ChatGPT decides which brands it is going to talk about before it retrieves a single page. Research published by Suganthan Mohanadasan on Search Engine Journal on 14 August 2026 found that in 21 of 27 tested conversations, ChatGPT wrote brand names into its own search query that the user had never mentioned, and that brands appearing inside that self written query were named in the final answer 68.9 percent of the time. Brands whose pages were fetched but which never appeared in the query text were named 2.1 percent of the time, a gap of roughly 33 times.
That ordering is the whole finding. Almost everything the search industry calls optimisation happens after retrieval has already started. If the model has written a shortlist into its own query first, a technically flawless page can still be competing for a place it was never eligible for.
What the research measured, and how it was done
Suganthan Mohanadasan published both the method and the numbers on Search Engine Journal on 14 August 2026. When ChatGPT runs a web search, it emits a JSON object carrying a search_queries field, which holds the actual query strings the model composes for itself. That object travels in ordinary network traffic and is readable in a browser developer console. The reproduction steps described in the article are short: open ChatGPT in Chrome, open DevTools, watch the Network tab, filter for the conversation request, open the response and search inside it for the queries field. Anyone with a ChatGPT account can run that inspection on their own session, which is unusual for a claim about model behaviour.
The citation analysis covered 57 conversations. The first query testing covered 27 conversations, built from 12 fresh category queries and 3 repeated ones. Across the citation set, 3,554 pages were labelled as either retrieved or cited, and 110 of those pages were actually cited, a citation rate of 3.1 percent. Of the 27 first queries examined, 21 contained a brand name the user had not typed anywhere in the prompt. Mohanadasan reported that 11 of 13 categories tested showed injected brand names, and that the sample leaned heavily toward software and AI tool categories.
The evidence he treats as primary is the raw query strings themselves, with the percentages described as approximate. That distinction is worth carrying forward. The query strings are observable artefacts. The percentages are counts drawn from a small, single account sample.
The two filters, in the order they run
The model of ChatGPT behaviour that Mohanadasan set out on Search Engine Journal on 14 August 2026 has two sequential gates rather than one. Filter one decides which brands enter the retrieval query at all. This is a pre retrieval decision, made from what the model already carries internally, and it is set by training data and by whatever general reputation signal the brand accumulated before the conversation began. Filter two is the ordinary citation contest: of the pages that come back, which ones get quoted, based on position in the result set, topical relevance and how well the page consolidates an answer.
Filter two is where the familiar work lives. Page structure, clean markup, answer density, internal linking, server response, all of it operates on pages that have already been fetched. Filter one runs before any of that is even loaded. The line Mohanadasan uses for it is short: the decision happens before anything touches your server.
The practical consequence is that two brands can publish equally good pages on the same topic and see completely different outcomes, because one of them was written into the query and the other was not. The page quality difference is not what separated them.
The numbers behind the 33x figure
The following figures are the ones reported in the Search Engine Journal article of 14 August 2026, and they are the whole quantitative basis for the claim. Nothing here is projected or modelled.
| What was measured | Reported figure |
|---|---|
| Conversations reviewed for citation analysis | 57 |
| Conversations tested for the first query ChatGPT writes | 27 |
| First queries containing a brand the user never mentioned | 21 of 27 |
| Brands named inside ChatGPT's own query that reached the answer | 68.9 percent |
| Brands whose pages were fetched but never named in the query | 2.1 percent |
The 33x headline is the ratio between the last two rows. It is a ratio between two proportions inside one sample, not a measured multiplier of business outcomes, and it should be quoted that way.
Where conventional optimisation still does work
Nothing in the Search Engine Journal research of 14 August 2026 says technical and on page work stopped mattering. It says that work operates on filter two. The 3.1 percent citation rate across 3,554 labelled pages is filter two in action: even among pages ChatGPT actually pulled, the large majority were never quoted. Whatever separates the 110 cited pages from the rest is exactly the territory that page level optimisation addresses.
So the honest reading is additive rather than replacive. If a brand is already inside the query, page quality decides whether it gets quoted. If a brand is not inside the query, page quality is being applied to a contest it entered late. Teams doing serious ChatGPT SEO work need both halves, and the second half is the one that has been measurable for years.
The limits of this evidence, stated plainly
This is one researcher, one account, and a sample of tens of conversations rather than thousands. Mohanadasan says so in the article himself, describing the results as directional rather than definitive and calling for multi account follow up testing. The sample is weighted toward software and AI tools, so the behaviour observed in that vertical may not hold in hospitality, healthcare, industrial supply or retail. Personalisation is an acknowledged confounder: the article notes location based variation in results, which means two people running the same prompt can see different query strings.
