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

What changed

In April 2026, OpenAI's GPT-5 changed the number of sub-queries it runs before generating an answer. This process is called query fanout. GPT-4 ran two to three sub-queries per search. GPT-5 runs five to twelve.

What that means: before GPT-5 writes a single answer, it runs up to twelve separate searches in the background. It pulls sources for each one. Then it synthesizes everything into one response.

The retrieval pool got four to six times wider overnight. Most B2B content libraries weren't built to cover that surface.

The change rolled out across OpenAI's ChatGPT between April 3 and April 9. Perplexity made a parallel shift in its query decomposition logic in the same window, moving from single-pass retrieval to a two-stage model for navigational and comparative queries. Google AI Overviews extended its entity corroboration threshold — meaning it now requires more external sources before citing a company it doesn't already have strong Knowledge Graph signals for.

All three platforms moved in the same direction in April: wider retrieval, higher corroboration bar, less tolerance for single-surface authority.

What it broke

Three visibility patterns that worked in Q1 stopped working in April.

Single-page dominance. One well-optimized page used to be enough to hold a category term in ChatGPT. GPT-5 now pulls from five to eight sources per answer. A single page gets diluted across the wider pull. Companies that built deep on one URL and thin everywhere else lost citation positions even when their core page didn't change.

Semantic inconsistency across pages. GPT-4 often consolidated varied language into one entity. Your site might say "AI marketing automation" on one page and "intelligent campaign management" on another — GPT-4 read them as the same company. GPT-5 treats each variation as a separate retrieval signal. You split your own authority. This is what I call Semantic Drift. It's a Layer 3 failure in the Algorithmic Authority Stack. The fix is not better writing. It is consistent language across every surface AI checks.

Missing entity anchors. Companies without a Wikidata entry or Knowledge Graph presence got retrieved in sub-queries and then skipped at the synthesis step. GPT-5 found them. It just didn't use them. The synthesis layer prioritizes sources it can corroborate externally. No external anchor means no citation, even when the content is directly relevant.

The companies that lost ground in April share one pattern: they had strong content in one place and weak or inconsistent signals everywhere else. The wider fanout exposed that gap.

What to do

Three structural adjustments restore citability under the new retrieval model. None of them require new content. All of them require consistency.

1. Audit your language across surfaces. Pull your homepage, about page, LinkedIn, and top three blog posts. List every phrase you use to describe what you do. If you have more than three variations of the same concept, you have Semantic Drift. Pick the canonical version. Use it everywhere. Update the pages that don't match. This is a one-time fix with permanent returns.

2. Build entity depth, not page depth. GPT-5's wider fanout means it's checking more places. One 3,000-word page is less useful than five consistent 800-word pages that all reinforce the same entity signals. Map your top category terms. Make sure you have a dedicated page for each one. Make sure every page uses the same language for your company and its core expertise.

3. Create an external anchor. If your company doesn't have a Wikidata entry, create one. It takes 20 minutes. It tells every AI retrieval system that your entity has been externally verified. If you're a founder or named expert, the same applies to your personal entity. A Wikidata entry for Maria Dykstra corroborates every page on mariadykstra.com. Without it, GPT-5 treats each page as a separate unverified source.

Run the test: open ChatGPT and ask "Who are the top [your category] experts?" If you're not listed and you should be, start with the language audit. That's the fastest fix and the most common gap I find in audits.

Next issue publishes May 2026. It covers what happened to Perplexity's citation behavior when it moved to two-stage retrieval, and which content formats are now getting prioritized in the first stage versus the second.