Hi everyone, this is Neo.

I came across a study the other day with a headline that cuts deep: ChatGPT Already Knows Who It’ll Recommend Before It Searches.

The author is Suganthan Mohanadasan, co-founder of SEO agency Snippet Digital. He spent two months digging through his own ChatGPT sessions, reading hundreds of the real search queries ChatGPT writes for itself, and reached a conclusion that should make the whole GEO industry uncomfortable:

Before ChatGPT fetches a single page, it already writes a batch of brand names into its own search queries. If your brand isn’t on that pre-decided shortlist, your website never gets visited — no matter how well it’s built.

That one sentence undermines a big chunk of the “AI search optimization” playbook that’s been sold for the past two years: structured data, page speed, llms.txt, schema tweaks — all of it is decoration if your name is never spoken in the first place.

Let me break down the study, and what it means for anyone running an indie site.

The method: reading ChatGPT’s inner monologue

First, the methodology — because it determines how much you should trust the findings.

Mohanadasan discovered that when you ask ChatGPT a question, it rewrites your question into its own search queries, runs them, reads the results, then writes an answer. Those queries sit in the JSON your browser downloads, under a field called search_queries (OpenAI renamed it from search_model_queries in early August 2026).

This isn’t a leak. Your browser needs that JSON to render the page. Open DevTools and you can read ChatGPT’s own search terms in about two minutes.

Example. He asked ChatGPT for the best AI note-taking app — seven words, zero brand names. The search query ChatGPT wrote for itself:

best AI note taking apps 2026 official pricing features Granola Notion AI Otter Fireflies Fathom Mem Limitless

Look at the tail: Granola, Notion AI, Otter, Fireflies, Fathom, Mem, Limitless — seven products, none named by the user, nothing fetched yet. Those names came out of the model.

Then it ran nine more searches:

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
site:notebooklm.google official features pricing 2026

See the pattern? One search builds the shortlist, then one search per name on that shortlist, each pointed straight at that company’s own website. The fan-out was never a search for candidates — it’s ChatGPT walking down a list it already has, one name at a time.

The numbers: 21 of 27, and a 33x gap

One example isn’t proof, so he tested it properly:

• In 21 of 27 conversations, the very first search query contained brands the user never typed (he controlled for the timing objection by taking the earliest query by timestamp — nothing had been fetched yet, so no earlier result could have influenced it)
• He ran 13 categories that have nothing to do with each other — 11 of 13 behaved the same way
• The robot vacuum category was the wildest: ChatGPT recalled Saros, Dreame X50, and Eufy S1 Pro — current model numbers, unprompted, in the first query

Then came the most valuable comparison: he split brands into two groups — those that appeared in a query ChatGPT wrote, versus those that were only fetched during the search without ever being named — and measured how often each group made it into the final answer.

Result: the query-named brands got cited roughly 33 times more often.

He also built a labeled dataset from 57 conversations: 3,554 retrieved pages, of which only 110 were cited. A 3.1% citation rate. ChatGPT reads about 600 pages to write one answer and credits about 30.

And one number that deserves its own highlight: 86 cases where a brand was recommended and its website was never fetched at all. A mention doesn’t need a crawl — the model can recommend you from memory.

68.9% vs 2.1%: the emperor’s new clothes of GEO

Here’s the part that stings:

• 68.9% of recommended brands were already known to ChatGPT before it wrote its queries
• Only 2.1% were “discovered” by the retrieval itself

Translation: most of what’s sold as GEO today is retrieval work — and retrieval work is the 2.1% column. The 68.9% column is decided before any of it runs.

The author says it flat out: “Most of what’s sold as GEO right now is retrieval work, which is the 2.1% column. The 68.9% column is decided before any of it runs.”

He also verified that llms.txt and similar “AI entrance files” are nearly useless for this problem — if your brand isn’t in the query, your server is never contacted, so the file never gets read. That lines up neatly with Ahrefs’ earlier finding that 97% of llms.txt files received zero requests.

Being read isn’t being cited: the brutal 3.1% funnel

Even if you make the shortlist and get crawled, you’re not done. Three factors decided who got cited out of that 3.1%:

1. Position, hard. ChatGPT groups results by domain, and where you sit in that group predicts nearly everything. Below the top two, citation is a rounding error. Being in the retrieved set isn’t a win if you’re ninth.

