A Norwegian search consultant read the network traffic behind 60 ChatGPT conversations and found that the assistant writes competitor names into its own search string before it fetches a single page. Brands that appear in that string reached the final answer 68.9 percent of the time. Brands that were merely retrieved during the search reached it 2.1 percent of the time.

Suganthan Mohanadasan published the analysis on 10 August 2026 at suganthan.com, then revised its headline on 17 August 2026 after further testing. The work sits on a narrow but unusually direct evidence base: raw HTTP responses captured from a logged-in ChatGPT Plus account in Dubai between 24 and 25 July 2026, covering 57 conversations for the citation measurements, 27 for what he calls the first-query test, twelve fresh category queries and three repeat runs.

The mechanism is the part he presents as solid. Everything expressed as a percentage, according to Mohanadasan, is directional rather than measured, because all of it comes off a single account.

What the browser already downloads

When a user submits a prompt, ChatGPT rewrites it into search queries of its own, executes them, reads the returned pages and composes an answer. Those rewritten queries travel back to the browser inside the response payload that renders the conversation. According to Mohanadasan, they currently sit under a JSON key named search_queries, which OpenAI renamed from search_model_queries in early August 2026.

Nothing about the observation depends on privileged access. The data is required by the browser to draw the page. Mohanadasan documents a four-step reproduction: open ChatGPT in Chrome, open DevTools, filter the Network tab for conversation, and search the response for queries.

The example that opens his article concerns note-taking software. He asked for the best AI note taking app and named no product. Before any fetch occurred, the assistant wrote a search string containing Granola, Notion AI, Otter, Fireflies, Fathom, Mem and Limitless. It then ran nine further searches, each one a site: probe pointed at a single vendor's own domain.

That sequence reframes what the industry has been calling fan-out. Rather than a broad hunt for candidates, the pattern Mohanadasan describes is a single query that names a shortlist followed by one confirmatory search per name on it. PPC Land documented the fan-out behaviour in June 2026, when Peec AI analysed 5 million query fanouts collected across ChatGPT, Perplexity and Grok and found the assistant injecting terms including best, reviews and the current year into strings where the user had typed none of them. The new work extends that finding from vocabulary to entities.

The timing test

The obvious objection is sequencing. If the assistant searched once, saw a set of brands, then wrote a smarter second query, the names would be a consequence of retrieval rather than a cause of it.

Mohanadasan tested for that directly. For each conversation he isolated the first user message and the first search query by timestamp, a point at which nothing has been fetched and no earlier result exists to have learned from. In 21 of 27 conversations, that first query contained brands the user never typed.

Two rows in his results table sit next to each other and differ only in phrasing. A request for the best AI based live chat support software produced a string naming Intercom Fin and Zendesk AI. A request for the best AI based live chat software produced Intercom Fin, Zendesk AI, Ada, Decagon and Sierra. The list stretched from three names to seven on a wording change, which he reads as evidence against a fixed lookup table.

He then ran twelve unrelated categories. Eleven of thirteen behaved the same way. Language learning returned Duolingo, Babbel, Busuu, Pimsleur, Speak and LingQ. Accounting software returned Xero, QuickBooks, Zoho Books, FreshBooks, Wave and Sage. Web hosting returned Hostinger, SiteGround, Cloudways and Kinsta.

The robot vacuum row is the one he singles out. The first query named Roborock Saros, Dreame X50, Narwal Freo Z10 and Eufy S1 Pro. Current model numbers, unprompted, in the string that precedes any fetch.

Electric SUVs broke the pattern in a different direction. For accounting, therapy and hosting the assistant named vendors and went to their pricing pages. For cars it named magazines: Car and Driver, Edmunds, Top Gear.

Instability between runs

Mohanadasan re-ran three categories to test whether any of this holds. Language learning barely moved, returning five of six names unchanged. Accounting collapsed from six vendors to a single targeted probe at QuickBooks. Web hosting dropped every vendor and went to a review site instead.

His conclusion is narrow. The vendor-versus-publication split is a tendency that can flip between runs rather than a fixed property of a category. Two things survived the repeats: the injection occurred every single time, and the category with the clearest market leaders kept its names.

That volatility echoes measurements taken by other methods. SISTRIX found in April 2026 that ChatGPT rotates 74 percent of its citations per week, against 56 percent for Google AI Mode. Semrush research covered by PPC Land put monthly citation-set change at roughly 50 percent with only 11 percent overlap between platforms.

