One Model Family, Three Sourcing Models
Three Google surfaces answer the same question in three different ways. If you work in search, the difference between grounded and generative is not trivia. It decides what you have to do to get cited.
People talk about "ranking in AI" as though it were one thing. It is not. Google runs one model family across three surfaces, and each one decides differently where an answer comes from.
That is the useful part. You are not building a second website for robots. You are making sure one site is retrievable, rankable and citable across every path an answer might take, instead of only the head query you happen to track.
Gemini (the App): Generative-First
How the answer is created. Responses are generated from the model's training data by default. Google Search grounding kicks in selectively, when the model decides a query needs fresh or verifiable information.
Where answers are sourced. Primarily from model weights, which is to say parametric memory: what the model already learned. When it does ground, it pulls live Search results and shows supporting links. But grounding is optional here, not the default.
Why it matters. Brand mentions in the Gemini app depend heavily on training-data presence, not on this week's rankings.
- Your brand is well represented in the training corpus: authoritative coverage, a Wikipedia presence, an established entity.
- Or grounding triggered for that query and you rank well in live Search.
- Most answers come from model memory. If your brand is new, niche, or thinly covered before the training cutoff, you do not exist to the model.
- Great rankings cannot save you on queries the model answers without grounding.
The move: build durable entity presence. Third-party coverage, consistent entity data, and content the model has actually had the chance to learn from.
AI Overviews: Grounding-First
How the answer is created. Retrieval happens first. A custom Gemini model synthesizes a summary from documents Google's ranking systems already retrieved, and it only shows when confidence is high.
Where answers are sourced. The live Search index, the Knowledge Graph and the Shopping Graph. The cited links point at the pages the summary was built from.
Why it matters. Classic retrievability still rules. If you cannot be retrieved and ranked, you cannot be cited.
- You are in the retrieved, rankable set for the query.
- A passage on your page directly supports a claim in the synthesized answer.
- Clean extractable text and freshness help. Ahrefs found AI-cited URLs skew about 26% fresher than organic results.
- No special schema is required for eligibility, per Google.
- You are not retrievable:
noindex,nosnippet, blocking, or simply weak retrieval. - Or no passage on your page supports a claim being made.
- Citation is passage-level. A top ranking no longer guarantees a citation, though it still strongly raises your odds: top-10 overlap ranges from about 17% to 76% across studies.
- Do not count on YMYL suppression. About 88% of healthcare queries now trigger an AI Overview.
The move: write citable passages. Clear, self-contained claims that answer a sub-intent on their own, rather than pages that merely rank.
AI Mode: Grounded Plus Fan-Out
How the answer is created. Query fan-out splits your question into multiple sub-queries, searched in parallel. Gemini then reasons across all of it to compose one conversational answer.
Where answers are sourced. The web index, the Knowledge Graph, and real-time data such as shopping, local and finance, synthesized with links out. Every response is grounded in retrieval.
Why it matters. You are no longer optimizing for one query. You are optimizing for the dozens of sub-queries it fans out into.
- You rank for any of the parallel sub-queries the fan-out generates, not just the head term.
- Pages ranking across fan-out queries are about 161% more likely to be cited, per Ahrefs and Surfer.
- Deep topical coverage and strong entity data widen your surface area.
- You optimized for one keyword while the answer was assembled from a dozen sub-queries you never targeted.
- Only about 20% of AI Mode citations come from top-20 organic results, per seoClarity.
- Passage relevance is the entry ticket, but authority still acts as a filter, especially in YMYL.
The move: map the fan-out. Cover the cluster of related intents around each money query, not just the query itself.
What to Actually Change
Three surfaces, three failure modes. The work splits cleanly along the grounded and generative line.
For the generative surface, build the entity. You cannot optimize your way into a model's memory this quarter. What moves it is durable third-party coverage, consistent entity data across the places Google reconciles identity, and enough substantive content that the model has something to learn. This is slow work and it compounds.
For the grounded surfaces, write passages, not just pages. A page that ranks is the price of entry. What gets cited is a passage that states a claim cleanly enough to be lifted out and still make sense. Self-contained, specific, and attached to a sub-intent someone actually has.
Stop measuring one query. If AI Mode decomposes a question into a dozen sub-queries, then tracking the head term tells you almost nothing about whether you are present in the answer. Map the sub-intents around the queries that matter commercially, and cover them.
None of that requires a separate AI strategy bolted onto your site. It requires the same site to be legible to systems that now assemble answers instead of listing links. And it requires measurement that can see what those systems actually did, which is a different problem, and one we have written about in the three signals worth monitoring.
See which AI surfaces are actually citing you.
WISLR measures the AI channel on your own first-party data: which crawlers read your pages, which answers cite them, and what that traffic does once it arrives. Server-level truth instead of a share-of-voice estimate.