Google AI Mode and AI Overviews cite different URLs (about 13.7% overlap). How each surface picks sources, and how to optimize for Google AI Mode and win both.
Google AI Mode and AI Overviews are different surfaces that cite different URLs, so ranking in one does not get you the other. Across 540,000 query pairs, Ahrefs found the two overlap on only 13.7% of the URLs they cite for the same query, and just 16.3% even among the top-3 citations. Win AI Overviews and you can still be invisible in AI Mode, the higher-intent conversational surface most search interest is moving toward. This post explains how each surface picks its sources and the specific moves that win both.
The reason this matters now: at Google I/O on May 19, 2026, Google upgraded AI Mode to Gemini 3.5 Flash as its new default model globally and said AI Mode had passed a billion monthly users about a year after launch, with queries more than doubling every quarter. AI Mode is not a box bolted onto the results page. It is its own surface with its own citation logic, and the answer engine optimization playbook has to account for it directly.
The two are different systems that happen to live under the same brand. One is a summary on the results page. The other is a full conversational surface running a different retrieval method. That difference is why their citation sets diverge.
An AI Overview is a generated summary at the top of the normal results page, built in a single pass. Google reads a handful of sources for one query, writes a short answer, and links the ones it used. It is the same retrieve-read-synthesize-cite pattern, applied inside the SERP. The on-page work that wins it is close to classic SEO with a higher bar for clarity, which is the whole subject of how to show up in Google AI Overviews. Keep that surface in mind as the baseline; AI Mode is the departure.
AI Mode is a separate, full-page surface where you ask a question, get a long synthesized answer, and keep going in a conversation. Under the hood it runs query fan-out instead of a single retrieval. Aleyda Solis describes how AI Mode deconstructs a query into its core intents and related sub-questions, then answers across many sources at once. That mechanism, not the chat interface, is what changes which URLs get cited.
The numbers track the difference. AI Mode answers run about 4x longer than AI Overviews and name 3.3 entities per response on average versus 1.3 for Overviews. AI Mode also omits citations far less often: 3% of its responses lack citations against 11% for AI Overviews. A longer, more entity-dense answer pulled from more sources is a structurally different target.
Query fan-out decides what gets cited by changing the unit of retrieval. Instead of finding sources for one query, AI Mode finds sources for many sub-queries and stitches them together. So a page that answers only the head term competes for a sliver of the answer, while a page that covers the sub-questions can be cited several times over.
AI Mode expands a single question into multiple simultaneous sub-queries covering different intents. Digiday's breakdown puts it plainly: the model breaks one query into multiple searches around related subtopics, all run in the background without the user typing them. Compound questions, comparisons, and constrained queries get split the most; a flat "what is X" may not fan out at all. Aleyda Solis catalogs the kinds of sub-queries it generates: related, implicit, comparative, and follow-up. The practical consequence is that the comparison, the constraint, and the use-case each become their own retrievable question with its own cited source.
AI Mode retrieves for every sub-query at once, then selects passages rather than whole pages. Retrieval here works at the chunk level: Lumar's explainer notes that passage-level retrieval pulls sections of roughly 100-300 words that semantically match a sub-query, instead of ingesting the full article. The model then assembles the answer from the chunks it kept and cites those. This is why a clearly segmented page, where each section is a complete thought, gets pulled into more answers than a long undifferentiated essay that happens to mention everything.
Your AI Overview wins do not carry over because the two surfaces select sources independently, and three separate studies confirm it. This is the data that turns "they feel different" into a measured gap you can plan around.
The divergence shows up in every independent measurement of it:
| Study | Sample | URL overlap | Domain overlap | Headline finding |
|---|---|---|---|---|
| Ahrefs | 540K query pairs | 13.7% (16.3% top-3) | - | Different citation sets for the same query |
| SE Ranking | 4,281 queries | 10.7% | 16% | "Major algorithmic discrepancies" in source selection |
| Victorious | 1,540 queries | - | 23% in both | No query produced identical citation lists |
The methods differ and the exact percentages move, but the conclusion is the same across all three: most of the sources cited in one surface never appear in the other. Victorious found 77% of cited domains showed up in only one surface, and AI Mode cited about 9 domains per query against 7.7 for AI Overviews. They also draw from different pools. AI Mode leans on Wikipedia about ten percentage points more often and cites Quora roughly 3.5x more, while AI Overviews lean harder on YouTube and video.
Here is the part worth sitting with: the answers are nearly the same, but the sources are not. Ahrefs measured 86% average semantic similarity between the two surfaces' answers, with 89.7% of response pairs scoring above 0.8, yet word-level overlap was just 16% and only 0.51% of responses were fully identical. AI Mode is reaching the same conclusions as AI Overviews while quoting mostly different people. That gap is open citation space. If your content already wins Overviews, the AI Mode citation for the same answer is sitting there for whoever structures for it first.
AI Mode citations go to pages that answer the sub-questions, not just the head term, in chunks a model can lift. The fundamentals from AI Overviews still apply, but fan-out adds requirements on top of them. Four moves matter most.
Map the fan-out and write for it. If your topic is a product category, the comparisons, the constraints, the use-cases, and the alternatives are all sub-queries AI Mode will generate, so each deserves its own answerable section. This is why broad topical coverage beats one tightly optimized page, and it is the same entity-and-coverage logic behind semantic SEO automation. A page that covers one slice competes for one citation; a cluster that covers the whole space competes for many.
Structure each section as a complete answer of roughly 100-300 words that reads correctly when quoted on its own. Lead with the answer under a heading that matches the question, then expand. Repeat the entity name instead of leaning on "it" or "this," because a chunk lifted out of context loses its antecedents. This is the same extraction-first discipline that wins citations in how to rank in ChatGPT: the model is quoting a passage, so write passages that survive being quoted.