Where the behaviour stops

Every query to that point had contained the word best, so Mohanadasan tried to break the pattern with twenty-four further queries across seven shapes that avoided it. The word turned out to be irrelevant. What determined the behaviour was whether the assistant had to supply the products itself.

Seven of the twenty-four never triggered a web search at all. Questions about how noise cancelling works, or what a vector database is, were answered from training. So was open-ended complaint. A prompt stating that a company was spending too much on customer support tooling produced no search.

Naming the brands removed the injection. A comparison between Xero and QuickBooks went straight to a probe at xero.com. A question about whether to use HubSpot for a small agency went to hubspot.com.

Leaving the candidates open produced the injection in ten of the eleven cases where a recommendation was requested without any product named. A prompt about meeting notes being a mess, which is a complaint rather than a shopping question, produced a first search for Granola pricing. A prompt about cleaning floors automatically, which avoided the phrase robot vacuum entirely, went to a specific Roborock model's own website.

Three displacement queries all pulled in new names. A request for alternatives to Zendesk produced Help Scout. A request for something to use instead of QuickBooks produced Zoho Books. A request for something like Duolingo but better for grammar produced Kwiziq and Babbel.

The 33 times gap

Mohanadasan sorted every brand into two groups: those appearing in a query the assistant wrote, and those retrieved during the search without ever being named. He then checked how often each group made the final answer.

Of 119 brands named in ChatGPT's own query, 68.9 percent were mentioned in the answer. Of 515 brands fetched but never named, 2.1 percent were. He also recorded 86 cases in which a brand was recommended without its website being fetched in that conversation at all.

He is explicit about a contamination risk in that comparison. The 68.9 against 2.1 figures use whole conversations, so a brand learned in turn one and queried in turn two counts as pre-known. The clean number, he writes, is the 21 of 27 first-query result. A gap of 33 times, in his reading, is too large for turn-order contamination to explain on its own.

The commercial reading is blunt. Most of what is currently sold as generative engine optimization operates on retrieval, which is the 2.1 percent column. Presence in the query is settled before any of that work executes.

A second filter, and a harsher one

Being named in the query is an entry ticket rather than a result. Mohanadasan built a labelled dataset from 57 conversations, treating every retrieved page as a row and citation as the label. Of 3,554 pages, 110 earned a citation. That is 3.1 percent. On his arithmetic the assistant reads around 600 pages to write one answer and credits about 30.

Three factors separated the cited from the ignored.

Position within the domain group predicted most of it. ChatGPT groups results by domain, and cite rates fell from 5.2 percent in first position to 4.6 percent in second, 2.4 percent in third, 1.7 percent in fourth, 0.6 percent in fifth and 0.3 percent from sixth onward.

Page volume from a single domain worked against the domain past a narrow optimum. One page in the group converted at 4.0 percent and two at 6.2 percent, but five fell to 1.9 percent and six or more to 1.7 percent.

Relevance qualified without selecting. Mohanadasan scored each cited page against every other page retrieved for the same claim, measuring how well its text matched the sentence being supported. The cited page sat in the top 5 percent of the pool. It was the single best match only 20 percent of the time, and its average overlap ran well below the best available.

One brand in his data was fetched 66 times and never cited once.

Two games, not one

The findings split AI visibility into what Mohanadasan frames as separate problems that the market treats as a single discipline.

The first is presence in the category vocabulary. If the assistant does not already associate a brand with its category, it will not write the name into the query, and the observed mention rate falls to around 2 percent. Schema markup, page speed and an llms.txt file cannot address that, on his reasoning, because the server is never contacted before the decision is made. What builds the association, in his account, is being written about, reviewed, compared and argued over across the open web until it exists in the training data.

The llms.txt point has independent support. Ahrefs server-log research across 137,000 domains found 97 percent of those files received zero requests in May 2026, with SEO audit tools rather than AI assistants accounting for the largest share of the requests that did occur. A March 2026 study put Fortune 500 adoption at 7.4 percent.

The second game is conversion once retrieved, and Mohanadasan treats it as real work: one tightly matched page per intent, the claim-bearing sentence early, facts and numbers in plain HTML text rather than in images or JavaScript-loaded elements, and no cluster of near-identical pages competing with each other.

He discloses a commercial interest at that point. Keyword Insights, a keyword clustering product he co-founded, is presented as suited to finding the intent overlaps the second game requires. FanoutFox, a free Chrome extension he built, reads the same session data the article analyses. His agency is Snippet Digital.

What the author says the data cannot support

Mohanadasan revised the article's title on 17 August 2026. The original wording, he writes, "could be read as claiming the outcome is decided before the search runs," a stronger claim than the article makes. Further testing, according to the update note, showed that brands never appearing in the model's own queries still reach the final answer, including the recommendation itself. He states that neither the numbers nor the mechanism changed on re-checking.

