Does Ranking in Google's Top 10 Get You Cited by AI?
Published estimates of the overlap between top-10 rankings and AI citations range from 12% to 93%. The spread is the finding: rank is neither necessary nor sufficient for citation.
Updated · Published
Frequently asked questions
Does ranking
Not reliably. In Ahrefs' March 2026 analysis of 863,000 keyword SERPs and 4 million AI Overview URLs, 37.9% of cited pages ranked in Google's top 10 for the same keyword, meaning roughly three in five citations went to pages that did not. BrightEdge, using its own parser, puts top-10 overlap near 17%. A top-10 rank helps and does not decide the outcome.
What percentage of AI Overview citations come from top 10 results?
It depends whose parser you ask and when they measured. Ahrefs reported 76.1% in July 2025 and 37.9% in March 2026; BrightEdge's figure is 16.7% across nine industries; CiteLens found 40% across 500 commercial prompts in June 2026. All agree on the direction of the conclusion even though the values differ by more than 4x: the top 10 accounts for a minority of AI Overview citations.
Why do studies disagree about rank–citation overlap?
Mostly because they answer different questions. 'Ranks in the top 10' and 'ranks anywhere in the top 100' differ by 38 points inside BrightEdge's own dataset (16.7% versus 54.5%). Beyond that, studies differ in which retrieval surface they measure, which language and market they sample, which citations they count, how many they sample per answer, and how their parser detects a citation at all. Ahrefs explicitly flags improved parsing as part of why its own number fell.
Is Google's AI Mode different from AI Overviews for citations?
On the only published comparison, sharply. CiteLens measured 93% of AI Mode citations inside Google's top 10 in June 2026, while its own AI Overviews study the same month put the figure near 40%. The two studies also differ in market and query style, so the surface is not the only candidate explanation. Either way, treat an overlap statistic as specific to the surface it was measured on rather than to Google generally.
Do AI assistants like ChatGPT cite pages that rank on Google?
Much less often than Google's own AI features do, though estimates vary widely. Ahrefs sent 15,000 long-tail queries to four assistants in August 2025 and found an average of 12% of assistant-cited links in Google's top 10, with ChatGPT's in-text citations at 8% and Perplexity at 28.6%. CiteLens, using 320 templated buyer queries in Turkish, reported 30% for ChatGPT and 89% for Perplexity. Query style and market plausibly explain most of that gap, and neither study settles it.
Does the language of the query change which sources AI cites?
Substantially, on the one dataset that tested it. CiteLens ran identical questions in Turkish and English across 126 categories and found AI Overviews shared only 22% of their cited sources between the two. If that generalizes even loosely, an overlap figure measured in one language is weak evidence about another, and English-language benchmarks say little about non-English markets.
Has rank–citation overlap been studied outside vendor research?
Yes, once. A SIGIR 2026 paper from researchers at NJIT, NTU Singapore, and Indiana University ran 11,500 queries against Google's traditional SERP, AI Overviews, and Gemini in December 2025 and published the dataset. Average source Jaccard similarity was 0.18 between AI Overviews and the SERP, 0.16 between the SERP and Gemini, and 0.11 between the two generative surfaces. Because all three return roughly nine sources per query, that converts to about 30% of an AI Overview's sources appearing among the ranked results — within eight points of Ahrefs' 37.9% vendor estimate, measured with a different instrument.
Does this mean SEO no longer matters for AI visibility?
No. It means rank is a weaker proxy than it looks, not that the underlying work is wasted. Your pages still have to be crawlable, parseable, and worth retrieving, and Google's own guidance is that its AI features run on core Search ranking systems. What the data undercuts is the inference step: reading a rank report and concluding anything specific about your AI citations.
Should I stop using rank tracking to measure AI visibility?
Stop using it as a substitute. Rank tracking cannot see the 31% of AI Overview citations that fall outside Google's top 100, and it cannot see citations to third-party pages carrying your brand. Measure citations and mentions directly on each engine, and keep rank data for what it actually reports.
