GPT-5 Query Fanout: What Changed and What It Broke

Ask OpenAI's ChatGPT to name the top companies in your category. If you held that position three months ago and you're gone now, something changed in April. Not your content. How GPT-5 reads it.

GPT-5 introduced a retrieval behavior called query fanout. Before generating an answer, it breaks your question into smaller sub-questions and runs each one separately. GPT-4 ran two to three sub-queries per search. GPT-5 runs five to twelve. The retrieval pool got wider. Most B2B companies weren't covering it.

Companies that held citation positions in GPT-4 lost them in April for three specific reasons.

First: single-page dominance stopped working. One well-optimized page used to be enough to own a category term. GPT-5 pulls from five to eight sources per answer. One page gets diluted across the wider pull.

Second: inconsistent language across pages broke clustering. If your site uses ten different words for the same thing, GPT-4 often consolidated them. GPT-5 treats each variation as a separate topic. You just split your own authority into fragments. This is what I call Semantic Drift. Using varied language across your content until AI reads you as multiple different companies instead of one.

Third: companies without a Knowledge Graph or Wikidata presence got skipped at the synthesis step. GPT-5 retrieved them in sub-queries. Then passed over them when assembling the final answer. Retrieved but not cited.

The companies still holding positions in GPT-5 aren't producing more content. They have consistent entity signals across every surface AI checks. That's the difference.

The full brief covers the exact mechanics of the fanout shift, the three content structures GPT-5 is now skipping, how to test your current citation position in under ten minutes, and the structural adjustments that restore it.