Other caveats are stated plainly. Every percentage comes from one account, weighted toward software and AI tools. His brand-matching script split Car and Driver into two tokens, once counted SaaS as a brand, and on a local query counted Arabian Ranches and Dubai as brands.

Personalisation appeared in the data. A meal kit query returned UAE, Dubai and a local company. A travel insurance query went straight to a UAE insurer's site. He never stated his location in either prompt.

There is also a moving-target problem inside his own capture window. In early August 2026 OpenAI renamed the JSON key, and the fan-outs in his captures dropped from twelve searches per answer to four.

Why this lands now

The research surfaced during a week in which the vocabulary of AI visibility was already under pressure. Rand Fishkin, co-founder of SparkToro and co-author of Zero Click Marketing, posted on LinkedIn that new research suggests the AI tool already knows which brands it is searching for before any citation appears. Advice to get into LinkedIn articles, be present in Reddit threads or follow the citations in a given industry, Fishkin wrote, "might sit somewhere between potentially misleading and totally useless." The post carried 421 reactions, 134 comments and 23 reposts when captured.

The claim cuts against a measurement industry that has grown quickly. The IAB published a framework in early August 2026 noting that only 16 percent of brands systematically track AI visibility, and requiring vendors to disclose whether their prompt libraries are synthetic or drawn from observed behaviour. Semrush analysed 126 million United States AI search prompts and found that only 36 of more than 1,200 tracked brands held top-100 mention status on every platform in every month, while treating mention and citation as structurally separate metrics.

Academic scrutiny has been colder still. A critical survey of 45 studies posted to arXiv on 15 July 2026 concluded that no reviewed generative engine optimization technique produces a stable cross-platform effect on discoverability or downstream traffic, and that body-only rewrites can cut a page's top-ten presence by 16 percent.

Mohanadasan's contribution is a different kind of evidence. Where most published work in this area infers behaviour from outputs, his reads the request stream that produces them. That method has limits a large sample would not fix, since a single account cannot separate model-level knowledge from personalisation. It also has a property no synthetic prompt library has: the queries are the artefact, not an inference about the artefact.

The distinction matters commercially because the two games carry different price tags and different vendors. Technical audits, structured data and machine-readable files are sold as AI strategy at a scale the retrieval column does not appear to justify. Digital PR, review-site placement, analyst coverage and category content are slower, harder to attribute and, on this evidence, closer to the mechanism that decides which names appear.

Similarweb clickstream research covered in June 2026 found AI-recommended brands 2.5 times more likely to receive a site visit within seven days, with 56 percent of that traffic arriving through branded search rather than a direct click. If the name in the query is what determines the recommendation, and the recommendation is what produces the branded search, then the measurable end of that chain sits three steps downstream from the decision.

Mohanadasan has requested a full ChatGPT data export to run the same analysis across his entire conversation history, and describes a multi-account category sweep as the study that would settle the question. If the same brands are injected for the same questions across different accounts, the knowledge is model-level. If they are not, personalisation is doing more work than the industry currently assumes.

Timeline

Summary

Who: Suganthan Mohanadasan, a Norwegian search practitioner based in Dubai and co-founder of Keyword Insights and Snippet Digital, conducted the analysis. Rand Fishkin, co-founder of SparkToro, amplified the underlying claim on LinkedIn. The subject is OpenAI's ChatGPT and the brands competing for mentions inside its answers.

What: An examination of raw network traffic from 60 ChatGPT conversations finding that the assistant writes brand names into its own first search query before fetching any page. Brands named in that query reached the final answer 68.9 percent of the time against 2.1 percent for brands merely retrieved, a gap of about 33 times. A separate labelled dataset of 3,554 retrieved pages found 110 cited, a rate of 3.1 percent, with position inside a domain group the strongest predictor.

When: Data was captured between 24 and 25 July 2026. The analysis was published on 10 August 2026 and its title revised on 17 August 2026 after additional testing.

Where: The capture came from a single logged-in ChatGPT Plus account in Dubai, United Arab Emirates. Personalisation appeared in the results, with local companies and UAE-specific terms surfacing in queries where no location was stated.

Why: The finding relocates the decisive moment in AI brand visibility to a point before any server is contacted, which places most technical optimization work downstream of the outcome it claims to influence. For marketers allocating budget between technical audits and the slower work of earning third-party coverage, the split between a 68.9 percent column and a 2.1 percent column is the operative number.