# How AI Overviews, AI Mode and Gemini choose whom to cite: what was measured and what the patents only describe

> Google's AI answers are written from results Search has already ranked, often for a sub-query nobody targeted. In RESONEO's Gemini corpus a third of cited passages start in the first 10% of the text, and the link sticks to the model's own sentence, not to the query. What was measured, what patents only describe, and what to do with the first screen.

_Source: https://seomatrix.ai/blog/how-ai-overviews-choose-sources/ · Updated: 2026-11-15_

---

Spoiler for skimmers: Google's AI doesn't look for "the best page on the topic". It takes results Search has already ranked, often for a sub-query nobody targeted, and pins the link to a sentence it is writing anyway, most often one that matches a passage near the top of the text. Below is what was measured, by whom and on what sample, kept apart from what patents merely describe.

In August we re-checked the 297 queries we track and counted who Google's AI Overview cited. Zero citations went to seomatrix.ai. We put that zero into [the study](/ai-overview-citations-study/) on purpose (a study that only shows other people's gaps is marketing). The reflex is to go write yet another guide. The useful move is to ask: who decides who gets the link, and by what rule?

So I read three RESONEO pieces back to back, two built on observed answers and one on Google's internal data schemas, and checked them against the patents in our knowledge base. Some of what the industry repeats holds up. Some of it turns out to be a patent dressed up as a fact!

## Three kinds of evidence. Don't mix them

Most arguments about AI citations break at the same spot: a schema field, a patent claim and a measured percentage get stacked into one confident sentence starting with "Google". Sound familiar? They're different things.

- **Measurements.** Someone captured real answers and counted. RESONEO's Gemini study: 76,450 publicly shared conversations (November 2024 – October 2026), skewed towards news, culture and English. Its AI Overviews / AI Mode / Web Guide study read results-page HTML via a browser extension in February and July 2026 (the July part is 16 captures).
- **Schemas.** Names of internal data types RESONEO published in October 2026. Names, not weights. RESONEO doesn't say where they came from, and their authenticity is unconfirmed. A field existing says nothing about how much it matters.
- **Patents.** What Google *could* do. Not proof that production works that way.

| Claim | Kind of evidence | Who, when, sample | What it does not prove |
|---|---|---|---|
| AI Overviews and AI Mode are written by one Search system, magi | schemas + RESONEO's reading | RESONEO, October 2026 | how magi weighs anything |
| 33% of Gemini's cited passages start in the first 10% of the text | measurement | RESONEO, 76,450 conversations, Nov 2024 – Oct 2026 | that moving a paragraph up earns a citation |
| The link goes after the part of the answer a document verifies | patent US 11,900,068 B1 | Google, priority December 2022 | that the product is tuned this way |
| Share of AI citations from the top 10: 76% → 38% | measurement, second-hand | Ahrefs, March 2026, 863,000 queries, via a GEOHOWTO SPB 2026 talk | why it fell |
| A "90-day freshness cliff" | vendor observations | Salespeak, a 35M-AI-Overview analysis, Trakkr | a Google threshold |

## Who writes the answer, and what it reads first

Per RESONEO's October piece, AI Overviews and AI Mode aren't written by the Gemini app: "This is not the Gemini app: it is magi, the Google Search system that writes these answers." Models, per RESONEO citing Google: Gemini 3 for AI Overviews worldwide since January 2026, Gemini 3.5 Flash for AI Mode since May 2026.

Before writing, magi receives search results with their ranking signals, output from about a hundred specialised providers (Shopping Graph, hotels, flights, reviews; 138 result types by RESONEO's count), and Knowledge Graph facts turned into sentences. All in one list, a graph fact sitting between two web pages. The key line for SEO: "Results reach it ranked by Search: the generated answer (LlmResponseMetadata) receives the RankLab ranking signals of a result and those of the query, the same as a SERP block." Google's guide says it more politely: its generative features are "rooted in our core Search ranking and quality systems".

magi then attaches 17 fields to each result (`PassageSignals`, `FreshnessSignals`, `DedupeSignals` among them); per RESONEO, a page enters the answer as a passage, not as a whole.

The entry ticket is official: "To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet." The current guide adds a third condition: the site must be included in the "Search generative AI features" setting in Search Console. Even then, nothing is guaranteed.

So is ranking still the gate? Mostly, yes! Just not the ranking in your position report: what counts is position for the sub-queries the question gets split into.

