Quick answer
An AI citation isn't an extra organic position. In March 2026, Ahrefs analyzed 863,000 SERPs and 4 million URLs present in AI Overviews: 37.1% of cited URLs also ranked in the top 10 for the same query, 26.2% between positions 11 and 100, and 36.7% outside the top 100. The divergence is compatible with the query fan-out Google describes: the system can search related subtopics and then select a source useful to one of them. It doesn't prove that a poor ranking favors a citation.
Key takeaways
- Measure the overlap with the same query, the same country, the same device, and the same moment.
- A URL outside the top 100 on the initial query can be visible on a subquery or in another modality.
- Ahrefs's 2025 and 2026 figures don't form a clean series, because the parsing method changed.
- Ranking is still a useful indicator of eligibility and relevance, but it isn't enough to predict selection as a source.
- The gaps become a diagnosis: "well ranked, not cited" and "cited, poorly ranked" call for different actions.
Two output systems, several shared steps
The classic result orders documents for a query. A generated answer builds a text and attaches sources to certain passages. Google explains in its documentation on AI features that AI Overviews and AI Mode can use query fan-out, meaning they launch several searches tied to different subtopics and data sources.
The operating model is therefore:
user query → subqueries → document sets → selected passages → answer → visible citations
The ranking observed by an SEO tool often covers only the first term. A citation can come from a document that's strong for an unobserved subquery. That's a plausible explanation supported by the stated behavior, not proof that every citation outside the top 100 comes from a particular fan-out.
What the Ahrefs study of March 2026 measures
In its updated study, Ahrefs looked for the same URL in the AI Overview and in the classic results of the same query. The sample covers 863,000 keyword SERPs and roughly 4 million AI Overview URLs. The published results are:
| Situation of the cited URL | Share | Correct interpretation |
|---|---|---|
| Ranked in the top 10 | 37.1% | Strong direct overlap for this query |
| Ranked positions 11 to 100 | 26.2% | The source is indexed and findable, but not on page one |
| Outside the top 100 | 36.7% | Not found in the first 100 observed results for the exact term |
Among the citations outside the top 100, 18.2% were YouTube URLs. That shows the possible importance of other formats, but not that YouTube automatically gets a bonus. The presence may also reflect the type of question, the need for a video demonstration, or the set of subqueries.
Ahrefs's 2025 study of 1.9 million citations indicated roughly 76% of citations were also present in the top 10. The 2026 version specifies that citation parsing was improved and that the sample more than doubled. A move from 76% to roughly 38% may result from Google's system, from coverage, or from parsing. It shouldn't be published as a certain drop over time.
Why other platforms give other overlaps
A Semrush study from July 2025 compared 5,000 random keywords and roughly 150,000 citations on desktop, across the Google top 10, AI Overviews, AI Mode, ChatGPT, and Perplexity. It reported roughly 54% domain overlap and 35% URL overlap between AI Mode and the top 10, versus roughly 86% of domains and 67% of URLs for the AI Overviews of that time.
These values don't contradict the Ahrefs 2026 study: they cover an earlier date, a different sample, a different collection tool, and sometimes a domain overlap rather than a URL one. A domain can be shared while the exact page differs. External systems like ChatGPT and Perplexity also have their own indexes, agreements, search mechanisms, and presentations.
The four box diagnostic matrix
| Classic ranking | AI citation | Priority diagnosis | Testable action |
|---|---|---|---|
| Strong | Yes | Dual visibility | Maintain freshness, check stability, and measure clicks |
| Strong | No | General relevance with no passage selection | Compare the cited passages, subquestions, evidence, and formats |
| Weak | Yes | A source useful to an expansion or modality | Identify the cited passage and the related queries where the page is strong |
| Weak | No | Low visibility on both outputs | Review demand, indexing, quality, differentiation, and architecture |
This matrix avoids a classic mistake: rewriting a well ranked page as if it were bad simply because it isn't cited across five tests. Non-citation can come from volatility, from a low number of repeats, or from a subquestion covered by another page.
Procedure: measuring your own overlap
- Select at least 50 prompts matching real tasks, separated by intent.
- Fix the platform, model, language, country, device, sign-in state, and date.
- Run each prompt at least five times, because a single answer doesn't measure stability.
- Record every cited URL, its domain, the passage it supports, and its display position.
- Record the top 10 and, where possible, the top 100 for the exact query at the same moment.
- Calculate URL overlap and domain overlap separately.
- Look for the stated or plausible subqueries, without claiming to reconstruct hidden internal reasoning.
- Place each page in the matrix and define a falsifiable hypothesis.
