SEO analyst Suganthan Mohanadasan captured ChatGPT's own search queries and found that in 21 of 27 conversations, the model wrote brand names the user never typed into its first query — deciding who to recommend before a single page was fetched, then searching largely to confirm.
2026-08-15 · by Moe Ameen
Around August 10, 2026, SEO analyst Suganthan Mohanadasan (co-founder of Snippet Digital) published an analysis of how ChatGPT actually picks the brands it recommends, and Search Engine Journal covered it on August 14. When ChatGPT answers a question in browsing mode, it does not search your words directly — it rewrites your prompt into its own search queries, runs them, reads what comes back, then writes an answer. Those rewritten queries sit in the response your browser downloads, under a JSON key currently named `search_queries` (previously `search_model_queries`), readable in browser DevTools. The revealing part: in 21 of 27 test conversations, that very first query already contained brand names the user never typed. The model appears to decide who it will recommend from its own prior knowledge, and then search mostly to confirm that choice rather than to discover candidates.
The citation data pointed the same way. Across 57 conversations, Mohanadasan reviewed 3,554 fetched pages and found only 110 were actually cited (about 3.1%). Brands that ChatGPT wrote into its own queries were cited roughly 68.9% of the time, while brands that were merely fetched during the search but never named in a query were cited about 2.1% of the time — close to a 33x gap. He documented 86 cases where a brand was recommended even though its website was never fetched in that conversation at all. Position mattered too: within a group of results, a page in the first slot was cited around 5.2% of the time versus about 0.3% by the sixth.
The findings are directional, not definitive. The captures came from a single account (based in Dubai, July 24–25, 2026), and OpenAI can change what it exposes at any time — related reporting noted that ChatGPT stopped surfacing some internal pipeline labels around July 21, 2026. Treat the exact `search_queries` mechanism and the specific percentages as a snapshot. The durable takeaway is structural: a large share of ChatGPT's recommendation is decided before, or independent of, the live web fetch — which changes where content discovery is actually won.
If ChatGPT largely decides who to recommend from what it already "knows," then the job is not to optimize one page — it is to become a brand the model has seen discussed, consistently, across many surfaces. That is a presence-and-cadence problem, and it is exactly what [Kompozy](/) is built to run. Kompozy generates net-new content — [Blog Articles](/glossary/output-buckets), [Carousels](/glossary/output-buckets), [Quote Graphics](/glossary/output-buckets), images, persona and avatar video, and [Email Newsletters](/glossary/output-buckets), not just repurposed clips — and publishes it across an owned blog plus eight social platforms and email on [Autopilot](/glossary/autopilot). Every piece stays in one voice through the [Persona Brief](/glossary/persona-brief) and one look through [HyperFrames](/glossary/hyperframes), so the footprint reads as a single, recognizable brand rather than scattered posts. That repeated, on-brand presence is the raw material a model's prior is built from.
The concrete near-term move is a post about the study itself: "ChatGPT decides who it recommends before it searches — here's what that changes." Feed the facts into Kompozy and it becomes a blog explainer, a carousel of the indexed-versus-recommended distinction, quote cards of the sharpest numbers, and a batch of captioned shorts, scheduled everywhere in a day. Then keep going — publishing specific, useful, answer-shaped content on a real cadence is how a brand earns the third-party mentions and category presence that [Generative Engine Optimization](/glossary/generative-engine-optimization) rewards. Kompozy can't force ChatGPT to name you, but it removes the reason most brands never enter that first query: they are simply not being talked about often enough, in enough places, for any model to have noticed.
An August 2026 analysis by Suganthan Mohanadasan suggests it often does. When ChatGPT browses, it rewrites your prompt into its own search queries, and in 21 of 27 test conversations that first query already contained brand names the user never typed — implying the model chose likely recommendations from its prior knowledge, then searched largely to confirm. The findings are directional (single account, July 24–25, 2026), so treat them as a snapshot.
It is a field in the response stream your browser downloads when ChatGPT searches — currently named search_queries (previously search_model_queries) — that holds the queries ChatGPT wrote for itself, readable in browser DevTools. Analysts used it to see that ChatGPT names specific brands in its own queries before fetching any page. OpenAI can change or hide this at any time.
Because the model tends to name brands from its prior knowledge, the leverage is off-site presence: being written about in reviews, comparisons, roundups, and third-party content your buyers read, plus publishing specific, answer-shaped content consistently on pages you own. Merely being crawlable is not enough — in the study, brands only fetched but never named were cited about 2.1% of the time versus 68.9% for brands ChatGPT named itself.
No. Being fetchable is necessary but far from sufficient. The same researcher's earlier data showed small sites are eligible to be indexed, but this analysis shows the recommendation is largely decided before the fetch — so the real work is building enough distributed presence that the model already knows your brand.