New to query fan-out? Start with our plain-English explainer. Already know it? This page is about what that means for your visibility, and what Google’s September 2026 research does and does not tell us.
Nothing is hidden. Sections outside your choice are dimmed, and hovering brings them back.
One question, a much wider investigation
Your customer asks one visible question. The machine may carry out a much wider investigation. Being strong for the first does not guarantee you survive the second.
For twenty years, visibility meant where you ranked. If you were first for the question your buyer typed, you were the answer. AI search splits that into two outcomes. There is still a ranking, and there is now also an answer, written by AI, that names some businesses and not others. The businesses it names are not always the ones at the top of the list.
Here is how that can happen. AI search can turn one question into several separate searches, each looking at a different part of what the buyer needs to know. This is called query fan-out. The answer the buyer reads is then written from what those searches brought back.
Try it below. In each example you rank first on Google. Keep an eye on your #1 card while the research comes back, and press Research it again a few times.
The Research Desk
AI search can turn one buyer question into several different research tasks, gather evidence for each, then write its answer from what came back. The business in blue ranks #1 on Google. Watch what happens at the desk.
A simplified teaching model, not a diagram of Google’s live systemSimplified model: the pile shows the idea, not how Google scores anyone.
Try adding a question to the desk
Not every question gets a research card. Pick one and see whether it passes the three checks.
Your buyer asked one question. The answer was built from the evidence for six or seven. Ranking for the question you can see is not the same as being the strongest evidence for the questions you cannot.
The three checks, in plain English
Google Research has published work on how to choose a good set of fan-out searches. Stripped of the terminology, it asks three things of every search in the set:
- Stays on brief. Every search still serves the buyer’s original question. The researchers call this alignment.
- Asks different things. Each search explores a genuinely different angle, instead of rewording one already asked. The researchers call this diversity.
- Finds real evidence. Each search points at something that actually exists and can be retrieved. The researchers call this groundedness.
These are the three meters on the Research Desk. Add a question to the desk and each meter rejects a different kind of bad question: one that drifts off the brief, one that repeats a question already asked, and one that nothing real could answer.
The same research teaches a much smaller model to reproduce that behaviour. It produces the whole set of searches in one pass instead of one after another, which Google reports is 12 to 20 times faster.
What Google demonstrated, and what it did not
This is where most AI search commentary blurs the lines, so we keep them separate.
The research is about choosing the searches. It is not research into how the final answer is written. We keep those two things separate on this page, and so should anyone quoting it.
What one study observed
In August 2026, an industry study of 120 buying searches across twelve categories, from CRM software to espresso machines, captured Google’s AI answer and the organic top ten from the same results page at the same moment.
Of the 79 searches with a usable AI answer, the site ranked first was named in the answer 41 times, which is 52%. In 78% of searches, the AI answer named at least one site from outside the organic top ten. The most common of those was YouTube.
Treat this as an illustration, not a benchmark. It is one day, in the US, on buying searches only, and its method has not been independently reviewed. The study itself says it shows how far apart the two lists sit, not why. What it does show is that the ranked list and the named list are not the same list.
Why this matters to your budget
If you judge visibility by rankings alone, you can be first and still be missing from the answer your buyer actually reads. Ranking remains an important route to visibility, but ranking and being named are different outcomes, and you need to measure both.
That changes the question you pay to answer. It is no longer only “do we rank for this?” but also “when a buyer’s question is researched, what evidence exists about us for each part of it, and where?” Most of what closes that gap is work you can see and control: specific answers on your own pages, and independent sources that confirm them.
Where this sits in our frameworks
In our five-stage model of AI recommendation (Recognition, Validation, Selection, Citation, Action), ranked but not named is a Selection outcome: the business was available to be chosen, and the evidence pointed elsewhere.
On the Lift, being named is a Floor 3 outcome: Trust and Selection, where AI has enough trust to name and recommend you, not just retrieve you. A business that ranks but is not named can be held up on Floor 2, where its content cannot be extracted cleanly, or on Floor 3, where independent evidence is too thin. The Lift is the way to find out which.
What to do: effort in, benefit out
These steps follow from the inference above, not from a confirmed change to how Google ranks pages.
| Step | Effort | What it gets you |
|---|---|---|
| List your research cards. For your five most valuable buyer questions, write down the separate things a buyer needs to know: price, process, proof, comparison, fit and risk. | 1 to 2 hours | A defensible list of distinct buyer facets worth making answerable |
| Check each card against your site. For every facet, is there a page that answers it specifically, with a real number, process or example? “Contact us for pricing” does not answer the pricing card. | Half a day | A gap list you can prioritise |
| Answer the missing cards. One clear, specific answer for each facet. Do not write ten pages for ten wordings of one question. | 1 to 2 days per facet | Evidence that can come back on the cards you were missing |
| Build independent evidence for the cards that matter most. Reviews, directory listings, comparisons and video are where claims get confirmed. | Ongoing | A better chance of being named, not just found |
| Test your list against real behaviour. Free tools such as Jean-Christophe Chouinard’s ChatGPT fan-out bookmarklet show the searches ChatGPT actually ran for a conversation. Search Console shows questions where you sit just outside the top positions. | 20 minutes | Your list of facets, checked against what systems actually do |
| Measure whether you are named. Ask your ten most valuable buyer questions in the AI tools your buyers use, every month, and record who is named and which sources are cited. | 1 hour a month | A visibility measure that matches what buyers now see |
Note what is not on the list. Because the searches change from one run to the next, there is no fixed set of six queries to reverse-engineer. The durable approach is to cover the facets that matter to your buyers.
If you are deciding where to invest
- If you rank well and are rarely named, the gap is usually evidence, not ranking. Steps 2 to 4 are where the budget goes.
- If you are neither ranked nor named, start with the foundations: see how AI decides which businesses to recommend.
- If you are named for some questions and not others, compare the cards. The difference is usually one or two missing facets.
Update, October 2026. Two independent sources now say the same thing from different directions. Wil Reynolds of Seer Interactive (2 October 2026) describes being visible and anti-recommended: an AI answer that names a brand only to tell the reader to look beyond it, which is a worse outcome than not being named at all. And Semrush’s AI Visibility Index, as reported by Leigh McKenzie in Search Engine Land on 24 September 2026, found that 40 to 60 percent of the sources cited across 2,500 prompts changed from one month to the next. Both are observations, not mechanisms, and the second is a vendor’s own data reported by a vendor-owned publication. They support the practical consequence of this page rather than change it: a single run of a question tells you little, repeated runs tell you whether you are on the right track, and the work is still evidence a buyer would trust.