Reusable formulas
URL overlap@10 = cited URLs also in the top 10 / verifiable cited URLs
domain overlap@10 = cited domains also present in the top 10 domains / cited domains
citation stability = runs where the URL is cited / total runs of the prompt
Keep the denominator visible. If ten URLs are cited in one answer and only one ranks, that isn't equivalent to ten answers each containing one URL.
Worked example: a shared domain, a different URL
For the prompt "how to choose SEO software for an agency," five repeats produce 40 citation slots and 18 unique URLs. SEOryon is cited in three runs, always via /logiciels-seo-ia/seo-software-for-agencies/. The domain is in the top 10 for the exact term, but with the page /logiciels-seo-ia/.
- Stability of the page's citation:
3 / 5 = 60% - Exact URL overlap:
0, because the cited page isn't the ranked URL - Domain overlap:
1, because seoryon.com is present on both surfaces
The diagnosis isn't "the AI ignores Google." The child page may answer a more precise subneed, while the hub better satisfies the classic query. The testable action is to strengthen the hub to child link and the child's specific evidence, then check whether stability changes. It would be an overreach to announce that internal linking caused a citation without a time comparison and enough repeats.
Falsifiable hypotheses for explaining a gap
- H1 (subquery): the cited URL ranks better on a subquestion than on the initial term. Test: record ten related queries and compare the ranks.
- H2 (passage): the page contains a data point or definition that's more directly usable. Test: annotate the proposition each citation supports.
- H3 (format): a video, a table, or a discussion thread provides suitable evidence. Test: segment citations by modality.
- H4 (freshness): the source was updated more recently. Test: compare visible and historical dates, without confusing a changed date with changed content.
- H5 (volatility): the gap disappears after repeats. Test: calculate the stability interval over several days.
What the data proves and doesn't prove
The studies prove a partial overlap: many citations come from well ranked pages, but a significant fraction doesn't rank in the top 10 for the exact query. They also show that domain and URL produce different readings.
They don't prove that ranking is useless, that organic factors no longer matter, or that a page outside the top 100 is selected "thanks to" its weakness. They don't identify the causal weight of length, schema, links, or mentions. The 2026 Google guide reiterates that SEO fundamentals remain relevant and that no special AI schema is required.
Common mistakes and stopping conditions
- Comparing domains on one side and URLs on the other.
- Mixing several platforms under one global "AI" rate.
- Running each prompt only once.
- Using today's ranking to explain historical citations.
- Treating a correlation as a content instruction.
- Stopping the test if the results aren't reproducible or if the tool doesn't reveal the exact URLs.
Reusable asset: the overlap register
Create one row per prompt × run × citation with the columns: date, country, device, platform, model, prompt, subquestion, cited_url, domain, exact_rank, subquery_rank, citation_position, supported_passage, brand, click. A pivot table then produces the four boxes without losing the source observation.
How SEOryon fits in
For the tracking functions actually available, SEOryon can bring visibility observations together with a property's organic data. The reconciliation must keep platform, date, prompt, and URL; an aggregated score without those dimensions would hide precisely the problem analyzed here. Check each integration's public availability before presenting it.
Measurable exercise
Test 30 prompts across five repeats and deliver the four box matrix for your domain and three competitors. Success criterion: at least 150 documented runs, URL and domain rates kept separate, and a hypothesis with a failure condition for each of the ten main divergences.
FAQ
Can a page be cited without being in the top 100?
Yes, the Ahrefs 2026 study observes this for 36.7% of cited URLs. That alone doesn't reveal the cause; subqueries, videos, indexes, and volatility are explanations to test.
Should you aim for the top 10 or for the citation first?
Work on usefulness, indexability, and organic relevance first. Then study the passages and subquestions that earn citations. The two goals share a foundation without being identical.
Why did the Ahrefs 2025 study say 76%?
The 2026 study reports improved parsing and a bigger sample. The system also evolved. With no constant method, the difference can't be attributed to Google alone.
Is schema enough to get cited?
No. Structured data can clarify entities when it matches the visible content, but Google promises neither citation nor ranking and states that no special schema is required for AI.
Read next
Put the figures back into the Observatory register, then map the subquestions with the query fan-out guide. Next, measure AI visibility with repeats before applying the protocol for getting cited by ChatGPT, Perplexity, and Gemini.
References
- Ahrefs: 38% of AI Overview citations come from the top 10- Ahrefs: 2025 study on rankings and citations- Semrush: comparing AI Mode, AI Overviews, and assistants- Google: AI features and query fan-out- Google: optimization guide for generative search
Method and update note
Reviewed 16 July 2026, translated and edited 22 July 2026. The Ahrefs 2026 proportions use the distribution table of cited URLs, not the rounded figure in the headline. Any update must separately note changes of model, parsing, country, device, and denominator.