When a shopper types “best three-row SUV under $50k that can tow a boat” into Google’s AI Mode, Google does not run that search. It runs a fan of searches the shopper never sees — tow ratings by trim, third-row legroom, current incentives, reliability data, what is in stock nearby — then assembles an answer from whichever pages win those smaller races. Google calls this query fan-out, and its own developer documentation now states that both AI Overviews and AI Mode may issue multiple related searches across subtopics and data sources to build a response. For dealerships, fan-out quietly rewrites who gets cited, and it explains why a store that “ranks fine” can still be missing from the answer.
One question in, a dozen queries out
Classic search was one-to-one: one query, one ranked list, one winner at the top. Fan-out is one-to-many. A reasoning model decomposes the prompt into sub-questions, retrieves results for each in parallel, and synthesizes the union of those result sets into a single response. The citations under an AI answer are not the top results for the shopper’s query. They are the top passages for the sub-queries the model decided to ask.
Scale makes this impossible to ignore. Google said at I/O 2026 that AI Mode had passed a billion monthly users, and Gemini 3 has been the default model behind AI Overviews since January. Every one of those sessions fans out. In a category as research-heavy as vehicles, where one purchase touches price, capability, cost of ownership, and local availability, the fan is wide.
What a dealer fan-out actually looks like
Google does not expose its sub-queries, but you can reverse-engineer them from the sources an answer cites and the follow-ups it suggests. A representative decomposition for a mid-funnel shopper looks something like this (illustrative):
Prompt: "best 3-row SUV under $50k that can tow a 5,000 lb boat, near Fort Worth" ├─ 3-row SUVs rated to tow 5,000 lb (2026 model year) ├─ [model] towing capacity by trim / tow package required ├─ [model] vs [model] third-row legroom ├─ [model] MSRP by trim ├─ [brand] incentives Texas this month ├─ [model] reliability and cost of ownership └─ [model] in stock near Fort Worth, TX
Notice where a dealer is eligible to win. The OEM site and the review publications own the spec and reliability branches. But the incentive branch, the trim-level towing branch (“which trim do I actually need?”), and the local-inventory branch are questions a dealer can answer more specifically and more currently than anyone else. Those are the branches worth building for.
Ranking still sets the floor, but citations are spreading
Fan-out does not make rankings irrelevant; it changes which rankings count. In a mid-2025 study, Ahrefs found roughly three-quarters of AI Overview citations came from pages ranking in the top ten for the query. Its early-2026 re-run, with improved citation parsing, put that figure near 38%. Part of the drop is methodology, but the direction matches what fan-out predicts: citations increasingly go to pages that rank for the sub-queries, not the head term. A Surfer analysis of roughly 173,000 URLs found that pages ranking for multiple fan-out queries were 161% more likely to be cited than pages ranking only for the main term.
The practical read: your model page ranking fourth for “Highlander for sale” is a floor, not a guarantee. The page that also ranks for “Highlander towing capacity by trim” and “Highlander incentives this month” is the one that gets pulled into the answer three different ways.
Your headings are your bid on each branch
AI retrieval works at the passage level. The model is not judging your page as a whole; it is extracting a chunk that cleanly answers one sub-question. That makes structure a ranking input in a way it never quite was before:
- Write question-shaped H2s and H3s that mirror real sub-queries. “How much can the Highlander tow, by trim?” beats “Capability.”
- Answer in the first 40–60 words under the heading, then elaborate. A passage that needs the three paragraphs around it for context is hard to extract and easy to skip.
- Use real HTML tables for comparative data. Trim-by-trim tow ratings, MSRPs, and seating are tabular facts, and a clean table is the easiest thing on the page for a model to quote accurately.
- Keep numbers next to their units and dates. “$3,000 bonus cash on 2026 models through September 30” survives extraction. “See current offers” does not.
One pattern we see constantly on templated dealer platforms: the most useful local content — incentives, in-stock counts, delivery options — lives inside JavaScript widgets or banner images. If it is not in the rendered HTML as text, it is not a passage, and it cannot win a branch.
Mapping your fan-out without an expensive tool
Google does not report fan-out queries in Search Console, but the long tail they generate leaves fingerprints. Three inexpensive ways to find them:
1. Read the citations. Run your ten most valuable model and service prompts in AI Mode from inside your market. Log every cited URL and ask which sub-question each one answered. The branches where a publisher, a marketplace, or a competitor is cited instead of you are your content plan.
2. Mine long queries in Search Console. In the Performance report, add a query filter using a custom regex to isolate conversational searches of eight words or more — the shape fan-out and AI-assisted searching produce:
^(?:\S+\s+){7,}\S+$
Sort the result by impressions. Long queries with impressions but weak clicks usually point to a branch you partially answer but never state outright.
3. Harvest the follow-ups. The suggested follow-up prompts at the end of an AI Mode answer, and the People Also Ask questions for the head term, are the closest public proxy for the branches the model considers related.
Don’t build a page per branch
The tempting response is to spin up a page for every sub-query. Resist it. Fan-out rewards one authoritative page that answers many related questions, because a single URL can then win several branches at once. Thin variants split your signals and invite trouble: in May, Google expanded its spam policies to explicitly cover attempts to manipulate AI Overviews and AI Mode, and mass-produced near-duplicate pages are exactly what those systems look for. The stronger architecture is a deep model hub — trims, towing, incentives, live inventory, FAQs — supported by a few genuinely distinct pages where intent truly differs, such as a head-to-head comparison or a maintenance-interval guide, all tied together with a clean entity graph in your schema.
Freshness is a branch of its own
Several sub-queries in any car-shopping fan-out are time-bound: this month’s incentives, what changed for the new model year, what is on the ground today. Engines favor sources that are demonstrably current for those branches. Put a visible “updated” date on model and offer pages, keep it honest, and mirror it in your structured data with dateModified. A dealer page that is accurate as of this week beats a publisher article from last spring on exactly the questions dealers are best positioned to answer.
Frequently asked questions
What is query fan-out?
Query fan-out is the technique Google’s AI Mode and AI Overviews use to break one question into multiple related sub-queries, search them in parallel, and combine the results into a single answer. Each sub-query has its own set of candidate sources, so a page can be cited for a sub-question even if it does not rank for the original query.
Can I see which fan-out queries my dealership’s pages were matched to?
Not directly. Google does not expose fan-out sub-queries in Search Console. You can approximate them by logging the sources cited in AI answers for your key prompts, filtering Search Console for long conversational queries, and studying follow-up suggestions and People Also Ask questions.
Should we create a separate page for every sub-query?
No. One comprehensive page with clear, question-shaped headings can win several branches of the same fan-out. Building near-duplicate pages for each variation dilutes your signals and risks running afoul of Google’s spam policies.
