Network traffic across 60 conversations: the model writes brand names from memory before fetching a page. Being named beats being findable by 33x.
Three names this week (Suganthan Mohanadasan’s research, Rand Fishkin’s video, Dixon Jones’s commentary) all pointed at the same thing and corrected the most expensive misunderstanding in GEO.
Suganthan read the network traffic across 60 conversations: ChatGPT writes brand names into its own first search query before it fetches a single page. Those names come from the model’s memory, not the web.
A brand that ChatGPT names in its own query gets mentioned in the answer 68.9% of the time; a brand that’s only fetched but never named, 2.1%. About 33 times the difference. So the fan-out isn’t a search for candidates, it’s the model walking down a list it already had.
Here’s what that means, plus the one test to run this week.
Suganthan Mohanadasan read the network traffic across 60 ChatGPT conversations this week and found something striking. Before ChatGPT fetches a single web page, it writes brand names into its own first search query. Ask it for the best AI note-taking app and, before any result comes back, it appends names like Granola, Notion AI, Otter, and Fireflies to its own search.
Those names can’t come from the web, because nothing has been fetched yet. They come from the model’s memory. In 21 of 27 conversations, that first query already contained brands the user never typed.
So the fan-out we discussed last week isn’t a hunt for candidates, it’s the model verifying a shortlist it already holds: name the list from memory, then run a site: probe against each brand’s own website.
And the number is brutal: a brand named in ChatGPT’s own query is mentioned in the answer 68.9% of the time; a brand only fetched, never named, 2.1%. Roughly 33x.
Which means most of what we sell as technical GEO (schema, llms.txt, speed) only improves that 2.1% column, while the real decision was made in the model’s memory before you ever saw the prompt.
The good news: these are two separate games, and you can play both, as long as you know which one you’re in.
Suganthan’s method is transparent and reproducible: he reads the search_queries field in the JSON your browser downloads to render ChatGPT’s answer. That’s where the model’s own search queries live. Because the first query is written before any result comes back, the brands inside it can only come from the model’s memory.
In 21 of 27 conversations they were already there. Across 13 unrelated categories, 11 did the same. For robot vacuums the model recalled current model numbers (Saros, Dreame X50, Eufy S1 Pro) unprompted. This “knowledge” runs all the way down to the product line.
Here’s a real example. When Suganthan asked for the best AI note-taking app, ChatGPT wrote itself this search before fetching anything:
best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless
Look at the seven products in the tail: Granola, Notion AI, Otter, Fireflies, Fathom, Mem, Limitless. The user named none of them, and no web result had come back yet. Those names came from the model.
Then it ran one site: probe per name, straight at each company’s own site:
So the fan-out was never a search for candidates. The model took a list it already had and verified each name on its own site. Miss that first list and your website never gets looked at, however well built it is.
You can see this yourself. Install a tool like Keywords Everywhere in Chrome (or Suganthan’s free FanoutFox extension); either one shows you the fan-out queries behind a search, verbatim. I did exactly that, and looked at the fan-out queries ChatGPT generated for “best project management software 2026.”
Rand Fishkin pointed at the same thing in a video this week: a large share of citations come from third-party sources (Wikipedia and Reddit in particular), and because the model already recognizes the brands, a bare “citation analysis” can mislead you.
Dixon Jones (inLinks/Waikay) put it in one line: by the time you see the prompt and measure the citation, the work is already done.
Suganthan split brands into two groups: those named in a query ChatGPT wrote, and those only fetched during the search without ever being named. Then he checked how often each group made it into the final answer:
| The brand’s situation | How many | Mentioned in the answer |
|---|---|---|
| Named in ChatGPT’s own query | 119 | 68.9% |
| Fetched, never named in a query | 515 | 2.1% |
About 33x. And in 86 cases a brand was recommended without its site ever being fetched in that conversation at all. You don’t even need to be crawled to be named; you need the model to know you.
This is the uncomfortable part for my own industry: most of what’s sold as GEO is retrieval work, the 2.1% column, while the deciding 68.9% column was settled in the training data before anyone touched a page.
