Quick answer
Rank and AI citation genuinely diverge, but the circulating “collapse” figures are not measuring the same thing. Ahrefs’ own AI Overviews number moved from 76.10% to 37.10% (blue links only) between July 2025 and March 2026 — while its parsing and its sample changed too. AI assistants are a separate measurement, not a later point on the same line: Ahrefs put that figure at about 12% in August 2025. The widely quoted “70% to under 20%” we could not trace to any published methodology, so we do not use it.
What we could actually trace
The figure changes with every retelling: 76%, 54%, 38%, 20%, 17%, 12%. Here is what sits behind each.
The two Ahrefs AI Overview studies: 76%, then 38%
In a study published 21 July 2025, Ahrefs analysed “1.9M citations from 1M AI Overviews … analyzing the top 3 most visible citations in each response”. The headline: “76.10% of AI Overview-cited pages rank in the top 10”, with 9.50% ranking 11–100 and 14.40% not ranking at all.
On 2 March 2026 Ahrefs published an update: “This time around, we analyzed 863K keyword SERPs, and a grand total of 4M AI Overview URLs—over double our last analysis.” Counting all returning result blocks, “37.9% of URLs cited in AI Overviews also appeared within the first 10 blocks”. Restricted to blue links only: 37.10% in the top 10, 26.20% at 11–100, 36.70% outside the top 100. Ahrefs summarised it as “~76% in July 2025 vs. ~38% today”.
That is the best evidence of divergence we found, but it is not a clean time series, and Ahrefs says so: “Since that last study we’ve also improved our parsing methodology in Ahrefs, so that we can see even more of the citations that appear in AI Overviews.” A second difference is our own reading of the two write-ups rather than a caveat Ahrefs draws: the 2025 study looked only at the three most visible citations per response, while the 2026 study covered 4M AI Overview URLs across 863K keyword SERPs — well over three per SERP. Restricting a sample to the most prominent citations should raise measured overlap, because the most prominent citation is the one most likely to already rank. On our reading, then, some unknown share of the drop is a change in what was counted. Ahrefs does not quantify that share, and neither can we.
The assistant number is a different measurement
The figure most often blended into the same paragraph comes from a separate Ahrefs study, published 11 August 2025, covering assistants rather than AI Overviews. Its method: “Using data from Ahrefs Brand Radar, our data scientist, Xibeijia Guan, searched 15,000 long-tail queries in Google and Bing, then asked those same questions to AI assistants.” The result: “In a dataset of 15,000 prompts, we found that—on average—only 12% of links cited by ChatGPT, Gemini, and Copilot appear in Google’s top 10 results for the same prompt.” Perplexity was the outlier at 28.6%; ChatGPT’s in-text citations hit 8.00%. Ahrefs’ conclusion: “AI Overviews follow the SERPs—AI assistants don’t.” The two surfaces should never be plotted on one line.
The number that points the other way
BrightEdge’s 16-month analysis, published 18 September 2025 and covering nine industries from May 2024 to September 2025, is headlined “AI Overview Citations Now 54% from Organic Rankings” and reports that “overlap grew from 32.3% to 54.5%”. The same page also says: “Only 16.7% of citations come from top 10 results, suggesting Google seeks diversity within ranked content.”
Both numbers come from one dataset on one page and differ by 37.8 points — because “ranks anywhere” and “ranks in the top 10” are different questions. That explains most of the variance in this year’s figures. When you see a bare “17%”, check which of the two it is.
The figure we could not trace
The most-shared version is a 5WPR press release of 4 May 2026, headlined “New 5W Research: Overlap Between Top Google Rankings and AI-Cited Sources Has Collapsed From 70% to Under 20%”. Its own figure line hands the number to someone else: “Drop in overlap between top Google ranking pages and AI-cited sources, per Brandlight analysis cited in the 5W report.” The 5W guide it points to lists that source only as “Brandlight analysis of Google/AI citation overlap” — no link, no sample size, no query set, no date range, no engines named. We checked Brandlight’s own blog and could not find the analysis published there. Separately, 5W’s State of AI Citations 2026 says in its own methodology note that it “is a synthesis of publicly available citation research published between August 2024 and April 2026” — a literature review, not a new comparison — and the claim does not appear in it at all.
So: this number is repeated widely and we could not find the study behind it. That is the finding, and we would rather publish it than launder someone else’s headline.
Why rank and citation really do diverge
Strip out the disputed figures and the mechanism is documented by Google itself. Its AI features and your website documentation states that “Both AI Overviews and AI Mode may use a ‘query fan-out’ technique — issuing multiple related searches across subtopics and data sources — to develop a response.” Five consequences follow.
- Retrieval is not ranking. An answer is assembled from many sub-searches and merged. The page that wins the merge shows up acceptably across several sub-queries, rather than winning a single SERP.
- The engine may never search your query. If the model rewrites “best CRM for a 12-person sales team” into eight narrower questions, the citations answer those — not yours.
- Freshness weighting differs. Rankings can sit still for months; a generated answer re-runs retrieval whenever it decides to search, so a new page can be cited long before it settles in the SERP.
- Extractability beats page-level relevance. Retrieval works on passages. A page whose answer is buried in paragraph nine can rank well and still lose to one stating the same fact in a self-contained sentence.
- Source-type preferences are category-specific. Ahrefs’ 2026 update found that among AI Overview cited pages not ranking in Google’s top 100 for the same keyword, “18.2% were YouTube URLs” — a source type most SEO reporting does not track.