"Rank in the top 10 and you'll get cited" was a reasonable summary of Google's AI Overviews in mid-2025. It is not one now. Ahrefs measured 76.1% top-10 overlap in July 2025 and 37.9% in March 2026, and draws the comparison itself: Google is "selecting far fewer pages straight from the original SERP." The conclusion is right. The evidence is messier than the headline, and the mess is the useful part: across five vendor datasets and now one peer-reviewed benchmark, headline estimates of how much AI citation flows through top-ranked pages run from 12% to 93%, and most of that range comes from studies asking subtly different questions.
Key takeaways
Ahrefs' March 2026 study of 863,000 keyword SERPs and 4 million AI Overview URLs found 37.9% of cited pages in the top 10, 31.2% in positions 11–100, and 31.0% outside the top 100 entirely.
Its July 2025 predecessor found 76.1% in the top 10 — but sampled only the three most visible citations per overview, the slots its own data shows rank highest.
"Top 10" and "top 100" are not the same question. Inside BrightEdge's dataset they differ by 38 points: 16.7% versus 54.5%.
Off Google, the link is weaker still. Ahrefs found 12% average overlap between assistant citations and Google's top 10, with ChatGPT at 8%.
The surface may matter as much as the engine. CiteLens measured 93% overlap on Google's AI Mode in June 2026 while its own AI Overviews study that month landed near 40% — two Google products, one vendor, a 53-point gap.
Language is a variable almost nobody controls for. Identical questions asked in Turkish and English shared just 22% of their cited sources.
Vertical variation rivals vendor variation. BrightEdge measured 75.3% overlap in healthcare against 22.9% in e-commerce, a 52-point spread inside a single dataset.
The first peer-reviewed benchmark now exists: a SIGIR 2026 study of 11,500 queries put source overlap between AI Overviews and the ranked SERP at a Jaccard of 0.18 — which converts to roughly 30% of an overview's sources appearing in the ranked list, within eight points of Ahrefs' estimate.
What survives every dataset: the top 10 is a minority pathway on the surfaces most buyers use, and rank tracking cannot see the citations that matter most.
Break the engine rows apart and the range runs from 6% to 93%, on what sounds like a single factual question. But most of those gaps close as soon as you read the column headers.
The first is the rank range. BrightEdge reports both 16.7% and 54.5% from the same data, because one figure asks whether a cited page reached the top 10 and the other asks whether it ranked at all. Quoting those as rival findings — which happens constantly — is quoting one study against itself.
The second is which citations get counted. Ahrefs' 2025 study analyzed the three most visible citations per overview and reported their median Google ranks as position 2, 4, and 5 for the first, second, and third citation slot. That is a monotonic decline: the deeper into an overview's source list you look, the lower the pages rank. The 2026 study counted 4 million overview URLs rather than a top-three sample, which means it sampled further down exactly that curve. Ahrefs is direct about the other half of the explanation too, attributing part of the drop to better detection of citations its earlier parser missed rather than to a pure change in Google's behavior.
So the honest reading of 76% → 38% is not "overlap halved in eight months." It is "a broader, better-instrumented look at the same phenomenon found a lot more low-ranking sources than the first look did."
Everything in the table above is vendor research — serious work in the best cases, but unreviewed, and in one case distributed by press release with the data available "on request." That gap closed in 2026. A SIGIR 2026 paper from researchers at NJIT, NTU Singapore, and Indiana University ran the academic version of the same measurement: 11,500 queries drawn from six public datasets, all three of Google's surfaces — the traditional SERP, AI Overviews, and Gemini — collected on 7–8 December 2025 from a simulated mobile device in Newark, NJ, with the query set and retrieval scripts published for replication.