There is no peer review here, no vendor replication, and no independent audit of the counts. The percentages are approximate by the author's own description. A finding this interesting invites overstatement, and it does not support any claim that brand mention volume causes citation, only that the two co occur strongly in this sample. Treat it as a well documented observation that deserves replication, not as a law of how generative engine optimisation works.
What the research did not say
The article did not say how a brand gets into the model's parametric knowledge, beyond the general statement that training data and reputation determine it. It gave no timeline, no threshold, and no method for influencing that. It did not test whether paid coverage, press volume, review site presence or social discussion move the needle, and it did not compare ChatGPT against Gemini, Claude or Perplexity. It also did not measure revenue, clicks or conversions of any kind. Anyone extending these numbers into a claim about pipeline is adding something the source did not contain.
It also did not say that the ordering is fixed forever. The behaviour described is the behaviour of one product at one point in August 2026, observed through a network response format that OpenAI can change without notice.
What a marketer can check on their own account today
The reproduction step is the useful part of this research, because it turns an argument into an observation. Working through it takes a few minutes.
- Open ChatGPT in Chrome, open DevTools, select the Network tab, then send a category question in your own market with no brand names in it, such as a request for the best supplier of something you sell.
- Find the conversation request in the network list, open its response, and search the payload for the queries field. Read the ChatGPT search queries the model composed for itself.
- Write down every brand name that appears in those query strings. Those are the brands the model already carries for your category.
- Repeat the same prompt several times, and from a different location if you can, because the article documents location based variation and one run proves very little.
- Compare that list against your own visibility work. If your brand never appears in the query text across repeated runs, the gap sits at filter one, and more page level tuning will not close it.
The output of that exercise is a list, not a strategy. But it does tell a team which of the two filters is actually blocking them, and that is a question most reporting cannot currently answer.
What this means for Thai marketers
The research did not test Thailand, Thai language prompts, or any Southeast Asian category, so nothing here can be read as a measurement of the Thai market. What can be reasoned about is the shape of the problem. Thai brands in categories where the international vendors dominate English language coverage are the ones most exposed to a pre retrieval filter, because parametric knowledge is built from training data that skews toward widely published English sources.
That points at unglamorous work. Being present and consistently named in the places that get indexed and quoted in your category, in both Thai and English, is the input to filter one, and it accumulates over quarters rather than weeks. Marketers running Thai and English versions of the same site also have a practical reason to check both language versions separately, since the query strings the model writes for a Thai prompt need not match the ones it writes for the English equivalent. Teams already investing in AI search visibility can add the DevTools check above to their monthly reporting at close to zero cost, which is a rare thing in this area.
FAQ
Does this mean SEO no longer works for ChatGPT?
No, and the research does not claim that. It describes two sequential filters and places conventional technical and on page optimisation at the second one, where a 3.1 percent citation rate across 3,554 labelled pages shows there is still a real contest to win. The claim is that page work alone cannot fix a brand that never enters the model's own query.
Is this behaviour live in Thailand?
The source did not state this. The testing was done on a single account weighted toward software and AI tool categories, with no country breakdown published, so there is no evidence in the article about Thai prompts or Thai users. The DevTools inspection described in the study can be run locally from Bangkok to see what the model writes for your own category.
Do I have to do anything right now?
Nothing urgent, because a single small study is not a reason to rebuild a programme. The low cost step is to run the network inspection on a handful of category prompts and record which brands appear in the query strings, which gives you a baseline before deciding whether anything needs to change.
How do I get my brand into ChatGPT's parametric knowledge?
The article did not answer this. It attributes filter one to training data and general reputation without naming a mechanism, a threshold or a timeframe, and it tested no tactic for moving it. Any specific method being sold as the answer to that question is going beyond what this research supports.
How reliable are the 68.9 percent and 2.1 percent numbers?
They are approximate counts from one researcher's single account sample of 57 and 27 conversations, described by the author as directional rather than definitive. The author treats the raw query strings as the primary evidence and the percentages as secondary, and he asks for multi account replication before anyone treats the ratio as settled.
If you want a second pair of eyes on where your brand sits in AI answers for your category, the team at Relevant Audience is happy to run the check with you and talk through what the numbers do and do not support. The original write up by Suganthan Mohanadasan is worth reading in full on Search Engine Journal.