2. Piling on pages hurts. When several pages from one domain show up in the same group, per-page conversion collapses. Two tightly matched pages is the sweet spot; past six, you’re mostly competing with yourself.

3. Relevance qualifies you; it doesn’t select you. Cited pages averaged in the top 5% of relevance for the claim being supported, but were the single best match only 20% of the time. Relevance gets you into the conversation; something else picks the winner.

So the full funnel is: name in the query (68.9% decided before this) → retrieved (~600 pages) → cited (~30 pages) → in the answer (~3.1%). Every layer leaks, and the top layer is exactly the one technical optimization can’t touch.

What to do: the four-step plan

The good news: the author didn’t just tear things down — he gave a practical path. Indie site owners can steal it directly:

Step 1: Check whether you’re on the list. Ask ChatGPT the “best [your category]” question your buyers ask, and read the first search query it writes. Run the same question five times — the list changes between runs, so a single run will mislead you. Present every time = you’re solid in the AI’s memory. Inconsistent = contested ground, where a push can move something. Absent in five runs = your problem isn’t technical, it’s existence.

Step 2: Not on the list? Spend on getting written about. The decisive factor is brand association in the training data. What builds that? Being reviewed, compared, recommended, covered by media, and listed in the roundups your buyers read — digital PR and category-defining content. It’s slow and unglamorous, but it’s the foundation work.

Step 3: On the list? Now do the technical work. This is when page-level optimization pays: one tightly matched page per intent, the claim-bearing sentence near the top in plain HTML text, facts and numbers in the body. Merge or redirect near-identical pages fighting over the same intent.

Step 4: Watch for “read but never cited” domains. His data had a brand fetched 66 times and cited zero times — that’s not an awareness problem. The engine kept going back and kept deciding there was nothing worth quoting. That’s a content-quality signal.

Neo’s take

Here’s what this study really did: it peeled back a corner of the emperor’s new clothes in GEO.

For two years, the market has been flooded with “AI search optimization” courses and tools — structured data, semantic optimization, page speed, llms.txt, entity graphs. None of it is useless, but all of it optimizes the retrieval stage. And this study proves, with the simplest possible method (open DevTools, read the queries), that the bulk of the recommendation decision happens before retrieval — in the model’s existing knowledge of your brand.

For indie site owners, my judgment:

1. “Brand building” is no longer soft — it’s the hard currency of the AI era. We used to build brands for premium pricing and repeat purchases. Now brand determines whether AI recommends you at all. Your “presence” in model training data is your AI ranking. And the way to build presence is deeply traditional: being written about, reviewed, compared, and cited. The slow work became the fast lane.

2. Reallocate your SEO budget. If you’re not on the AI’s pre-decided shortlist, spending on technical optimization is building on sand. Do the existence work first (PR, reviews, listings), then the page work. And don’t get harvested by “AI visibility audit” tools — per the author, you can check this yourself in two minutes with DevTools.

3. Don’t be spooked by single-study numbers. The limitations are obvious: one person, one account (a ChatGPT Plus account in Dubai), a few hundred conversations, software-heavy categories. The percentages are directions, not precise measurements. But the direction is clear: brand awareness beats retrieval optimization. That aligns with multiple independent studies on citation rates and AIO citation distribution.

4. Beware “recommendation as truth.” One detail in the study: ChatGPT’s picks were heavily personalized — his meal-kit query returned a local UAE company. Model recommendations are shaped by user profiles; they’re not neutral rankings. AI recommendations are becoming editorials with a point of view. Your job isn’t to game the algorithm — it’s to be a name worth remembering.

Here’s my honest bottom line: in the AI search era, the biggest barrier was never technical. It’s being remembered. Technical optimization is the how; brand existence is the whether. This study’s real gift is putting those two back in the right order — first make ChatGPT know you before it searches, then worry about how you perform after it finds you.

That’s all for today. Go open DevTools yourself and see whether your brand shows up in the first query ChatGPT writes for your category — and feel free to tell me what you find.