Their headline metric is source Jaccard similarity: of all the sources returned by either surface for the same query, the share returned by both. The averages are 0.18 for AI Overviews against the SERP, 0.16 for the SERP against Gemini, and 0.11 for AI Overviews against Gemini. Order-aware overlap is no kinder — rank-biased overlap of 0.23, 0.21, and 0.15 for the same pairs. Google's two generative surfaces agree with each other less than either agrees with the ranked list the company has published for twenty-five years.
A Jaccard of 0.18 sounds irreconcilable with Ahrefs' 37.9%, and isn't — it's the same finding in a stricter metric. The three source lists are nearly the same length (8.75 sources per query for the SERP sample, 9.24 for AI Overviews, 9.68 for Gemini), so a Jaccard of 0.18 works out to an intersection of roughly 2.7 sources per query, which is about 30% of an AI Overview's source list appearing among the ranked results. That conversion is our back-of-envelope from the paper's published averages, not a figure the authors report — but it puts the peer-reviewed measurement within eight points of the vendor one, on a different corpus, with a different instrument, at 75x smaller scale. For a literature whose headline numbers span 12% to 93%, that is what convergence looks like.
The paper also produces three findings the vendor studies structurally couldn't:
Generative surfaces reach further down the web's popularity curve. Looking only at the top source per query, a site in the Tranco top 1,000 holds that slot for 52.7% of queries on traditional Search, 40.0% on AI Overviews, and 32.6% on Gemini. Lesser-known domains gain share exactly on the surfaces rank tracking can't see.
The generative lists are less stable than the SERP. Repeat the identical query and AI Overview sources overlap at 0.66 Jaccard against the SERP's 0.78; make cosmetic edits — contractions, abbreviations, punctuation — and AI Overview overlap drops 29% against the SERP's 14%. That is independent, peer-reviewed confirmation of the churn our citation volatility data shows from the outside.
Blocking Google's AI crawler has a measurable, surface-specific price. All 21 popular publishers the authors checked — the New York Times, CNN, BBC, ESPN, and others — block Google-Extended in robots.txt. Gemini never cited any of them, and their AI Overview citations were reduced too, while their ordinary rankings were untouched. The trade-off in blocking AI crawlers now has published numbers on Google's own surfaces.
The caveats are the mirror image of the vendor studies'. This is one two-day snapshot from one simulated device and location, taken in December 2025 — before the Gemini 3 upgrade to AI Overviews that Ahrefs credits with part of the early-2026 shift. And "sources returned" for an AI Overview means the full source list, not the two or three visible citations a user actually sees, which Ahrefs' slot data shows are the highest-ranking ones. What it settles is the layer underneath: at retrieval, the ranked list and the generative surfaces are already mostly different documents.
The third is the retrieval surface, and it is the one the 2026 data added. CiteLens ran both kinds of study within weeks of each other: 500 commercial prompts against AI Overviews, where 60% of cited domains fell outside Google's top 10, and a four-engine benchmark in which Google's AI Mode drew 93% of its citations from the top 10. Roughly 40% and 93%, one vendor, one month. The two studies differ in market and query style as well as surface, so the gap is not cleanly attributable to AI Mode alone — but it is a vendor's own pair of numbers, and no plausible split of the credit leaves the surface out. Whatever else is true, "Google's AI cites top-ranking pages X% of the time" is not a well-formed sentence until you say which Google product you mean.
Strip out the definitional noise and one real conflict remains: direction.
Ahrefs' two readings describe overlap falling — 85.6% of cited pages somewhere in the top 100 in July 2025, 69.1% by March 2026. BrightEdge's 16-month series describes it rising, from 32.3% of citations ranking organically at launch to 54.5% by the end, which it reads as Google progressively committing to ranked content.