Say you make the list. There’s another filter. From 57 conversations Suganthan built a labelled set of 3,554 retrieved pages: only 3.1% (110 pages) earned a citation. ChatGPT reads around 600 pages to write one answer and credits about 30. The gap between being read and being credited is where the work is.
Three things separate the cited from the ignored: position within the domain group (1st place 5.2%, 6th or later 0.3%), not piling on pages (two tightly matched pages is the sweet spot; six or more and you compete with yourself), and relevance (which qualifies you for the shortlist but doesn’t pick the winner).
In short: one tightly matched page per intent, the answer sentence near the top in plain HTML text, with the numbers in it.
Every finding lands on one split, and that split cuts GEO in two.
Game one is being in the category vocabulary. If the model doesn’t already connect your brand to your category, it won’t write you into the query, and you’re looking at a 2% chance. Schema won’t fix that, nor will speed or an llms.txt, because the decision is made before your server is ever contacted.
What builds it is slow and unglamorous: being written about, reviewed, and compared across the open web for years, until the association exists in the training data. That’s digital PR and category-defining content. It’s exactly the Off-Site GEO and the “Recognize” door we’ve covered before.
As Jan Caerels (TUI) put it in that LinkedIn thread: this isn’t retrieval, it’s the model’s foundational knowledge; changing it means waiting for a retrain, a longer game than classic SEO.
Game two is winning the citation once you’re in. That’s the 3.1%, and it’s entirely in your hands: one tight page per intent, the claim sentence up top, numbers in plain HTML, no cluster of near-identical pages fighting each other.
I saw the clearest proof of this split in our own data. We’ve been tracking our own brand across five AI engines for a while now; this is week eleven.
Last week was instructive: when I asked Gemini about our brand in a fresh chat, it searched the web (the retrieval layer) and got us exactly right, founding year and all.
But when it answered from memory alone, without searching, it didn’t know us at all, and even mistook the name for “Stradivarius” and declined to answer. That’s Suganthan’s split, live: being found and being known are two different things.
We’re strong on the retrieval side (when ChatGPT and Gemini search, they know us and even quote our measurable evidence), but on the memory side, game one, we still have a gap. That’s the work we keep flagging in the tracker: a consistent entity identity and a Wikidata record.
The model’s memory is trained mostly on English and global sources. Brands in smaller-language and emerging markets are even less likely to be written into that first query, so they fall straight into the 2% game. To make it concrete, I ran the same test in Turkish, on the Turkish version of the same category.
The bad news is the thin list. The good news is that nobody has filled it yet. In a smaller-language category, the first brand that gets written about, reviewed, and recorded in Wikidata is the first brand to enter the model’s memory. That’s not being late, it’s being early.
Let’s be honest: Suganthan’s percentages come from one account, so they’re directional, not measurements. Even the fan-out count is unstable: in early August, after OpenAI renamed a field, his captures dropped from 12 searches per answer to 4, while Nectiv’s API measurement averaged 7.61. So don’t anchor on a single number.
What’s solid is the mechanism: the model decides who to recommend from memory, before you see the search. You can verify that on your own account in two minutes.
Run one test this week, ten minutes: find out whether the model knows you by heart.
Ask ChatGPT (or Google AI Mode) the “best [your category] 2026” question your buyers ask. Then read the fan-out queries the model writes in the background.
Easiest path: install a tool like Keywords Everywhere, or Suganthan’s free FanoutFox extension; both show the fan-out queries directly, verbatim. Prefer manual? Open DevTools > Network and search the response for queries. Ask the same question five times, because the list shifts between runs.
Now the key question: is your brand in that first query?
One line: before you polish the page, measure which game you’re in, because if you’re not in the model’s memory, page-level work is polishing the handle of a door that never opens.
If you want me to map where your brand stands in both games (in the model’s memory, or merely findable) across five engines, that’s what the GEO Score Card is for. Every applicant gets a free AI visibility roadmap.
English: stradiji.com/geo-score-card
Turkish: stradiji.com/tr/geo-skor-karti
Every application strengthens the benchmark. Thank you in advance.
Subscribe to our newsletter for weekly SEO insights, join the discussion in our community, or engage with professionals on our Twitter group.