Ranking signals versus citation signals
| Dimension | Classic SEO ranking | AI citation |
|---|---|---|
| Unit that competes | The page, against other pages, for one query | The passage, across many rewritten sub-queries |
| What it rewards | Topical authority, links, intent match and technical health for the target keyword | A clean, self-contained answer to a sub-question, and presence in the source types that category draws on |
| How it is measured | Position for a tracked keyword; clicks and impressions in Search Console | Share of prompts where you are cited, and who is cited beside you. Google’s generative AI report publishes impressions, not positions |
| Sample needed for a reading | One check per keyword is roughly repeatable | Many runs. Semrush found only 25.6% of cited domains overlap between ChatGPT’s own reasoning modes on identical prompts |
| How fast it moves | Slowly and directionally; a trend is meaningful over weeks | Fast and noisily; only the trend across a fixed prompt set means anything |
Measure the overlap on your own query set
Industry averages cannot tell you whether your rank and your citations agree — the answer is category-specific. Run it yourself.
- Freeze a prompt set. Twenty to fifty prompts real buyers type. Fix the wording — reword one and the series restarts.
- Fix the run conditions. Same engine, mode, country, logged out, personalisation off. Record the mode: Semrush’s 30 June 2026 study of 100 prompts run twice through GPT-5.2 found “only 25.6% of cited domains overlap between minimal and high reasoning for the same prompts”, with minimal reasoning running 245 web searches against high reasoning’s 1,130. Comparing an Instant-mode week to a Thinking-mode week measures your settings, not the market.
- Record every cited domain and URL per prompt, with the date, then run the same prompt as a plain Google search and record the top 10 organic URLs.
- Compute two overlaps. URL overlap (same page in both) and domain overlap (same site). They differ substantially, and reporting only one is how the public figures got so far apart.
- Repeat weekly, same day and time, reading the trend rather than the level. Our note on how often to sample AI answer citations covers cadence.
Add two columns: whether you were cited with a link or merely named, and whether the answer recommended you — separate outcomes, as we have covered in the gap between being cited and being recommended. Google’s generative AI performance reports, announced 3 June 2026, are a useful second source for impressions; we cover what they do and do not show separately.
What to do when they diverge
Divergence is not a reason to abandon rankings — Google’s documentation is explicit that “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” It changes where marginal effort goes.
Ranking but not cited usually means extractability or coverage: the page answers the head term but not the fan-out’s sub-questions, or not in liftable form. Put the direct answer in a standalone sentence under the relevant heading, then expand, and cover adjacent questions in the same cluster.
Cited but not ranking is a page doing work your rank tracker cannot see. Do not prune it as underperforming: flag it, keep it current, cite your own sources on it.
Neither means checking which source types get cited for your prompts first. If the answers draw on forums, documentation, review sites or video, more blog posts will not change the outcome.
FAQ
Does ranking in Google still help you get cited by AI?
Yes, but not on its own. Google’s AI features and your website documentation (last updated 10 December 2025) states: “The best practices for SEO remain relevant for AI features in Google Search (such as AI Overviews and AI Mode). There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.” The measured overlap between rank and citation is only partial, so treat ranking as groundwork rather than a guarantee.
Why do published overlap figures range from 12% to 76%?
Because they answer different questions. Some measure AI Overviews and some AI assistants; some count a cited page as overlapping if it ranks anywhere, others only if it ranks in the top 10; some compare URLs and some domains. BrightEdge’s September 2025 analysis reports 54.5% overlap with organic rankings and 16.7% overlap with the top 10 from the same dataset. Check two figures define overlap the same way before comparing them.
Is the “70% to under 20%” collapse real?
We could not verify it. The figure is attributed to a Brandlight analysis in a 5W press release and practice guide, but neither publishes a sample size, query set, date range or engine list, and we could not locate the underlying study. We have excluded it from this article. The best-documented decline we could trace is Ahrefs’ AI Overviews figure moving from 76.10% in July 2025 to 37.10% in March 2026 on blue links — and Ahrefs discloses that it improved its parsing methodology between the two studies. That the sample definition also changed is our own comparison of the two write-ups, not a point Ahrefs makes.
How often should I re-measure?
Weekly, on a fixed prompt set, at a fixed time. Single readings are close to meaningless: Semrush’s study of 100 prompts run twice through GPT-5.2 found only 25.6% of cited domains shared between the model’s own minimal- and high-reasoning modes. If an engine disagrees with itself that much, one run cannot tell you whether your visibility changed.
Can Search Console tell me whether I am cited?
Partly. Google announced its Search Generative AI performance reports on 3 June 2026, giving “dedicated views of your impressions within generative AI features on Search, such as AI Overviews and AI Mode, as well as generative AI features in Discover”. The published metric list covers impressions, pages, countries, devices and dates, and Google says they are rolling out “to a subset of websites”. That tells you a page appeared; it does not tell you what the answer said about you, and it does not cover engines outside Google.
Where this leaves you
Rank and citation overlap partially, the overlap differs by engine and category, and nobody has published a figure clean enough to plan a budget around. Measure your own, week over week, and treat any borrowed statistic without a sample size and a date as marketing rather than data. Zian AI builds autonomous sales agents — SmartReach AI™ and PrecisionPitch AI™ — for teams that would rather run the measurement than repeat the statistic. We are in partnership-application beta: Apply For Partnership to talk to us.