpages” line that lists the pages Google ties to the entity. A wrong founding year, an old site or a former job title reach the answer without any page showing them. Updating the entity’s Wikidata item, claiming the knowledge panel, publishing the facts on an official page and keeping them consistent from one site to the next is how they get corrected.

2. Publish your facts on your own pages to win back the link

   A fact known only to the Knowledge Graph enters the answer with no URL: it brings no click. The same fact stated on a page, in a passage that answers the question on its own, gives magi a web result that can become the cited link next to it.

3. Count on structured data for the graph, not for the answer

   The model receives text: snippets, tables, sentences taken from the graph. Structured data acts upstream, when Google ties the page to an entity and builds its facts. Google puts it its own way: no special schema.org markup is needed to appear in AI Overviews or AI Mode.

4. Rank for the sub-questions, not only for the query

   To be cited, a page must be indexed and snippet-eligible, on a site that stays included in the “Search generative AI” setting of Search Console; the answer then receives Search’s ranking signals. Fan-out multiplies the queries: a page that answers one of them precisely (the price, the season, the route) stands a chance even if it does not rank for the original question.

5. Stay the original source, current, adding something else

   magi records every result’s freshness, what it repeats from another source and its outlinks. A page that copies what others already say loses its spot to the original; a dated, updated page that links to its own sources keeps its place.

6. Choose your settings knowing what they reach

   The “Search generative AI” setting in Search Console removes the site from AI Overviews and AI Mode, and cuts off the traffic they send. Google-Extended only reaches model training, the Gemini app and Vertex AI. Before deciding, the generative AI performance reports in Search Console show the site’s impressions in these features.

7. In shopping, local and travel, feed Google’s own data

   The Shopping Graph, Google reviews, hotel rates and merchants’ explanations are answer sources in their own right, served by their own providers. The Merchant Center feed and the Business Profile weigh as much as the site’s pages.

An AI Mode answer mixes pages, which can become links, with facts Google holds itself, delivered with no URL. A site’s share of the answer is won on both sides: through pages that answer passage by passage, and through accurate facts where Google takes them from.