The comfortable version of AI search went like this: keep ranking well, and the AI layer will keep quoting you, because it quotes what ranks. It was a reasonable assumption. It is now hard to defend.
In an analysis of 863,000 keywords and roughly 4 million AI Overview URLs published in March 2026, Ahrefs found that 38% of cited pages also appeared in the top 10 organic results for the same query. In their earlier run of the same study, that figure was 76%. The remainder splits almost evenly between pages ranking 11–100 (31.2%) and pages ranking beyond position 100 (31.0%).
One honest caveat before the conclusions
Ahrefs attributes part of the drop to their own improved parsing, which now catches citations the earlier method missed. So do not read 76 → 38 as a clean measurement of the world changing by half. The number to act on is the current level, not the delta: today, roughly six in ten AI Overview citations point at pages that do not rank in the top 10 for the query being answered. That holds regardless of what the old figure really was.
The mechanism is query fan-out
Google does not simply answer the question you typed. It decomposes it into related sub-questions, retrieves results for each, and builds the answer from the pages that keep showing up across that set. Ahrefs points to this fan-out as the likeliest driver: the pages that recur across many sub-queries get cited, and a page can recur across a dozen narrow sub-queries without ever winning the broad one.
Which flips the competitive picture. You are not competing against the nine other results on the page anymore. You are competing across the fifteen questions hiding inside the one the buyer asked — and the page that ranks 43rd for the headline term while being the single best answer to three of those sub-questions is the page that gets quoted.
Rank is a bet on one question. Citation is a bet on all the questions behind it. A page can lose the first badly and still win the second.
What this actually breaks
- Rank as a proxy for visibility. "We hold position 2" no longer implies the AI answer mentions you, and Search Console's AI report will not tell you either — it shows impressions without the query.
- The pillar-page reflex. Long generic hub pages are built to win one big term. Fan-out rewards the opposite shape: narrow pages that answer one question completely.
- The improvement loop. Teams push a page from 8 to 4 and see nothing change in AI answers, then conclude AI search is unmeasurable — when the actual problem is that they moved the wrong lever.
- Competitive reads based on the SERP. The brands cited above you in an AI answer are frequently not the brands ranking above you, so a rank-tracker view of the category quietly omits your real rivals.
What to do instead
- Write the fan-out yourself. Take your main buying query and list the fifteen sub-questions a model would have to resolve to answer it: pricing, limits, integrations, migration, who it is not for, how it compares. That list is your content plan.
- One answerable question per page, with the answer in the first sentence. Setup paragraphs are where citations go to die — a retrieval system lifts the conclusion, not the throat-clearing before it.
- Prefer specific over comprehensive. A 700-word page that fully resolves one sub-question outperforms a 4,000-word guide that touches it in a subheading.
- Look at which of your pages actually get cited, not which rank. It is very often not the page you optimised, and that mismatch is the cheapest insight available to you. If nothing of yours is cited at all, work the four failure modes before rewriting anything.
- Track citations and mentions as their own metric, on their own schedule, next to rank rather than downstream of it.
Rank still matters — it is just no longer the scoreboard
None of this argues for abandoning classic SEO. Ranking well still means being crawled, indexed, trusted and retrievable, and those are preconditions for being pulled into an answer at all. What has ended is the convenient assumption that rank and citation are the same measurement wearing different clothes. They correlate loosely now, they are produced by different processes, and if you only instrument one of them you are flying on the wrong instrument.
