Only about 12% of the URLs that AI assistants cite are ranking in Google’s top ten for the question that prompted them. For ChatGPT specifically it drops to 8%. That figure comes from Ahrefs, which ran 15,000 long-tail queries through Google and Bing, put the same questions to four AI assistants, and mapped every citation back against its ranking position.
The obvious conclusion — that rankings no longer matter — is wrong, and acting on it will cost you. The accurate conclusion is narrower and more useful: your page is being judged against a question nobody typed.
Three studies, one finding
This is not a single anomalous result. Independent teams using different methods have landed in the same territory.
| Study | Sample | Finding |
|---|---|---|
| Ahrefs, July 2025 | 15,000 long-tail queries, 4 assistants | 12% of AI-cited URLs rank in Google’s top 10 |
| Surfer SEO, December 2025 | 173,902 URLs across 10,000 keywords | 68% of pages cited in AI Overviews were not in the top 10 |
| Position Digital, July 2026 | 278 prompts, 3,508 citations, B2B SaaS | 58% of Google top-10 results were also cited somewhere |
That third row looks like it contradicts the first two. It does not, and understanding why is the whole point.
The studies are asking opposite questions. Ahrefs asks: of the URLs the AI cited, how many were top-ten pages? Position Digital asks: of the top-ten pages, how many turned up somewhere in the citations? The first measures how much of the citation pool comes from high rankings. The second measures how much of the top ten gets used at all.
Both can be true simultaneously, because an AI answer draws on a far wider pool than ten results. Ranking well makes you eligible. It does not make you selected. Most of what gets cited was never in the running by conventional search standards.
The per-engine breakdown matters too, because “AI search” is not one behaviour:
| Assistant | Citations that rank in Google’s top 10 |
|---|---|
| Perplexity | 28.6% |
| Gemini | 8.6% |
| Copilot | 8.2% |
| ChatGPT (in-text) | 8.0% |
| ChatGPT (references) | 6.1% |
Perplexity leans on conventional rankings roughly three and a half times more than ChatGPT does. If your buyers use Perplexity, your SEO investment transfers reasonably well. If they use ChatGPT, it mostly does not.
What is actually happening: query fan-out
When you type a question into an AI assistant, it does not search for that question. It decomposes your prompt into a set of related sub-queries, runs them in parallel, and synthesises what comes back.
Typical volume is eight to twelve sub-queries for a standard prompt. Google has said its Deep Search mode fires hundreds. These are synthetic queries — generated by the model, never typed by any human, absent from every keyword tool you own.
Google’s patents on the technique — filed under the less catchy name “query variant generation” — describe eight classes of variant:
| Variant class | What it does | Example from “best sales intelligence software” |
|---|---|---|
| Equivalent | Rephrases the same question | top B2B contact data platforms |
| Follow-up | Asks the logical next question | how much does sales intelligence software cost |
| Generalisation | Broadens the scope | what is sales intelligence |
| Specification | Narrows to detail | sales intelligence tools with verified mobile numbers |
| Canonicalisation | Standardises the phrasing | B2B sales intelligence platform comparison |
| Entailment | Asks what the question implies | how accurate is B2B contact data |
| Clarification | Resolves ambiguity | sales intelligence for SMB or enterprise |
| Translation | Retrieves across languages | — |
Look at that column of examples. A page built to rank for “best sales intelligence software” targets exactly one of eight retrievals. Seven of them go looking for something it never set out to answer.
How the results get merged
The sub-query results are combined using reciprocal rank fusion, a merging method with one property that changes everything about how you should write: it rewards consistency across queries, not peak position on any single one.
A page sitting around position six for five of the eight variants will outrank, in the merged pool, a page sitting at position one for a single variant and nowhere for the rest. Breadth beats peak. That is the mechanical reason a #1 ranking so often fails to convert into a citation, and it is not a quirk to be patched — it is the design.
Why the sixth-place page wins
Consider two pages competing in the same category.
- Page A is a tightly optimised commercial page. It ranks #1 for “best sales intelligence software.” It says nothing about pricing, nothing about data accuracy, and never names a competitor.
- Page B is a broader guide. It ranks around #6 for the head term, but it also covers what the category is, how the data is verified, what it typically costs, and how four named vendors differ.
Page A appears in one of eight retrievals. Page B appears in five or six. After fusion, Page B is cited and Page A is not — despite Page A holding the ranking that every dashboard in the company is celebrating.
This is why the “one page, one keyword” discipline that served SEO for twenty years now works against you. Fan-out rewards pages that answer a cluster of adjacent questions well enough to be retrieved repeatedly.
How to find your own fan-out queries
You cannot see the synthetic queries directly. You can approximate them closely enough to act, and this takes an afternoon per priority topic.
- Write the eight variants by hand. Take your head term and walk the table above — equivalent, follow-up, generalisation, specification, canonicalisation, entailment, clarification. Fifteen minutes gets you a workable approximation of what the model will generate.
- Harvest People Also Ask and AlsoAsked. These expose real adjacent questions Google already associates with your term. Overlap with synthetic variants is high, because both are generated from the same underlying understanding of the topic.
