On 2 November I sat down to build a query fan-out tool. The pitch writes itself: paste a keyword, see the hidden searches Google’s AI runs behind it, get a score for how many of them your page answers. Free versions already exist and the advice that comes with them is close to universal (cover the sub-queries, get cited). Before I put a number like that in front of a client I wanted to know whether it meant anything, so I wrote the test down first, with the decision rules, and then ran it. It failed. I am publishing it anyway, because the failure is the useful part.
Why test fan-out before building a tool for it?
The mechanism itself is not in doubt. Google’s Search Central guidance on AI features says AI Overviews and AI Mode
may use a ‘query fan-out’ technique, issuing multiple related searches across subtopics and data sources
Google Search Central, AI features and your website
What Google does not say is that a page answering those searches is more likely to be cited. That is the step the whole tool category rests on, and as far as I could find nobody had tested it while holding rank constant. Every SEO knows the trap: pages that rank well also tend to be long, thorough and well linked, so almost any content measure “predicts” citation until you control for rank. So the pre-registration, written before any data was pulled, set one bar. Coverage had to improve out-of-sample prediction of which pages get cited, within the same query, by at least 0.02 of AUC on top of rank, length and domain rating. If it did not, we would not ship a coverage score.
What exactly did we measure?
- 235 non-branded keywords with AI Overview citations on record, drawn at random from 46 domains in our own citation corpus: 162 in the UK and 73 in the US.
- Observed fan-out: the searches Gemini actually ran with Google Search grounding (three runs per keyword, 705 calls) and the searches Claude ran with its web search tool (two runs per keyword, 470 calls). The APIs log these, so nothing is guessed.
- Guessed fan-out: a model asked to predict what AI Mode would search, which is what most free tools show you.
- Citation labels from Ahrefs AI Overview data, the head-query top ten, and the results pages for 539 sub-queries from DataForSEO.
- Every candidate page read with our own audit engine’s extractor and compared with each sub-query using bge-small embeddings.
The model is a conditional logistic regression within each query, so a page is only ever compared with the other pages competing for the same keyword. 158 queries, 1,190 pages and 670 citations made it into the final models. The model and results-page calls cost under 50 US dollars in total, which is worth saying out loud: this is not expensive research, it is just research hardly anyone runs before selling the answer.
Does answering the fan-out predict AI Overview citation?
No. Here is the primary result (the rank odds ratio sits below 1 because a lower rank number is a better position).
| Factor | Odds ratio | AUC change out of sample | Verdict |
|---|---|---|---|
| Better head-query rank (per SD of log rank) | 0.43 | baseline | Supported, p < 0.0001 |
| Longer page (per SD of log words) | 1.30 | baseline | Control, p 0.015 |
| Domain rating | 0.94 | baseline | Control, p 0.59 |
| Best heading matches the query | 0.81 | +0.004 | Not supported, p 0.11 |
| Ranks top ten for an observed sub-query | 1.30 | −0.001 | Not supported, p 0.65 |
| Share of observed sub-queries answered | 1.29 | −0.003 | Not supported |
| Observed sub-queries instead of guessed ones | n/a | +0.001 | Not supported |
Read the coverage row carefully, because it is the one people will want to quote. Inside the sample, pages answering more of the sub-queries were cited a little more often: pages covering 76 to 100% were cited 70% of the time, against 53% for pages that answered none. Out of sample it added nothing. The AUC change was −0.003, with an interval of −0.009 to +0.003, nowhere near the +0.02 bar. Coverage is riding on things rank and length already tell you.
And the guessed sub-queries did at least as well as the observed ones (odds ratio 1.52, p 0.002), which is the giveaway for me. If the real searches carried a mechanism, they should beat a model’s guess. They did not, so coverage looks like a proxy for a thorough page rather than a lever you can pull.
How much fan-out do AI systems actually run?
