- The fact: since switching to GPT-5.3 on March 4, 2026, ChatGPT Search cites an average of 15 unique domains per response versus 19 before — a 20% drop.
- The mechanism: fan-out queries let ChatGPT fire 5–10 parallel sub-searches per response, but only toward sources already known in parametric memory.
- The Bigfoot Effect: AI visibility concentrates on a shrinking pool of authority domains. If your brand isn't in LLM memory, it isn't even a candidate.
- What to do: build parametric visibility (press, Wikipedia, authority sites) before the next training data update window.
Independent research published May 14, 2026 on Search Engine Land reveals how ChatGPT's internal search engine actually works — and the numbers will surprise any marketing team that assumed their site was "visible" on AI.
The research team reverse-engineered ChatGPT Search's web.run system — the fundamental architecture powering all web retrieval during conversations — and documented how fan-out queries redistribute visibility across the web.
Fan-out queries: how ChatGPT decides what to search
When you ask ChatGPT Search a question, the model doesn't fire one search. It executes a fan-out strategy: multiple targeted queries in parallel, each designed to retrieve a specific angle from specific sources.
For a typical query, ChatGPT generates between 5 and 10 sub-queries. For product queries, the team discovered a previously undocumented type: browse_rewritten_queries. When a user asks for a product recommendation, the model first runs a query to identify candidates, then individual queries for each product to retrieve specs, reviews, and pricing.
GPT-5.4 Thinking goes further: it uses site: operators to restrict searches to domains it considers trusted. If your domain isn't on that implicit list, it simply won't be queried.
Key finding from the research: "The model formulates its web queries by targeting sources it already knows. A brand absent from parametric memory won't even be considered as a search candidate." — Source: Search Engine Land, web.run system analysis, May 14, 2026.
The Bigfoot Effect: the concentration of AI visibility
The team named it the "Bigfoot Effect": a growing concentration of citations on a limited pool of authority domains, mirroring the Google Bigfoot update of 2012.
The numbers are precise:
- Before GPT-5.3 (before March 4, 2026): 19 unique domains cited on average per response, 24 URLs
- After GPT-5.3: 15 unique domains, 19 URLs — a -20% reach drop
- Consequence: some pages have simply stopped being crawled by the ChatGPT-User bot
The effect is amplified by GPT-5.4 Thinking which, per the research, "amplifies concentration further" through its targeted site: operators.
For teams working on their GEO strategy in 2026, this confirms what we've observed for months: AI visibility is not linear. It follows a power law — the early movers capture the majority.
Parametric vs. dynamic visibility: the key distinction
The research introduces a critical framework for understanding why some brands get cited and others don't:
Parametric visibility
This is the authority encoded in the model's training data — stable, measurable through one-shot API audits. It's shaped by presence in Wikipedia, press coverage, and high-authority site citations in your sector. Think of it as E-E-A-T for LLMs.
Dynamic visibility
This is real-time retrieval during conversations — volatile and model-dependent. The same prompt produces different citations depending on whether the user accesses GPT-5.3 Instant, 5.4 Thinking, or 5.4 Extended. A single model update can wipe out your dynamic visibility overnight.
The strategic conclusion is clear: parametric visibility is the only stable foundation. And training data update windows — roughly annual — represent the strategic opportunity to influence that LLM memory. We may be in one of those windows right now.
The web.run system: ChatGPT Search's internal plumbing
Beyond the Bigfoot Effect, the research documents the architectural evolution of web.run, the system driving all ChatGPT web searches.
Before March 4, 2026, web.run sent compact text commands separated by pipes. Since GPT-5.3, it transmits structured JSON objects with typed parameters — a significant evolution reflecting deep changes in query formulation.
The system now supports 12 operations (up from 4): search_query, open, find, click, screenshot, product_query, plus specialized widgets for sports, finance, and weather. This expansion signals that ChatGPT Search is becoming a real retrieval engine, not just a web search overlay.
The team also uncovered internal policy details: Reddit receives an exemption from copyright word limits; a banned-products list exists; verbosity operates on a 1–10 scale. These details were obtained by circumventing filters through formulations that avoided the term "system prompt."
What this means for your AI visibility strategy
If you're a marketing director, SMB CEO, or SEO consultant, here are three concrete implications:
- Monitor your visibility across multiple GPT versions. The same prompt produces different citations depending on the model. Your GEO audit must cover GPT-5.3, 5.4 Thinking, and 5.4 Extended separately.
- Invest in parametric visibility now. Training data update windows are the strategic opportunity to durably influence your LLM memory footprint. Press publications, Wikipedia, authority sites — these signals accumulate.
- Structure your content for dynamic retrieval. Direct answers at the top of articles, dated data, named sources, self-contained citable passages — the ingredients that make a page retrievable by the fan-out. This is what Cicero calls GEO-ready content.
The key number to retain: if your brand isn't in parametric memory, it won't even be in the candidate pool for citation. Fan-out doesn't discover unknown brands — it retrieves, it doesn't explore.
Limits of this analysis
What this article does not cover:
- The precise reverse-engineering methodology used by the research team (not fully disclosed)
- Comparative data for Perplexity, Google AI Overviews, or Bing Copilot — behavior may differ significantly
- Impact on paid vs. free GPT plans — the research notes free-plan users may be routed to lighter models
- Non-English language sites — data is primarily from English-language sources
Further reading
- OpenAI ChatGPT Ads Manager and GEO in 2026
- Google AI Overviews: 38% click-through impact
- OpenAI GPT Realtime Voice and SEO content
Frequently asked questions
The Bigfoot Effect describes the growing concentration of ChatGPT Search citations on a limited number of high-authority domains. Since switching to GPT-5.3 in March 2026, the model cites an average of 15 unique domains per response versus 19 previously — a 20% drop. Brands absent from the model's parametric memory are no longer considered as citation candidates at all.
Fan-out queries are ChatGPT's strategy of launching multiple targeted searches rather than a single broad query. For one user question, the model generates 5 to 10 parallel sub-queries, each designed to retrieve specific information from specific sources. GPT-5.4 Thinking goes further by using site: operators to restrict results to domains it considers trusted.
Parametric visibility is the authority encoded in the model's training data — stable and measurable through one-shot API audits. It depends on your brand's presence in Wikipedia, press coverage, and high-authority sites. Dynamic visibility is real-time retrieval during conversations — volatile and model-dependent. A single model update can wipe out your dynamic visibility overnight.
Three levers: (1) build parametric visibility through Wikipedia mentions, specialist press, and authoritative sites in your sector; (2) structure your content for dynamic retrieval with direct answers, dated data, and named sources; (3) monitor your visibility across different GPT versions as one model may cite you while another doesn't. Training data update windows are the strategic opportunity to influence parametric memory.
web.run is ChatGPT's internal search system — the fundamental architecture powering all web retrieval during conversations. Since GPT-5.3 on March 4, 2026, it transmits structured JSON objects with typed parameters. It now supports 12 operations: search_query, open, find, click, screenshot, product_query, and specialized widgets for sports, finance, and weather.
Growth and SEO content strategist, I founded Cicéro to help businesses build lasting organic visibility — on Google and in AI-generated answers alike. Every piece of content we produce is designed to convert, not just to exist.
LinkedIn