4 min read
For twenty years, one number answered the question “how are we doing in search?” It was your position for a keyword. That number still exists, still matters, and is now measuring something narrower than most people think it is.
The reason is straightforward. The pages an AI assistant quotes are largely not the pages Google ranks for the same question.
The number
12%of the URLs cited by AI assistants also appear in Google’s top ten for the original prompt. Ahrefs ran 15,000 long-tail prompts through ChatGPT, Gemini, Copilot and Perplexity and matched every citation against the ranking results, July 2025.
Put the other way round: 88% of what gets cited does not rank in the top ten. And 80% of it does not rank in the top hundred at all.
There is a real methodological caveat and it is worth stating rather than burying. The comparison is against Google’s results for the conversational prompt itself, and long, spoken-sounding prompts often have no meaningful results page to compare against. Part of the low overlap is that the two systems are being asked genuinely different questions. That does not make the finding less useful — it is exactly the point — but it does mean the number describes a mismatch of questions as much as a mismatch of pages.
The counter-evidence, because it exists
Seer Interactive published a study in early 2025 finding that 87% of SearchGPT citations matched Bing’s top organic results — close to the opposite conclusion. Ahrefs, on the same relationship, found 8%.
We weight Ahrefs, because its sample is roughly thirty times larger and its matching is stricter. But anyone who quotes one of these numbers at you without mentioning that the other exists is selling, not reporting.
AI Overviews are a different animal from assistants
This is where most of the confusion in the market comes from. “AI answers” is not one surface. Google’s AI Overviews draw far more heavily on conventional rankings than the chat assistants do — 37.9% of AI Overview citations appear in the top ten, against 12% for the assistants.
Ahrefs reported 76.1% for that same AI Overview measure in July 2025, then 38% in March 2026 — a revision it attributes partly to improved parsing and partly to Google’s query fan-out. When a publisher moves its own headline finding by half in eight months, the honest response is to date every figure you quote and expect it to move again.
What this changes about the work
Three things, in order of how much they matter.
- Coverage beats position. A prompt becomes several searches — see query fan-out. What you want is to be a credible answer across many phrasings of the buying question, not the top result for one of them.
- Retrievability beats optimisation. A page that states a claim plainly, once, in the customer’s language can be quoted. A page that takes four paragraphs to arrive at the point can be ranked and never quoted. See grounding.
- Your identity has to resolve. A recommendation is a statement about a business, not a document. If the systems cannot resolve scattered mentions of you into one entity, there is nothing for them to recommend. See entity resolution.
What it does not mean
- It does not mean SEO is finished. Ranking still drives the majority of high-intent clicks, and the same content that ranks is the substrate the models read.
- It does not mean you should stop measuring rank. It means one number is no longer the whole report.
- It does not mean anyone can promise you a position in an AI answer. These systems are non-deterministic; the same prompt twice can return different names. Anyone guaranteeing placement is describing something they do not control.
The measure that replaces it
Not a rank. A rate. Take a fixed set of real buying questions for your category and city, run them repeatedly across the assistants, and count the proportion where you get named. That proportion is your share of AI voice, and unlike a rank it survives the fact that the system is non-deterministic. It is a floor, not a score.
Sources: Ahrefs, AI Search Overlap Study, 2025 · Ahrefs, AI Overview Citations Study, 2026 · Seer Interactive, SearchGPT citation overlap, 2025.
Definitions: AI Overviews, retrieval-augmented generation, share of AI voice. Or see it measured on your own business with a free AI Visibility Snapshot.