Those trend lines point opposite ways, and the tidy resolution is that they cover mostly different periods. BrightEdge's window runs May 2024 to September 2025; Ahrefs' first datapoint lands in July 2025, and its second sits well past the end of BrightEdge's series. A sequence where overlap climbed through 2024 and early 2025 as AI Overviews matured, then fell in early 2026 as Google leaned harder on query fan-out — Ahrefs' own explanation, alongside the Gemini 3 upgrade to AI Overviews in January — fits both datasets without either being wrong. It is a plausible reconstruction, not an established one. Nobody has run one parser across the whole 24 months.
Two details from the 2026 data support the fan-out reading. Of the cited pages that did not rank in Google's top 100 for the keyword, 18.2% were YouTube URLs, and YouTube accounts for 5.6% of all AI Overview citations. A video that ranks for nothing in a text SERP is a natural product of a system that retrieved against a different sub-query than the one you typed.
The assistants are a separate problem and a starker one. When Ahrefs sent 15,000 long-tail queries to Google and Bing and the same prompts to four assistants in August 2025, the average overlap between assistant citations and Google's top 10 was 12%, and roughly 80% of assistant citations did not rank anywhere in Google for the original query.
Engine
Cited links in Google's top 10
Perplexity
28.6%
Copilot
8.6%
Gemini
8.2%
ChatGPT (in-text)
8.0%
ChatGPT (references)
6.1%
Perplexity's relative closeness to conventional search results is consistent with it behaving most like a search product. ChatGPT's 8% is the number worth sitting with: on the engine most buyers actually talk to, the pages it cites are almost never the pages ranking for the question.
Ahrefs' own caveat is the important one, and it cuts against reading these figures as a scoreboard: assistants "don't rank results in the same way search engines do." They rewrite and fan out the prompt before retrieving anything, so "does this page rank for the prompt the user typed" may simply be the wrong denominator. The overlap is low partly because the comparison is unfair — and from outside the system, the fair comparison is unavailable.
CiteLens's July 2026 benchmark puts the same engines far higher: Perplexity at 89% against Ahrefs' 28.6%, ChatGPT at 30% against 8%. A 60-point gap on Perplexity is not a rounding difference, and it is worth being explicit that nobody has reconciled it.
Engine
Ahrefs, Aug 2025
CiteLens, Jul 2026
Perplexity
28.6%
89%
ChatGPT
8.0%
30%
Claude
—
53%
Google AI Mode
—
93%
Three design differences plausibly account for most of it, and all three run in the same direction. CiteLens used 320 templated buyer queries where Ahrefs used 15,000 long-tail ones, and long-tail is precisely where a ranked SERP is thinnest and retrieval wanders furthest. CiteLens sampled a single national market in Turkish. And its methodology note says citations were "normalized to registrable domains" with "search engines and social platforms excluded as noise" — which removes the source class that CiteLens's own AI Overviews study found dominant, with 74% of answers citing YouTube and 84% citing forums or user-generated content. Strip out the sources least likely to rank and overlap rises mechanically.
That is a reading of the published summaries, not a replication. CiteLens distributed its findings by press release and offers the underlying dataset "on request" rather than publishing it, so the benchmark should carry less weight than the studies that show their working. What it does establish is that a defensible-looking method can produce 89% for Perplexity, which is a useful check on treating 28.6% as settled.
The most quietly destabilizing number in the 2026 batch has nothing to do with rank. CiteLens asked identical questions in Turkish and English across 126 categories and found AI Overviews shared 22% of their cited sources between the two.
If that holds anywhere near generally, it means roughly four in five sources change when the language does — and every headline overlap statistic in this article was measured in English, mostly on US SERPs. For a brand selling into multiple locales, an English-language benchmark is not a global read on AI visibility. It is one market's read, reported without the qualifier.
Surface and language are two axes that dwarf the disagreement between vendors. The third is the vertical, and it was the first one anyone measured.