## How much comes from the top 10, and what comes from elsewhere

Each number below is a measurement on its own sample, so please don't add them up:

- **Ahrefs** (March 2026, 863,000 queries, presented by Alexander Runov at GEOHOWTO SPB 2026; I haven't seen the primary study): 38% of AI citations come from Google's top 10, down from 76% a year earlier; 31% of AI Overview citations come from beyond the top 100.
- **AirOps**, as retold by DrMax: about 60% of citations sit outside the organic top 20.
- **Ashmanov & Partners** (same conference): in AI Overviews about 60% of sources come from the original query, in AI Mode about 15%. The full comparison is in [our AIO vs AI Mode breakdown](/blog/aio-vs-aim/).
- **RESONEO**, July 2026, one query: 16 domains in organic results, 12 in AI Overviews, 11 in AI Mode, 18 in Web Guide. Only 5 appeared in all four.

Where does the rest come from? The boring answer is query fan-out: the question is split into sub-queries, each with its own ranked list. Natzir Turrado, testing the Gemini app: "What matters is the position, because the sentence goes in the same list as the web pages." With many parallel queries each brings fewer results, and low-ranked ones fall out. So a page that nails one sub-question (a price, a season, a route, in RESONEO's examples) has a shot without ranking for the original query. More in [how query fan-out works](/blog/query-fan-out/).

The less boring answer: some sources aren't web pages at all. Under an answer Google logs each shown source with one of nine types, and only one, `LINKED_SOURCE`, is a link to a page. The rest are Knowledge Graph facts, Google-generated content, Google reviews, Shopping Graph data, UGC and merchant notes. In shopping, local and travel queries a site shares the answer with Google's own data.

Then there are graph facts with no URL. Turrado asked AI Overview his height and got an answer with no source: he had entered it into Freebase himself in 2014, and no web page states it. Dan Petrovic asked AI Mode his age, and the only visible citation went to a footballer namesake's Wikipedia page (awkward!). The schema has the mechanism: a graph fact keeps its source URL only if marked `REQUIRES_CITATION` or if the process is explicitly allowed. In Petrovic's words: "A fact that only exists in the Knowledge Graph reaches AI answers without a link to you." Two tests, one query each, so anecdotes. But they match the schema comment.

## AI Overviews vs AI Mode: one writer, different habits

Here we owe a correction, and an awkward one! Our July post presented the Ashmanov & Partners hypothesis that the two surfaces run on different retrieval engines. The October schemas complicate it: per RESONEO both are written by magi, and AI Mode gets Search ranking signals through the same system. It does have its own, though: `CorroborationRankingSignals` (ranks sources that corroborate the answer's claims) and `ClaimGroundingToolResult` (a tool tying each claim to results) appear only for AI Mode.

An alternative explanation for AI Mode's small share of sources from the original query: fan-out across about a hundred providers, no separate vector engine needed. So the empirical differences are real and the explanation is open. (Better to say it out loud than to quietly rewrite the July post.)

What RESONEO observed in February and July 2026:

- AI Mode runs 8–12 sub-queries per turn by Google's own figure and shows 20–40+ citations; AI Overviews show 5–15. By July the sub-queries ran fully server-side, so the count is Google's account, not extractable data.
- Web Guide, a US Search Labs test, shares a pipeline with AI Mode: a third to a half of the URLs AI Mode cites are in the corpus Web Guide read. AI Overviews don't share that layer. Web Guide reads 30–166 documents and shows only 50–60%, and Google prints that on the page ("Deeply analyzed 21 results, 12 selected").

## Where the link lands: after the part that checks out

The intuitive assumption is that the cited page is "the most relevant result for the query". The patent describes something else. In US 11,900,068 B1, from Google's "Generative summaries for search results" family (priority December 2022), the summary is split into parts, each part's embedding is compared with the embedding of a document's content, and if the distance passes a threshold, the document "verifies" that part and the link goes right after it. One part can carry several links. How it's tuned in the product isn't disclosed.

A patent isn't proof. But the Gemini measurements point the same way. The cited passage is the closest match to Gemini's answer sentence in 54% of cases and to the search query in 10%; a random pick from the roughly 68 passages on a page would give 2%. 70% of answer sentences share no four-word sequence with the cited passage, so Gemini paraphrases. Until April 2026 each citation carried a score for whether the passage supports the sentence (99% agreed); from May that score vanished from the data. RESONEO's reading: the page is picked while writing, the exact passage matched and checked afterwards. An inference, not a direct observation.

In practice: a page needs a sentence that could *be* a sentence of the answer. Not "a page about X" but "X costs Y as of Z". The link sticks to a claim, not to a topic!

> **Three strengths of evidence, not one** — The patent describes a possible mechanism. The schema names exist, but their origin is undisclosed. The 54% comes from the Gemini **app**, which per RESONEO has its own pipeline, inherited from Google Assistant, with AI Mode itself and the results prepared for AI Overviews among its routes. Three sources pointing the same way is a reason to act. It isn't "Google confirmed".