Name the tools, standards, numbers, and dates a model can pick up. AI Mode names 3.3 entities per answer on average versus 1.3 in AI Overviews, and its answers run about 4x longer, so it has room for specifics and rewards content that supplies them. Vague, entity-thin prose gives a longer answer nothing distinctive to cite. Concrete nouns, real figures with dates, and named products are what a fan-out sub-query latches onto.
Connect the related pages so fan-out keeps landing on your domain across sub-queries. When AI Mode splits one topic into a dozen questions, a well-linked cluster means several of those questions resolve to pages you own. That is the compounding case for internal linking automation: the links spread authority across the sub-queries a single topic fans out into, instead of leaving each new post orphaned.
AI Overviews still reward classic fundamentals: answer-first structure, featured-snippet-shaped passages, verifiable dated facts, and page-one presence. None of that goes away. A page that states the answer near the top, backs it with a current source, and sits under a question-shaped heading is what Google can safely summarize in a single pass. The depth on each of those moves is in the AI Overviews guide; the point here is narrower.
The fundamentals win Overviews, but they are not enough for AI Mode. Overviews reward a clean, citable answer to one query. AI Mode rewards that plus coverage of the sub-questions, passage-sized chunks, and entity density. Do only the first and you win the lower-intent surface while the conversational one cites someone else.
You win both surfaces by doing all of it on every post: fan-out coverage, passage-sized chunks, entity density, verified facts, and internal links, without exception. That is the catch. Any single post can be structured this way by hand. The discipline collapses across a content program, because the work is repetitive and unforgiving, and a model quietly drops a source whose facts have gone stale or whose structure got sloppy. Verifiable, dated, entity-specific facts are exactly what both surfaces cite, which is why fact-checking has to be a hard gate rather than a final glance.
That consistency problem is what Lyra is built to carry. She writes for extraction by default: answer-first sections sized to be quoted, question-shaped headings that map to fan-out sub-queries, named entities and dated facts, fact-checking before anything ships, and internal links that tie each post into the cluster, on every post. Then she opens each one as a pull request you review and merge, so nothing auto-publishes and you stay the editor. The same compounding logic applies whether the citation lands in Overviews or AI Mode, which is the broader SEO for SaaS case for structuring once and earning in both places.
AI Mode and AI Overviews cite different URLs, so winning both means structuring every post for query fan-out and single-pass synthesis at once. Lyra writes that way by default and opens each post as a PR you merge.
Step by step
Map the query fan-out
List the sub-questions a topic implies (comparisons, constraints, use-cases, alternatives) and give each its own retrievable answer, not one page about the head term.
Write passage-sized chunks
Make each section answer one question in roughly 100-300 words so it reads correctly when an AI Mode answer quotes it on its own.
Raise entity density
Name the specific tools, standards, numbers, and dates a model can pick up. AI Mode names about 3.3 entities per answer versus 1.3 in AI Overviews.
Verify every fact and link
Both surfaces favor sources they can attribute with confidence. Check each claim against a current source and confirm every link resolves before publishing.
Build the cluster and internal links
Connect related posts with descriptive anchors so fan-out keeps landing on your domain across the many sub-queries one topic produces.
FAQ
AI Overviews is a generated summary at the top of the normal results page, built in a single pass from a handful of sources for one query. AI Mode is a separate, full-page conversational surface that runs query fan-out: it splits your question into many sub-queries, retrieves for each in parallel, and stitches a longer answer. They are different systems and cite mostly different URLs for the same query.
Mostly not. An Ahrefs study of 540,000 query pairs found the two surfaces overlap on only 13.7% of the URLs they cite for the same query, and 16.3% even among the top-3 citations. SE Ranking found 10.7% URL overlap, and Victorious found 77% of cited domains appeared in only one surface. Ranking in one does not get you the other.
Query fan-out is the retrieval technique behind AI Mode. It expands one query into multiple simultaneous sub-queries covering related, implicit, comparative, and follow-up intents, retrieves from the live web and knowledge graph in parallel, then synthesizes one cited answer. It rewards broad topical coverage over a single optimized page, because each sub-query can land on a different source.
Cover the sub-questions a topic implies, not just the head term. Write self-contained, passage-sized sections (roughly 100-300 words) that read correctly when quoted alone. Be entity-rich and specific, since AI Mode names about 3.3 entities per answer versus 1.3 in AI Overviews. Then build the cluster with internal links so fan-out keeps landing on your domain across sub-queries.
No. The surfaces share only about 13.7% of cited URLs, so an AI Overview win leaves you mostly invisible in AI Mode, which is the higher-intent conversational surface. The fundamentals that win Overviews still help, but AI Mode adds sub-question coverage, passage-sized chunks, and entity density on top. You have to structure for both.
Built by the tool you're reading about
Lyra finds the topics worth ranking for, writes them in your repo's voice, fact-checks every claim, and opens a pull request scored and ready to merge. You review and hit merge. Want to see what she'd write for you? Tell us about your blog and the founder will walk through it with you.
Keep reading
How to get cited by ChatGPT, Perplexity, and Claude. One AEO checklist won't win all three engines, so here's the engine-by-engine playbook, backed by 2026 data.
Schema markup for AI search is hygiene, not a lever. What Google says, what the one controlled test found, and which structured data to actually ship in 2026.
Does Google penalize AI content? Across 600k pages the correlation with rankings is 0.011. The real risk is scaled emptiness. 4 edits that de-risk your AI blog.