We sell AI visibility work, so we ran our own audit on ourselves first. The result was a useful mistake: our pages were written for the words a buyer types, and ChatGPT searched with different words.
What a fan-out query is
When you ask ChatGPT a question with web search on, it does not search for your question as typed. It writes its own search queries, reads the results, and builds the answer from the pages it picked. Those background searches are the fan-out queries. If your page does not match the words in them, it may never be read, however well it answers the original question.
How we ran it
- 10 buyer questions, written the way a buyer might ask for AI visibility help: four "best for" questions for the UAE, India, Southeast Asia and the US, one on experts, three "how do I" problems, one on audit cost and one comparison (GEO agency or fractional CMO).
- ChatGPT with web search on, in a temporary chat, in English, on 22 October 2026 between 23:30 and 23:49 IST. Our brand was never in a question.
- For each answer we logged the searches ChatGPT ran, the brands and people it named, and every URL it cited.
The mistake we found on our own site
RemoteCMO was named in 0 of the 10 answers, and none of the 72 cited URLs was on remotecmo.io. Our domain did not appear in the search results either.
The reason was partly vocabulary. The 4 pages we checked said "AI visibility" throughout. They said "GEO" zero times and "fractional" zero times. ChatGPT, meanwhile, searched with both. On 23 October we changed our titles and headings to use the words ChatGPT searched with, where they are true for us.
1. GEO is the word ChatGPT searches with
Behind the 10 answers ChatGPT ran 26 searches. 15 of them used "GEO" or "generative engine optimization". Only 2 of the 10 questions used the term, so ChatGPT added it to 11 searches on its own, across 6 questions that never said it.
One example. The question was: Who are the best AI visibility consultants for brands in Dubai and the UAE? The first search ChatGPT ran was: best AI visibility consultants agencies Dubai UAE GEO generative engine optimization brands.
2. It adds "agency" and the year, too
12 of the 26 searches said "agency" or "agencies". In 7 of those, from 5 questions, the buyer had not said agency. 9 searches added "2026", and 18 of the 72 cited URLs had 2026 in the page title. A buyer asking for a consultant can still be shown pages about agencies, and dated pages for the current year.
3. Lists decide the market shortlists
15 of the 72 cited URLs were "best" or "top" lists. For the best-agency questions about the UAE, India and Southeast Asia, 9 of the 34 sources cited were lists of agencies, and three of the agencies ChatGPT named had their own list cited. 36 of the 72 cited URLs were on the named brands' own sites.
No agency or consultant was named in more than one answer. Each market had its own set of names, drawn from its own lists, so each market has to be won on its own.
4. Some questions get no vendor at all
In 3 of the 10 answers, ChatGPT named no agency or consultant at all: two of the "how do I" questions and the GEO agency or fractional CMO comparison. 13 of the 14 citations in those answers went to OpenAI, Google Search Central or arXiv. Across the audit, 6 of the 26 searches went looking for official documentation from OpenAI, Google Search Central or tool vendors. Those slots go to the platforms' own pages, and a brand page will not win them.
5. "Fractional CMO" is a search term in its own right
5 of the 26 searches used "fractional CMO", from the comparison question and the US question. If you offer senior marketing leadership on a part-time basis, the page that says so in plain words is the one ChatGPT can find.
What to do with this
- Log the searches, not only the answers. The answer tells you who won. The searches tell you which words decided it.
- Use their words where they are true. Put them in your page titles, headings and first sentences. Do not claim a term that does not describe what you do.
- Find the lists for each market. List the "best" and "top" pages cited for your category in each country, and check whether you are on them and described correctly.
- Date your pages. Show when a page was published and updated, and update it only when something real changes.
- Skip the slots you cannot win. Questions answered from official documentation are not worth a brand page.
The afternoon AI visibility audit shows a simple version of the method, and the AI visibility audit page shows how we run it for a brand, by market.
Limits
This is one run of 10 questions, in English, on one day, from one signed-in temporary chat. Location was not recorded. ChatGPT's searches and answers vary from run to run, so treat these as early signals, not rates. We classified "lists" from the cited page titles, which is a judgement call. We will repeat the run and publish the second result here, with its date.