## Gemini reads from the top

This is the number everyone will quote, so carefully: **33%** of cited passages start in the first 10% of the page's text (the first 100 words of a 1,000-word article), **80%** in the first half. RESONEO measured extracted text with navigation stripped, not the first screen. Half the passages are 43 words or shorter, and 99% of the displayed fragments reproduce the start of the passage. The authors' own caveats: moving a paragraph up doesn't guarantee a citation, and 43 words is a median, not an optimal length.

Before searching, Gemini rewrites the question, which changes which pages it can even find. 55% of exchanges with a non-English question include an English query (26% in October 2026). 37% of queries add a year that wasn't in the question (13–17% in August–October). 84% are keyword strings, not questions.

And a sobering one: since May 2026, 43% of exchanges where Gemini searched Google show no sources at all (79% in January). Read and shown are different states. Was a page read and then quietly dropped? From the outside, there's no way to tell.

## Freshness: a named signal, an unproven threshold

Freshness is where the "data" gets shakiest, so sorted by evidence type:

- **Schema:** every result reaching magi carries `FreshnessSignals`, and `STALE_SNIPPET` is a defect type in Google's internal quality grid. Names, no weights.
- **Schema, RESONEO's reading:** Glue counters sum satisfaction by the age of the date shown in the snippet (under a day, 3, 30, 90, 365 days, plus "has a date"). A page with no visible date drops out of those counters.
- **Measurement:** Gemini adds a year to 37% of its rewritten queries.
- **Vendors:** Salespeak says 50% of AI-cited material is under 13 weeks old; an analysis of 35 million AI Overviews says AI-cited content is 25.7% fresher than organic; Trakkr (108,000 citations, 8 models, 10 months) puts a citation's median life at 0 days, with 73.5% appearing once and vanishing.

The "90-day cliff" is a pattern on different samples, not a Google threshold, and RESONEO's data neither confirms nor refutes it. Bumping only the date is another story: the 2024 leak has an `unreliableDatesScore` field for dates the system doesn't trust. No penalty is described, and a field existing proves nothing about how it's used. The refresh routine is in [where AI engines get their answers](/blog/ai-citation-sources/).

## What kinds of pages get cited

The Gemini corpus is the most detailed source here, with its skew (English, news, culture):

| Source | Share of Gemini answers with sources | Cited when Gemini read it |
|---|---|---|
| Wikipedia | 32.1% (28.8% in 2025) | 45% |
| News sites | 55 of 100 answers (41 in 2025) | not reported |
| Reddit | 5.2% since February 2026 (~0 in 2025) | 36% |
| Other sites | not reported | 32% |
| YouTube | read in 18% of answers, cited in 5% | 19% |

The YouTube row is my favourite: video gets read often and cited rarely, while text pages with clearly written facts get cited more readily. Health and science sites fell hard (nih.gov from 8.1% to 2.0%). And 51% of sourced answers cite at least one site outside the top million by popularity (Chrome UX Report, not SEO authority). Big sites don't own this game!

Our [August snapshot](/ai-overview-citations-study/) is a different surface (AI Overviews), a different niche (AI-SEO tools) and a single day: product and landing pages took 52% of citations, guides 23%. That doesn't contradict Gemini's lean towards Wikipedia. I think both say the same thing: the page that plainly states the fact gets the link.

## What I'd do with the first screen and the structure

My subjective synthesis, not a Google rule. Google's guide says there's no need to chunk content or write in a special style for AI, and "There's no ideal page length." Fair enough. None of this is an "AI version" of a page; it's putting the fact where a reader, and a model, finds it first.

- [ ] Entry conditions: indexed, snippet-eligible (no `nosnippet`, no `max-snippet:0`), site included in Search Console's generative AI setting.
- [ ] The first ~100 words of main text hold the answer and the key fact, with a number and a date. No warm-up.
- [ ] Every H2/H3 opens with a self-contained sentence: entity, attribute, value. The 43 words is a median, not a target.
- [ ] Write the sentence the answer should contain. The link attaches to a claim.
- [ ] Facts live as text on the site's own page, not only in the Knowledge Graph, an image or JSON-LD. Markup serves the graph and rich results, not the answer.
- [ ] Sub-questions become sections of one strong page, not a page per rephrasing (made to game rankings or AI answers, Google calls that scaled content abuse).
- [ ] A visible date, and updates that change content, not just the date.
- [ ] Multilingual project? Key pages get an English version: Gemini often searches in English.
- [ ] Per key query, track three states: no overview / overview citing you / overview without you. Seer Interactive (April 2026) counted about 33,500 clicks per million impressions without an overview, 20,700 when cited, 9,400 when not.

For the third item there's a quick mechanical check: the [answer capsule checker](/tools/answer-capsule/) shows whether each section opens with a self-contained answer or with a warm-up.

