Your website is being demoted from destination to data source
In July 2025, 76% of the pages Google cited in an AI Overview also ranked in the top ten for that search. By February 2026 it was 38%. Here is what changed, what it means for a small business, and the six things worth doing about it — four of which are free.
The number that changed
Ahrefs studied 863,000 keywords and 4 million AI Overview URLs. In their July 2025 run, 76% of the pages Google cited in an AI Overview also appeared in the organic top ten for that same query. In the version published on 2 March 2026, that had fallen to 38%.
The rest is the part worth sitting with. Of the pages now being cited, 31.2% rank between positions 11 and 100, and 31.0% rank beyond position 100.
Nearly a third of the pages being quoted to your customers rank nowhere at all.
If you have spent money getting to page one, that is an uncomfortable sentence. It does not mean ranking stopped mattering. It means ranking first for a question is no longer the same thing as being in the answer to it.
Why it happened
The mechanism has a name: query fan-out. When a search triggers an AI Overview, Google can split the original question into several related sub-questions, run those, and assemble an answer from the pages that appear most consistently across all of them.
So a page can be cited without ranking for the question you typed — it ranked for one of the sub-questions you never saw. And a page can rank first for your question and still be absent from the answer, because it did not appear across the rest of the fan.
Worth being precise: Ahrefs describe fan-out as the likely cause. Google has not confirmed any specific change to fan-out behaviour. This is a measured effect with a well-supported explanation, not an announced one.
What it does to a small business website
For most of the last twenty years a website had one job: get found, then persuade the person who arrived. Everything followed from that — the hero image, the layout, the button colour, the story above the fold.
Increasingly, nobody arrives. A model reads your pages, writes an answer, and your customer acts on the answer. Your homepage was never involved. The visit that used to begin the relationship now happens somewhere you cannot see, in a paragraph you did not write, about a business you own.
That does not make design worthless. It makes design insufficient, and it promotes a set of unglamorous things that used to be somebody else’s problem: can a machine read this page, can it tell what the page is about, can it tell who you are, and does the rest of the internet agree.
What actually correlates with being cited
Ahrefs also looked at 75,000 brands and compared two things against AI Overview visibility: how often a brand is mentioned across the web, and how many backlinks it has.
Brand web mentions: 0.664. Backlinks: 0.218. Roughly three times stronger, by Spearman correlation.
For a small business that translates into something concrete and slightly old-fashioned. Being talked about beats being linked to. A spa named correctly in a local guide, a map listing, a supplier’s page and three tourist blogs is far easier for a machine to resolve into one real business than the same spa with a handful of links and its name spelled five different ways.
Correlation is not cause. Well-known businesses get mentioned more and get cited more, and both may follow from being well known. The practical advice survives either way, because the work — be named consistently, be mentioned in real places — is worth doing regardless of which direction the arrow points.
Be careful who you pay to fix this
A market has appeared selling your “rank in AI”. Before buying one, look at what happens when the same question is asked twice.
SparkToro tested 2,961 AI responses across twelve prompts and three assistants — ChatGPT, Claude and Google’s AI — with 600 volunteers running them. Ask the same question a hundred times and the chance of getting back the identical list of brands is under 1 in 100. The identical list in the same order is closer to 1 in 1,000.
There is no position to track. A dashboard showing you one is showing you noise with a number on it.
What the same researchers do accept is a visibility percentage measured across dozens or hundreds of prompts, each run several times — how often you appear at all, rather than where you sit. That is a real measurement. If a vendor cannot tell you which of the two they are selling, that is your answer.
Six things worth doing, cheapest first
In the order a small business should actually do them. The first four cost nothing but time.
- Check whether an assistant can read your site at all
Not whether Google indexed it — whether the page delivers its content in the HTML rather than assembling it in the browser afterwards. If your text only appears once JavaScript has run, some crawlers see an empty page. Free, and about ten minutes with your browser’s view-source. - Give every page a description and one clear heading
The oldest advice in search, and the reason a page gets summarised correctly rather than guessed at. Free, and half a day for a small site. - Answer one question per block, in the words a customer would use
Fan-out means your page is matched against sub-questions, not only your headline. A heading that asks the question with a paragraph that answers it directly is far easier to lift than the same information spread across a page. Free. - Add structured data that matches what you actually are
A restaurant marked up as a restaurant, with its menu and its prices, is unambiguous. A page with no structured data is a guess. Free, a few hours. - Make your name, address and phone number identical everywhere
Every directory, every map, every social profile. Inconsistency is why a machine cannot tell whether three listings are one business or three — and it is the most common fault found in assessments run so far. Free, and mostly tedium. - Get mentioned by name in places that are not yours
Local guides, directories, suppliers, partners, press. This is the one that correlates at 0.664, and it is the slowest of the six. Free to ask for; it costs time and relationships rather than money.
Notice what is not on this list. Nothing here is a new tool, a subscription, or a rebuild. This is the same work that has always made a business easy to recommend — it is the thing doing the recommending that changed.
What would prove this wrong
Three things would take most of the urgency out of this article, and it is only fair to name them.
AI referral volume plateaus. The share of citations changed sharply. The number of people actually acting on AI answers is a separate question, and a smaller number than the noise around it suggests.
Google reverts. Advertising revenue depends on people clicking. There is a commercial argument for pulling back toward conventional results, and it is not a weak one.
The measurement problem turns out to be unsolvable rather than merely unsolved. If nobody can measure AI visibility reliably, the sensible response is to do the fundamentals and stop paying anyone to report on it.
If any of those happen, that gets written too. The monthly reports Aroc sends clients always include what did not work. This is held to the same standard.
Sources
- Ahrefs — AI Overview citations and organic rankings. 863,000 keywords, 4 million AI Overview URLs. The July 2025 run compared with the version published 2 March 2026.
- Ahrefs — brand web mentions versus backlinks. 75,000 brands. Spearman correlation 0.664 against 0.218. Louise Linehan and Xibeijia Guan.
- SparkToro — consistency of AI brand recommendations. 2,961 AI responses, 12 prompts, ChatGPT, Claude and Google AI, 600 volunteers.
Every figure in this article is quoted from the study that produced it, with its sample size. Where a claim is an interpretation rather than a measurement, it says so.