SEO & AI Tip: A single Query Fan-Out misses most of what Google and ChatGPT are actually searching. Fan-outs are non-deterministic. One run is one sample from a much larger distribution. Running a query fan-out once shows you a sample, not the space. Most SEOs assume one fan-out captures what Google AI Overviews and ChatGPT are searching on a user's behalf. It doesn't, fan-outs are non-deterministic, and one run is just one draw from a much larger distribution. We regularly run the same prompt through Google over and over again and watch how the generated fan-out queries change on each request. We started capturing every unique sub-query until new runs stopped producing anything we hadn't already seen. Then we cluster the full set to see what topical directions the models were actually exploring. The screenshot below shows the result. What stands out: Single runs were thin. Any one fan-out only surfaced a slice of the sub-queries the model would generate for that seed. Saturation took repetition. It took multiple runs before new unique fan-outs stopped appearing, that's the signal that topical coverage for the space has been captured. The modifier is the signal. Every fan-out shares the broad topic. The differentiators: reviews, by year, cost, near me, etc are the topical directions you actually need to cover. We found that clustering works better if you mean-center the embeddings first. Otherwise everything collapses into one mega-cluster because the main topic moves each fan-out into a similar vector space. Why this matters - Map fan-outs from a single run and you're optimizing against a sample, not the topical space. - Once you cluster to saturation, you see exactly which modifier directions the models care about. The starting point to determine whether your site covers each one. What to do next - Pick your highest-intent seed queries or prompts. - Run the fan-out repeatedly until no new sub-queries appear. That's your full topical space. - Embed each unique query fan-out, mean-center the vectors of each fan-out to help isolate what's unique about each one, and cluster the unique embeddings with silhouette-picked k. - Map each cluster to existing pages. Empty clusters are content gaps. Overcrowded ones are consolidation opportunities. Fan-out saturation plus clustering is how you find the topical directions your content may not cover that are important to an LLM when generating a response. Weighting those that drive revenue as top priority and after filling the topical content gaps you'll likely increase how often your company is recommended by AI Overviews and ChatGPT.… more