## In short

- Per RESONEO's schemas, magi writes AI Overviews and AI Mode from results Search already ranked, plus providers and graph facts.
- The original top 10 matters less (Ahrefs: 76% → 38%); sub-query rankings matter more.
- A graph fact keeps its URL only with `REQUIRES_CITATION` or an explicitly allowed process, so facts belong on your own page.
- The patent puts the link after the part a document verifies; Gemini's data agrees (54% vs 10%), and 33% of its cited passages start in the first 10% of text.
- Freshness is a named signal; "90 days" is a vendor pattern, not a Google threshold.

As for our zero from August: the first paragraph of this post was written to be quotable. The next snapshot will tell whether anyone quotes it. Fingers crossed!

## Sources

- RESONEO (Olivier de Segonzac), "Understanding Google", part 4: what AI Overviews and AI Mode read, October 2026, with tests by Natzir Turrado (27 October) and Dan Petrovic (28 October): [think.resoneo.com/ai-overviews-ai-mode/](https://think.resoneo.com/ai-overviews-ai-mode/)
- RESONEO, AIO / AI Mode deep dive, February 2026: [think.resoneo.com/aio-aim-deepdive/](https://think.resoneo.com/aio-aim-deepdive/); July 2026 update with Web Guide: [think.resoneo.com/aio-aim-deepdive/v2/](https://think.resoneo.com/aio-aim-deepdive/v2/)
- RESONEO, how Gemini searches and picks citations (76,450 conversations), October 2026: [think.resoneo.com/gemini-convo/](https://think.resoneo.com/gemini-convo/)
- Google Search Central, guide to optimizing for generative AI features in Google Search, May 2026 (current version, with the Search Console requirement)
- Google patent US 11,900,068 B1, "Generative summaries for search results" family, priority December 2022
- Ahrefs, AI citations vs the top 10, March 2026 (863,000 queries), as presented by Alexander Runov (MGCom) at GEOHOWTO SPB 2026, May 2026
- Ashmanov & Partners (Anastasia Kurdyukova, Mikhail Volovich), AIO vs AI Mode sources, GEOHOWTO SPB 2026, May 2026
- AirOps; Salespeak; an analysis of 35 million AI Overviews — as cited by DrMax (DrMax SEO Telegram channel, 2026)
- Trakkr, "The Half-Life of AI Citations", 2026: [trakkr.ai](https://trakkr.ai/trakkr-research/citation-decay/answers/how-long-do-ai-citations-last)
- Seer Interactive, clicks per million impressions with and without AI Overviews, April 2026, as cited in Dzhilavdarov's Google SEO course
- seo·matrix, "Who AI Overview actually cites: 297 queries measured", 21 August 2026: [/ai-overview-citations-study/](/ai-overview-citations-study/)

## FAQ

### Do you need to rank in the top 10 to be cited in AI Overviews?

Not necessarily. Ahrefs data (March 2026, 863,000 queries, presented at the GEOHOWTO SPB 2026 conference) puts the share of AI citations coming from Google's top 10 at 38%, down from 76% a year earlier, with 31% of AI Overview citations from pages beyond the top 100. Ranking still matters: per schemas published by RESONEO, results reach the AI already ranked by Search. It's just ranking for the sub-queries the question is split into, not only for the original query.

### How is AI Mode different from AI Overviews in choosing sources?

Per RESONEO's October 2026 schemas, both are written by the same Search system, magi, but AI Mode has its own signals, including one that ranks sources corroborating the answer's claims and a tool that ties each claim to results. RESONEO observed 8–12 sub-queries per AI Mode turn (Google's figure) and 20–40+ citations, against 5–15 citations in AI Overviews. The hypothesis that the two run on separate retrieval engines remains unconfirmed.

### Where does Google put the citation link in an AI answer?

Patent US 11,900,068 B1 describes a mechanism: the summary is split into parts, each part's embedding is compared with a document's content, and when a document verifies a part, the link is placed right after it. A patent is not proof of how the product works. RESONEO's Gemini data is consistent with it: the cited passage is closest to Gemini's answer sentence in 54% of cases and to the search query in 10%.

### Does Gemini really cite the beginning of a page?

In RESONEO's corpus of 76,450 shared Gemini conversations (November 2024 – October 2026), 33% of cited passages start in the first 10% of the page's text and 80% in the first half. Half the passages are 43 words or shorter. The corpus skews English and towards news and culture, and the authors warn that moving a paragraph up does not guarantee a citation.

### Does schema markup help get cited in AI Overviews or AI Mode?

Google says no special schema.org markup is needed for its generative AI features, and Natzir Turrado concluded from his tests that schema doesn't directly influence AI Overviews or AI Mode answers. Markup helps build the Knowledge Graph and earns rich results; the model itself receives text — snippets, tables and sentences.

