MJ Cachon found 189 branded ChatGPT prompts produced 1,797 sub-queries

Search Engine Journal published Greg Jarboe on 2026-09-11 reporting on a dataset study by MJ Cachon of how ChatGPT fans a single branded prompt out into many sub-queries. According to the SEJ article, Cachon ran 189 branded prompts through ChatGPT in August and observed 1,797 sub-queries the model issued in response, an average of 9.5 per prompt. Jarboe writes that within a single fan-out run, ChatGPT's first sub-query is plain conversational language and later ones narrow with the site: operator and take exact quoted phrases, with quote usage rising 25-fold from the first search of a run to the last.

The pattern the study describes

The SEJ piece describes each fan-out run as a narrowing sequence. The first sub-query is broad. The next few narrow to a specific site. The last few take an exact quoted phrase from a page and search for it again as an exact string, apparently to verify that a source contains the wording the model thinks it does. Cachon's dataset puts a number on how much heavier that verifying use of quoted phrases is at the end of the run than at the beginning, and Jarboe treats the 25-fold ratio as the main finding.

The other numbers Jarboe cites

Jarboe sets the fan-out figure alongside two other numbers from named sources. The first is Google's 2019 statement, restated in March 2025 by John Mueller at Search Central Live NYC, that 15 percent of Google's daily queries have never been seen before, a share that has stayed near constant even as large language models spread. The second is a May 2026 Google usage number that Jarboe himself reported earlier: the average AI Mode query in the United States is now three times the length of a traditional search query. Cachon's dataset adds that a single branded prompt in her run fanned out into sub-queries that averaged seven words each.

What Jarboe says to do about it

The article prescribes three habits for a content team, and each one is a statement about page structure a search team can apply the same day. Build a page around the nested version of a target phrase, so the copy contains both a three-word core and its four-word or five-word extension, because AI systems narrow their queries from short to long inside a single run. Publish as fast as the news happens, because the 15 percent of Google queries that have never been seen before are disproportionately tied to something that just happened, and a same-day publishing schedule is where new phrasing appears first. Write the literal answer as a standalone quotable sentence, because the last sub-queries in a fan-out run search for exact phrases and take them word for word.

For a marketing team measuring generative-search visibility, this is a research finding more than a product. The commercial fan-out trackers already sell dashboards on ChatGPT sub-queries. Cachon's contribution is that the shape of a run is repeatable and quantifiable across a real dataset. A team that has been tuning pages for one branded prompt can now decide, using her ratio, whether the sub-queries the model actually issues on that prompt are ones its own copy contains at all.

Source

Reporting: Greg Jarboe, Search Engine Journal, 2026-09-11. Underlying study: MJ Cachon, "Study query fan-out ChatGPT brand", August 2026.