Before ChatGPT fetches a single web page, it composes a query on its own, and that self-written query already contains brand names nobody typed. Search consultant Suganthan Mohanadasan tested this on one ChatGPT Plus account and published the results as a contributor column in Search Engine Journal, using 27 category questions for the pre-search query test and a separate set of 57 conversations for a citation analysis. Brand names showed up in the model’s own query, before any result had been retrieved, in 21 of those 27 tests.
Ask ChatGPT which AI note-taking app is best without naming one, and the query it silently generates already lists seven products: Otter, Mem, Fathom, Granola, Notion AI, Fireflies and Limitless, according to the source. The same injection turned up for language-learning apps, accounting software, online therapy and robot vacuums, in one case down to a specific named model. Electric SUVs broke the pattern. There, the self-written query named review outlets such as Car and Driver rather than manufacturers, which suggests the model’s stored association can point at reviewers instead of vendors depending on the category.
Mohanadasan then split every brand his data touched into two groups: those named in ChatGPT’s own query and those that surfaced only after retrieval. Brands named in the query were mentioned in the final answer 68.9 percent of the time, across 119 instances. Brands fetched during the search but never named in the query were mentioned just 2.1 percent of the time, across 515 instances. That is roughly a 33 times gap between the two groups.
That gap does not prove the pre-search naming causes the citation that follows it. A brand’s presence in ChatGPT’s self-written query and its odds of being cited later could both trace back to the same underlying prominence in the model’s training data, rather than one driving the other. Mohanadasan flags a related limitation himself: his aggregate percentages count a brand learned in one conversation turn and queried in a later turn as pre-known, which is why he treats the cleaner first-query result, 21 of 27, as the stronger claim.
A second filter operates on whatever clears that first bar. Working from a labeled set of 57 conversations and 3,554 retrieved pages, Mohanadasan found that only 110 pages, 3.1 percent, earned a citation. Position within a domain’s result group predicted most of the outcome: a first-position result converted at 5.2 percent, second at 4.6 percent, third at 2.4 percent, and sixth or later at 0.3 percent. Below the second slot, citation odds are close to a rounding error.
The percentages throughout are directional, not measured. They come from one ChatGPT Plus account logged in from Dubai over two days in late July 2026, weighted toward software and AI-tool queries because that is what Mohanadasan asks about, and he says a different mix of questions would move the numbers. The mechanism itself, brand names appearing in a pre-search query before any page is fetched, is independently verifiable in about two minutes through Chrome DevTools, according to the source.
For a search team, the finding splits AI-visibility work into two budgets instead of one. A brand absent from ChatGPT’s own query needs digital PR, comparison coverage and review placements that build the training-data association no crawl can create. A brand already appearing in the query should shift effort toward the individual pages losing out below the second result position, where citation odds fall under 2.5 percent.
Suganthan Mohanadasan reported these findings in a contributor column published by Search Engine Journal on August 14, 2026.
