Martin Alexandru

AI Search Field Data: What ChatGPT Searches Before It Answers

This is my AI search field data: numbers from tests I run myself on ChatGPT and Google, published with the date and the method. Each entry answers one question about how AI search picks what to show, and I add a new one every month.

Published

Only my own measurements go here. When I quote someone else's study, I name it.

Last updated

Latest entry
17 October 2026: 88 ChatGPT runs, 204 fan-out searches
Next entry
November 2026: the same questions again
Entries so far
2

October 2026: What ChatGPT Searches For When a Buyer Asks

On 17 October 2026 I asked ChatGPT 58 questions a buyer might ask before hiring someone for AI search: who to hire, what it costs, consultant or agency, which case studies to trust. Web search was on for every run. The five questions closest to a purchase ran three times, the rest once or twice. That's 88 runs.

Before ChatGPT answers, it rewrites the question into one or more web searches. These are called fan-out queries. I saved all 204 of them.

What I measuredResult
Fan-out searches per answermedian 2, maximum 6, never zero
Length of a fan-out searchmedian 11 words
Fan-outs that contain a year46% (mostly 2026)
Search results returned per answermedian 21
Fan-outs aimed at platform documentation (OpenAI, Google Search Central)43 of 204
Fan-outs that checked one named company9 of 204
Fan-outs using a site: search4 of 204

ChatGPT picks its own words. When the buyer's question didn't mention the term, ChatGPT still added "generative engine optimization" or "GEO" to 41% of its searches. It added "AI SEO" or "AI search optimization" to 8%, "AEO" to 6% and "AI visibility" to 3%. So a page that only says "AI SEO" is missing the words ChatGPT searches with most.

Three directions carry the buying questions. Weighted by how close each question is to a purchase, the searches that mattered most were consultant plus case studies, consultant plus a "best" list, and consultant or agency plus pricing. ChatGPT looks for proof, for someone else's ranking, and for a price.

Who got named. First Page Sage was named in 14 of the 88 runs and iPullRank in 12. Independent consultants do get named: Aleyda Solís, Kevin Indig and Lily Ray each came up several times. I was named in none of them, and none of my pages came back as a result. That's the baseline.

October 2026: OpenAI's Search Crawler Read My robots.txt 48 Times and No Pages

OAI-SearchBot is the crawler that fills ChatGPT's search index. In the 14 days to 17 October 2026 it requested my robots.txt 48 times and not one page. I checked the visits against the IP addresses OpenAI publishes, so these were real. Nothing on my side blocked it: the robots.txt allows it and the firewall's AI-bot blocking is off.

My site has a sitemap, and that wasn't enough. My reading is that the crawler finds pages through links from pages it already knows. So the fix isn't technical: the site needs links from places OpenAI already crawls. I'll report here when the first page gets fetched.

What This Means If You Want ChatGPT to Recommend You

  1. Use the words ChatGPT searches with. Put "generative engine optimization (GEO)" in the title and first sentence of the pages that matter, next to the term your buyers type into Google.
  2. Publish proof with numbers. Case studies were the strongest buying direction.
  3. Publish a price. Pricing searches came up again and again, and pages with real numbers get cited.
  4. Get onto the lists. "Best consultant" answers are built from other people's rankings, not from your homepage.
  5. Check that the crawler can reach you. Look at your server or CDN logs for OAI-SearchBot. If it only reads robots.txt, you need links, not settings.

Method

ChatGPT runs go through DataForSEO's ChatGPT endpoint with web search forced on, so every answer searches the web. Most runs were from the United States and the United Kingdom, with single runs from Singapore, the UAE and Australia. Each question has a weight from 1 to 5 by how close it is to a purchase. Fan-outs are tagged with fixed word lists, not by an AI, so every number traces back to a saved search. The raw answers are kept with their date.

One month of data from one engine is a snapshot. The point of this page is the series: the same questions, every month.

What is a fan-out query?

A fan-out query is a web search that ChatGPT (or Google's AI Mode) writes itself before it answers. It turns your question into one or more searches, reads the results, and builds the answer from them. In my October 2026 test, ChatGPT ran a median of 2 fan-out searches per answer and at most 6.

Does ChatGPT use the same words as the person asking?

Not always. In my October 2026 test, ChatGPT added "generative engine optimization" or "GEO" to 41% of its searches even when the question never used the term.

Get the Next Entry

Get each new entry by email when it's published. No other emails.

Want the same test run on your company? Start with an AI visibility audit. My own progress from zero is logged in the live experiment.