- Read the AI Overview itself. Search your head term and note which sub-topics the generated answer chose to cover. That is the fan-out, partially revealed — the model is showing you which angles it considered load-bearing.
- Map your entity network. Your primary entity, adjacent concepts, attributes, and sub-entities. Wikipedia category pages and the Knowledge Graph are unglamorous but effective for this.
- Score your coverage. List the variants, mark which of your pages addresses each, and calculate the percentage covered. Aim for 80% of core themes. The gaps are your content plan, and they are far better prioritised than a keyword list sorted by volume.
What to change in how you write
Four adjustments follow directly from the mechanism.
Write in retrievable passages
Retrieval operates at passage level, not page level. Each section should answer one sub-question completely enough to stand alone if lifted out — because that is precisely what happens to it. A section that only makes sense after reading the three above it cannot be cited. Front-loading matters too: 44.2% of AI citations come from the first 30% of a page (SparkToro, January 2026).
Cover the cluster, not the keyword
Deliberately answer the pricing question, the accuracy question, the comparison question and the definitional question — even where each has a stronger dedicated page elsewhere on your site. Cannibalisation was a ranking concern. Under fusion, appearing in more retrievals is the objective.
Name the entities
Cited pages average roughly ten named entities per thousand words against eight for uncited ones, and pages naming six or more brands averaged 2.13 citations against 1.21 for those naming none. Naming competitors is how a model learns which category you belong to.
Keep it current
The median cited page is 3.9 months old and 69.7% were published within twelve months. Refreshing a page that already covers the cluster beats publishing a new one that covers a slice of it.
What this does not mean
Three honest caveats, because the 12% figure is being used to sell some bad advice.
It does not mean abandon SEO. Ranking is what makes you retrievable in the first place. A page that ranks nowhere for any variant is not in the pool for any of the eight retrievals. The finding is that ranking for one query is insufficient, not that ranking is worthless.
The study has a stated gap. Ahrefs measured citations against the original prompt’s rankings. They did not test whether those same pages ranked well for the likely fan-out sub-queries. It is entirely possible the cited pages were ranking — just for questions the study never checked. That would strengthen the argument for cluster coverage rather than weaken it, but it is an open question and worth saying so.
The numbers will move. Ahrefs ran in July 2025, Surfer in December 2025, Position Digital in July 2026. Retrieval architectures change quarterly. Treat the mechanism as durable and the percentages as a snapshot.
What to measure instead of rank
| Old metric | Replace with | How to get it |
|---|---|---|
| Position for head term | Mention rate across 30 buyer prompts | Manual logged-out audit, monthly |
| Keywords ranking top 10 | Fan-out coverage percentage | Variant list vs page inventory |
| Backlinks | Cited-URL inventory for your category | Record which URLs the AI cites |
| Organic sessions | AI referral sessions and their conversion rate | GA4 referral filters |
The cited-URL inventory is the sleeper. Every domain the model quotes for your category is a publication it already trusts on your topic — which makes that list the best-qualified outreach target you will assemble all year.
Frequently asked questions
Do Google rankings help at all with AI citations?
Yes, but indirectly and unevenly. Perplexity draws 28.6% of citations from top-ten pages; ChatGPT draws 8%. Ranking makes you retrievable across the sub-queries you happen to rank for. It is a necessary condition far more often than a sufficient one.
Can I see the actual fan-out queries for my topic?
Not directly — they are generated at query time and not exposed. People Also Ask, AlsoAsked, and the sub-topics an AI Overview chooses to cover are the closest available proxies, and in practice they overlap heavily with what the model generates.
Does this make keyword research obsolete?
It changes the unit. Individual keyword volumes matter less; the completeness of your coverage across a topic’s question cluster matters more. Keyword tools still tell you which clusters have demand — they simply cannot show you the synthetic variants inside each one.
Should we still build separate pages per keyword?
Build separate pages where search intent genuinely differs, but stop stripping adjacent context out of each one to avoid overlap. A page that touches pricing, accuracy and comparison is retrieved more often than four pages that each touch one. Some internal overlap is now an asset.
How long until changes show up in AI answers?
Retrieval indexes refresh continuously, so a substantially expanded page can be picked up within weeks. The constraint is usually your publishing cadence rather than the index. Re-run your prompt baseline monthly and watch mention rate rather than checking daily.
The takeaway
The 12% figure is not evidence that search optimisation stopped working. It is evidence that the unit of optimisation changed underneath it. You are no longer competing for a position on one query. You are competing to be retrieved across eight to twelve questions that nobody typed and no tool reports.
The page that wins is not the sharpest answer to the head term. It is the page that turns up, credibly, in most of the retrievals — and consistency across the cluster is a different thing to build than a #1 ranking.
Pick your most commercially important topic, write out the eight variants, and score how many your current pages actually answer. Most teams find the number is one.
Rankings are only half the picture. The signals that decide whether a model recommends you at all are mostly off your own domain — we covered those in why your website is the weakest signal in AI search.
Request a demo and we will run a live search against your best-fit account profile. Or explore the Sales Bundle, and read more in our B2B Growth Hacks topic hub.