Original post by Suganthan Mohanadasan
30-second rundown
Key learnings:
- ChatGPT may arrive with a shortlist: In one account, 21 of 27 first searches contained brands the user had never mentioned, showing that the model often recalls candidates before it fetches any webpage.
- Brand recognition opens the first gate: Brands named in ChatGPT’s own queries appeared in the final answer 68.9% of the time, compared with 2.1% for brands that were merely found during retrieval, although these directional figures come from one personalized account.
- A focused page must win the second gate: Only 3.1% of retrieved pages earned a citation, so consolidate overlapping pages and put the clearest answer, evidence, and numbers near the top in ordinary webpage text.
This week: Ask ChatGPT the same buyer recommendation question five times and record which brands and buying criteria appear in its searches. Identify the most repeated criterion, then update your strongest relevant page so its answer and supporting evidence appear clearly near the top. Remove or redirect one weaker page that competes for the same question, leaving the machine one obvious source to cite.
Coming 3 months: Build an automated AI Search monitoring workflow that runs your priority buyer questions regularly and records which brands appear in the initial shortlist, which pages are retrieved, and which sources earn citations. Route persistent brand gaps into an outreach workflow for reviews, comparisons, and expert coverage, while sending citation gaps into a content workflow that strengthens or consolidates the relevant pages. Assign one owner to review changes and act on recurring patterns each month.
Suganthan Mohanadasan went digging through the network traffic of his own ChatGPT account and found a velvet rope where the open web was supposed to be. Before ChatGPT fetches a page, its first search query often contains a brand shortlist drawn from what the model already associates with the category.
1. What Mohanadasan typed: best AI note taking app. He did not mention Granola, Notion AI, Otter, Fireflies, Fathom, Mem, Limitless, or any other product.
2. ChatGPT’s first broad search: best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless. ChatGPT added the year, the buying criteria, and seven brand names itself. Because this was the first search query, those names could not have come from pages retrieved earlier in that conversation.
3. ChatGPT’s follow-up shortlist searches: it then fanned out into nine narrower checks, usually sending a separate site: query to a shortlisted company’s own domain. Mohanadasan showed examples including site:granola.ai pricing features AI meeting notes 2026, site:otter.ai pricing AI meeting notes 2026, site:fathom.video pricing AI meeting assistant 2026, site:notion.com product AI Meeting Notes official 2026 pricing, site:fireflies.ai pricing official AI meeting notes 2026, site:mem.ai pricing AI notes official 2026, and site:notebooklm.google official features pricing 2026.
The first search created a working guest list; the later searches investigated names already on it. The web search that follows is real and can still change the result, but some companies arrive at the casino with their names already pencilled onto the guest list.
1. The Search Starts With a Memory Test
ChatGPT rewrites a user’s question into its own searches, reads the results, and then assembles an answer. Those searches are visible inside the response downloaded by the browser under a field currently called “search_queries,” so this is not a stolen system prompt or a moonlit séance with the machine.
Across the controlled first-query comparison, 21 of 27 conversations contained brands the user had never typed; among recommendation prompts where he supplied no candidates, it happened in 10 of 11 searches. The model was not discovering the entire field from scratch. It was reaching into memory and bringing its own suspects.
The list is not fixed. Similar wording produced different brands, and repeat tests sometimes switched from vendor websites to review publications, so one answer is a lousy oracle. Mohanadasan also revised his original headline after further tests showed that brands absent from the initial query can still reach the final recommendation, which means the shortlist is an advantage, not a sealed verdict from the desert courthouse.
2. A Name in the Query Gets Thirty-Three Times More Oxygen
Mohanadasan divided brands into two groups: those ChatGPT named in a query it wrote, and those it only encountered after fetching search results. Among 119 pre-named brands, 68.9% appeared in the final answer; among 515 retrieved-only brands, just 2.1% survived. That is roughly a thirty-three-fold gap, large enough to make every technical-audit salesman suddenly develop an interest in the carpet.
He also found 86 recommendations where the brand’s own website was never fetched during that conversation. Being known as part of the category can matter far more than merely having a crawlable page, because the model may nominate a brand before its server is ever contacted.
This is where the first game is won: reviews, comparisons, digital public relations, analyst coverage, expert discussion, and category-defining content across the open web. Schema markup, page speed, and an llms.txt file can help machines interpret a site later, but they cannot retroactively stuff an unknown brand into a query that was written before the website entered the room.
3. The Second Gate Is a Citation Meat Grinder
Recognition only buys admission to the next bad neighborhood. Mohanadasan labelled 3,554 retrieved pages across 57 conversations and found that 110 earned citations, a rate of 3.1%. Almost every page was ignored, so appearing in the search results is not the same as becoming evidence in the answer.
Position inside each domain’s group mattered sharply. The first page was cited 5.2% of the time and the second 4.6%, while pages in sixth place or later managed 0.3%; two tightly matched pages from one domain performed best at 6.2% per page, while six or more fell to 1.7%. Flooding the machine with near-duplicates is not coverage. It is a family knife fight in a locked filing cabinet.
The cited page usually ranked in the top 5% for relevance to the supported claim, yet it was the single closest textual match only 20% of the time. Relevance gets a page considered but does not fully explain the final choice. The practical move is still mercifully concrete: keep one strong page for each buyer question, fold or redirect competing pages into it, and place the plain-English answer with its evidence near the top in visible Hypertext Markup Language, the ordinary text code a browser and an AI system can both read.
The machine is playing two games in sequence. First it asks whether your brand belongs in the category at all; then it asks whether one of your pages deserves to be quoted.
If you are absent, earn category memory beyond your own website; if you are present, give the machine one clear page worth citing. Keep checking monthly, because Mohanadasan watched the search format and the size of the query fan-out change within weeks, and anything this alive will rearrange the furniture while you are still measuring the room.
