Reverse EngineeringFanout Queries
ExposedHow Important
Web Exposure Is

The Fanout Query

When a person types a prompt into ChatGPT, the model answers in one of two ways. It can answer from what it already learned in training. Or it calls a web search, looks up the latest information, and answers from what it finds.

ChatGPT takes the prompt and writes fanout queries from it. A fanout query is a web search the model generates from that prompt. Those fanout queries are sent to Google’s index, Bing’s index, and OpenAI’s own index. The pages that come back are what the answer is written from.

This study looks only at the start of that pipeline, the fanout query. The indexes, the pages, and the written answer all come after it.

Live fanout queries have been picking up brand names. For a prompt like “best email marketing software”, the live fanout query was “best email marketing software 2026 Mailchimp Kit beehiiv HubSpot official pricing features”.

The question is which brands get added into fanout queries, and why ChatGPT leaves out brands it clearly knows.

Search-Off Recall

AWeber is missing from the live fanout query. On 50 email prompts, the live fanout query names AWeber 0 times. Before calling that a gap in what ChatGPT knows, we checked whether the model already knows the brand.

That check is search-off recall. Web search is off, so the model cannot look the brand up. We ask which brands come to mind for the category, and we read the answer for the name. If the name is there, the model knows the brand. This is model probing. It is a separate test from the fanout query.

On the next prompt the order was swapped: the category line first, then the same instruction. We counted a brand when its name showed up anywhere in the answer.

The model does know these brands. It still leaves them out of the live fanout query. The same probe in other categories: Hasselblad on 44 of 50 camera prompts, Basecamp on 45 of 50 project-management prompts, ClickUp on all 50. Kodak was on 0 of 50 camera prompts, so the probe also shows when a brand is not connected with the category.

The Fanout Query Simulator

We built a fanout query simulator, powered by GPT-5.6 Luna via the API. It receives the same prompt, with web search turned off, and one instruction written beforehand: write the web search ChatGPT would write.

We took 919 live fanout queries extracted from the API, each with the prompt it came from. For every prompt, the simulator wrote a fanout query, and we compared that simulated fanout query with the live one.

The exact wording matched on 0 of 919. Out of 919 live fanout queries, 782 contained at least one brand name, so the brand comparison is 782 live fanout queries against 782 simulated fanout queries. On 100 of 782 prompts, both contained the exact same set of brands. Both shared at least one brand on 722 of 782. Across the fixed brand panel, how often each brand shows up lines up at 0.965.

The simulator agrees with the live system on which brands get picked over many prompts. It does not reproduce each query word for word, and it does not keep the same brand order. That agreement is what makes the simulator worth trusting for the question the live system will not answer.

“Brands in both” is how many brands in the live fanout query also appear in the simulated fanout query. On the VPN prompt, the live fanout query names review sites and no VPN brand, so the overlap is none.

Increasing the Budget

AWeber is associated with “best email marketing software”. If the model knows the brand and connects it with that prompt, why didn’t the live fanout query include it?

The live fanout system will not let us change how many brands the live fanout query may name. In the simulator, we can. We raised the budget: the number of brands the simulated fanout query is allowed to name. The simulator had to name exactly 1, 2, 3, 5, 8, or 12 brands. We recorded when each brand first appeared.

The brands missing from the live fanout query came back once the budget was high enough. AWeber is absent from 50 of 50 live fanout queries, appears in 2 of 50 simulated queries at a budget of 8, and in 47 of 50 at 12. Campaign Monitor is present in 1 of 50 live fanout queries, and its entry point is 8. The live fanout query writer uses a short list. Brands below the cutoff are left out, but they appear when the simulated fanout query is allowed to name more brands.

We could raise the budget only in the simulator, not in the live system. This is an inference from the close match between the live and simulated fanout queries, not a setting we changed in the live system.

Common Crawl Correlation

So far, two things are clear. The brands in the fanout query are a shortlist. The model may know a brand well and still leave it out of the live fanout query, because of that cutoff.

The next question is what made Mailchimp and the other fanout-query brands win, while AWeber and Campaign Monitor did not. We compared the brands with the Common Crawl January 2026 index. Common Crawl is used by OpenAI, as stated in court documents and in the GPT-3 paper. For each brand we counted websites where the name appears near words about its product category, and ranked brands only against rivals in that category.

It matches very strongly. Category web presence and the simulated fanout queries move together.

Across all 29 brands, the within-category rankings match at +0.76. Restricting the comparison to the 25 brands with a measured simulator budget gives +0.85. In a held-out e-signature category, fixed before looking at the results, the match is +0.986. DocuSign is first by web presence and first in the simulated fanout query.

Cleaning the pages does not make the match stronger: +0.762 on raw pages, and +0.730 after a quality filter and near-duplicate removal. US branded search volume also lines up on the software brands, so the crawl result does not separate cleanly from popularity.

Summary

1. 1.The simulator is a good proxy for the live fanout query. It does not copy the exact wording, but it picks mostly the same brands. On 722 of the 782 prompts where the live fanout query named a brand, the simulated fanout query shared at least one of those brands. Across the brand panel, how often each brand appears matches the live fanout query at 0.965.
2. 2.Brands missing from the live fanout query come back when the simulator may name more brands. What matters is the budget at which a brand enters. Mailchimp, Salesforce, Zendesk, 1Password, Asana, and Canon enter at 1 name. Campaign Monitor and Basecamp enter at 8. AWeber and Hasselblad enter at 12. Those missing brands were not unknown. They were further down a list that the live fanout query cuts short.
3. 3.Common Crawl presence tracks where a brand sits on that list. Within each category, brands with more category-related websites enter the simulator at a smaller budget. The match is 0.76 across 29 brands, 0.85 for the 25 brands that enter the simulator, and 0.986 on electronic signature, a category set aside before any of these numbers were looked at.