Far less than the tools show. Most of what gets sold as fan-out is a model imagining fan-out.
| Measure | Value |
|---|---|
| Sub-query clusters per keyword, observed | 2.35 |
| Sub-query clusters per keyword, guessed | 16.1 |
| Gemini runs with no search at all | 331 of 705 (47%) |
| Keywords where Gemini never searched in three runs | 77 of 235 |
| Gemini clusters that appeared in all three runs | 16.5% |
| Claude runs with exactly one search | 433 of 470 |
So a guessed list is around seven times longer than what actually ran, and what ran changes from one run to the next. Building a content plan around one capture of one run is building on sand. The guessed lists are not useless (58% of observed clusters had a close guessed match), but they are content research, not a map of what the machine did.
What nearly fooled us?
Three things, and the first two would each have made a headline.
First, DataForSEO now returns many result URLs as google.com/goto wrappers, 64% of the sub-query results in our pull. Matched naively, a wrapped URL never equals the cited page, which quietly deflates every sub-query measure. An independent check of the first analysis (a separate AI reviewer that had not seen the build) caught it. We resolved every wrapper and re-ran, and the decision held.
Second, the same-day replication (fresh head results and AI Overview references from DataForSEO on 3 November) first showed an odds ratio of 7.8 for sub-query ranking. That was the wrappers again, this time on the organic side. Resolved, it fell to 1.60 (p 0.051), and adding it to rank moved AUC by −0.012. A 7.8 would have been a very shareable chart and completely wrong.
Third, smaller but humbling: Ahrefs and DataForSEO, pulled a few weeks apart for the same keywords, shared only 13.5% of top-ten URLs and 30% of domains. If you are quoting anyone’s citation study, ours included, ask which snapshot it came from.
Why does this disagree with other studies?
Surfer’s study of 10,000 keywords found pages ranking for fan-out queries were 161% more likely to be cited than pages ranking only for the main query, and I do not doubt the count. The difference is the question. Pages that rank for lots of related searches are usually stronger pages, so a comparison across pages bundles rank, depth and authority together. Our test compares pages for the same query with rank held fixed, and the effect goes away.
Ahrefs’ March 2026 analysis found only about 38% of AI Overview citations come from the top ten, and our sample agrees: 54% of cited pages sat outside the head-query top ten. The tempting story is that fan-out explains those outsiders. Here it does not, because only 14% of them ranked top ten for any observed sub-query. AirOps’ ChatGPT work, which also logs real follow-up searches, found a third of cited top-20 pages were reached only through fan-out. That is about retrieval in a different system, and it deserves the same within-query test.
What should you do with sub-queries instead?
- Keep ranking work first. Head-query rank was the strongest predictor in the main run and in the replication.
- Treat sub-queries as content research. They are a decent list of what a thorough answer covers, which is why our engine keeps them as briefs and content gaps, labelled as research with no claim about citation.
- Write question headings because they help a reader scan, not because they move citation. Heading match did not.
- Do not build a page for every sub-query. Spinning out thin variant pages is how a site drifts towards what Google’s spam policies call scaled content abuse.
- Be suspicious of any score that ties coverage to citation, including one we might have built.
That last one is the point for me. The decision rules said a failed test meant no coverage score, so there is not one. We built the observation side (it captures and clusters the fan-out real systems run for a client’s keywords, for the audit and content planning) and stopped there. The free tool I sat down to build on 2 November is not getting built.
What can this study not tell you?
It is observational, so even a pass would have shown association, not cause. Gemini’s grounded search is a stand-in for what AI Overviews run internally, not the same pipeline. 235 queries rules out large effects, not small ones. Every page outside the top ten is in the candidate set only because it was cited, so the models really compare top-ten pages (a refit on ranked pages alone gives the same answer). And it is one point in time, autumn 2026, with systems that change month to month. I would rerun it in six months before relaxing the rule.
Where can I read the paper?
The full write-up, with the methods, all eleven deviations from the plan and the robustness runs, is Answering the fan-out does not predict AI Overview citation: a pre-registered test across 235 queries. You can read it in the browser or download the PDF. It is deposited on Zenodo as doi.org/10.5281/zenodo.23931637 and listed on my ORCID record, with the code, analysis outputs and query sample so anyone can rerun the numbers. If you have a dataset that could break this result, I would genuinely like to see it.