Industry
AI Overview citations that rank organically
Change over 16 months
Healthcare
75.3%
+12.0 pp
Education
72.6%
+53.2 pp
Insurance
68.6%
+47.7 pp
E-commerce
22.9%
+0.6 pp
BrightEdge's spread from healthcare to e-commerce is 52 points — wider than the entire disagreement between Ahrefs and BrightEdge on top-10 overlap. In a YMYL category with heavy authority filtering, rank and citation still travel together closely. In e-commerce they barely relate, and coverage there went down rather than up.
Stack the axes and the published averages stop meaning much. A healthcare brand reading AI Mode and an e-commerce brand reading ChatGPT are not looking at slightly different versions of the same number. The nearest published figures for them are 93% and 8%, and nothing in the data says either is wrong.
Which means the practical answer to "does ranking get me cited" is that nobody can tell you from a published average. It is a measurement you take in your own category, in your own languages, on the surfaces your buyers actually use.
None of these studies establishes causation in either direction. They are overlap counts: the share of cited pages that happen to rank. A page can rank well and be cited because both follow from being genuinely good on the topic, with neither causing the other. Discovered Labs' regression on 2 million citation observations found prompt-content alignment (β = +0.37) far ahead of any page-level proxy, which is what you would expect if rank and citation are two symptoms of the same underlying fit.
It also does not license the popular overcorrection. "Rankings don't matter for AI" is not what a 37.9% top-10 overlap says; a pathway that carries nearly two in five citations is not a rounding error, and Google's guidance on its AI features is explicit that they run on core Search ranking and quality systems. The defensible claim is narrower and more useful: rank is neither necessary nor sufficient for citation, so you cannot infer your AI visibility from a rank report. That is a measurement conclusion, not a strategy one.
Finally, this is only the part of the problem that happens on pages you could plausibly rank. Most AI citations point at domains you do not own at all — 84% of them, in the largest dataset published — and no amount of ranking reaches those.
The instrumentation is straightforward, and it replaces an inference with an observation.
Record every cited URL for your prompts, on every engine you care about, not just the citations that land on your domain.
Join each one to its Google rank for the same query, captured the same day. Stale rank data joined to fresh citations measures nothing.
Report the top 10 and the top 100 as separate figures. Collapsing them is the single most common way this stat gets mangled.
Split your own domain from third-party pages. Whether your citations come from your ranking pages is the version of the question that changes what you work on next.
Segment by prompt type. Definitional, comparison, and long-tail commercial prompts pull from different pools and will not share an overlap figure.
Keep surfaces and languages apart. AI Overviews and AI Mode have been measured 53 points apart in the same month, and translating a prompt can turn over most of its sources. Blend them and you get an average that describes no one.
Re-measure after engine changes. Overlap moved when the model behind AI Overviews changed. Any number you carry forward for a year is a number about last year.
If your own overlap comes back high, your rank data is a decent leading indicator and you can lean on it. If it comes back at ChatGPT's 8%, you have been reading a dashboard that cannot see the thing you are trying to manage.
Rank tracking answered a question with a stable shape for twenty years: where do I sit in an ordered list. AI citation does not have that shape. There is no list, the retrieval query is often not the query the user typed, and a third of the sources in a Google AI Overview are pages that do not rank in the top 100 for it.
The instrument did not break. It just stopped being a proxy for the thing everyone now cares about, and the studies above are best read as successive attempts to measure how far apart the two have drifted rather than as competing facts. Each new dataset has added an axis rather than a verdict: first the rank range, then which citations get counted, then the vertical, then the surface and the language — and now, from the peer-reviewed side, stability and popularity. That is what an unsettled measurement problem looks like from the outside, and it is a reason to instrument your own case rather than wait for the field to converge on a number.
Elmo is an open-source, self-hosted AI visibility platform that runs your prompt sets across ChatGPT, Claude, Gemini, Perplexity, and Google's AI surfaces, recording every mention, cited URL, and competitor named alongside you on a schedule. Because it stores the full citation set for every run and the data is yours to query directly, the overlap analysis above is a join against your own rank data rather than a vendor's average.