Most research about AI search is speculation dressed up with numbers. This one is not.
In April, Ahrefs published an analysis of 1.4 million ChatGPT prompts, tracking which pages the model retrieved and which of those it went on to actually cite. It is the largest look at this I have seen, and unlike most of what gets written about AI visibility, it produces advice you can act on this afternoon.
Three findings stood out. One of them changed something in how I write pages.
Finding one: being found is only half the battle
Only about half of the pages ChatGPT retrieves end up cited. Just under 50%, across the whole dataset.
I have written before that ranking has become the easy part, and this puts a number on it. Your page can be relevant enough to be pulled into consideration and still never appear in the answer. Retrieval gets you into the room. Citation is a separate door.
The gap is even starker in places. Reddit pages were retrieved over sixteen million times in the dataset and cited less than 2% of the time.
Which is worth remembering next time somebody tells you to go and post on Reddit for AI visibility. It gets read constantly. It rarely gets credited.
Finding two: your title needs to match the sub-question, not the question
This is the mechanical one, and it is the most useful.
When you ask ChatGPT something, it does not search for your question. It generates its own sub-questions, called fan-out queries, and searches for those instead.
The study measured how closely page titles matched, and the pattern is clear:
- Titles matching the original prompt: cited pages scored 0.602, non-cited 0.484
- Titles matching the fan-out queries: cited pages scored 0.656
So alignment with the fan-out query is a stronger signal than alignment with the question the human actually typed.
That is a genuinely different way to think about titles. You are not writing for the question somebody asks. You are writing for the smaller, more specific questions the machine generates on their behalf.
And there are fewer of them than people claim. You will see figures like “three to twelve searches per prompt” thrown about. Ahrefs found an average of 1.78, with 75% of prompts triggering exactly two searches. That is far more manageable, and it means you are aiming at a small number of specific sub-questions rather than trying to blanket a dozen.
What that means in practice
Think about what somebody actually needs to know, not just what they asked.
Ask an AI “who should fit my bathroom in Lytham” and it is not going to search that phrase. It will break it into things like what bathroom fitting costs, who works in that area, how long it takes.
Those are your page titles. Not the big question, the small ones underneath it.
Which is why a page titled “What a bathroom costs on the Fylde coast in 2026” will out-perform one called “Bathroom Installation Services.” One matches a sub-question. The other matches a brochure.
Finding three: the one that surprised me
Here is the finding I have not seen anybody talking about, and it is embarrassingly easy to act on.
Pages with natural language URL slugs were cited 89.78% of the time. Pages without were cited 81.11%.
Nearly nine points, purely from what the address of your page looks like.
By natural language slug, they mean a URL that reads like words a person would say:
- Good:
yoursite.co.uk/bathroom-fitting-costs-lytham - Bad:
yoursite.co.uk/services/page-id-4471 - Also bad:
yoursite.co.uk/s/b/2026/04/post-name-final-v2
It makes sense once you see it. The URL is one of the few things the model sees before deciding whether to open a page at all. It gets a title, a snippet and an address, and makes a judgement from those. A readable address is another signal of what the page is about. A meaningless one is a wasted opportunity.
And it costs nothing on a new page. Every page you publish from now on can have a clean, readable, descriptive URL.
Be careful on existing pages, though. Changing the URL of a page that already ranks means redirects, and it means losing some of the authority that page has built. On an established page earning its keep, leave it alone. Nine percentage points is not worth breaking something that works.
What this changes about how I write a page
Putting the three together, the practical version:
Write the title for the sub-question. Specific, plain, phrased the way somebody would actually ask it. Not the category, the question.
Make the URL readable. Words, hyphens, no identifiers, no dates, no version numbers. Set it right when the page is created.
And remember that appearing is not winning. Half the pages retrieved never get cited. Being relevant enough to be considered is the beginning, and everything after that is about whether your page contains something clean and specific enough to be quoted.
The bit worth stepping back for
There is something reassuring buried in these findings.
Match your title to what people are actually asking. Give your page a sensible, readable address. Be specific rather than vague.
That is not new advice. Anybody doing search properly has been saying it for fifteen years, and it was true long before ChatGPT existed.
Which is broadly what Google itself has said about all this: optimising for generative AI search is optimising for the search experience, and it is therefore still SEO.
The mechanism has changed. The instruction has not. Be clear about what you do, in the words people use, in a place that is easy to find and easy to read.
I would love to tell you there is something cleverer than that. Fifteen years in, there mostly is not.
Want to know whether your pages are set up for any of this? Get your free audit, or email hello@inkon.co.uk. I read every message myself.