ChatGPT Writes the Shortlist Before It Fetches Anything: 68.9% If You Are In the Query, 2.1% If You Are Not

On 10 August 2026, search consultant Suganthan Mohanadasan published an analysis that Search Engine Journal republished under his byline on 14 August, and on 17 August he walked the headline back a notch. The method is unglamorous: open Chrome DevTools and read the JSON payload ChatGPT sends to your own browser. The result is a stark pair of numbers. Brands appearing in the search string ChatGPT writes for itself reached the final answer 68.9% of the time. Brands merely fetched, never named in a query, reached it 2.1% of the time. A gap of roughly 33 times.

That stings because the entire AEO/GEO tooling market of the past 18 months is built on the second half. The IAB measurement framework published in early August reports only 16% of brands tracking AI visibility systematically, with the remaining 84% treated as addressable market. Ahrefs scanned 137,000 domains and found 97% of llms.txt files received zero AI requests in May 2026, with Fortune 500 adoption at 7.4%. A survey of 45 studies posted on 15 July 2026 is blunter still: no reviewed GEO technique produces a stable cross-platform effect, and body-only rewrites can cut a page's top-ten presence by 16%. Tools are sold on winning after retrieval. This research says the decision sits before it.

Adjacent measurements use different methods and point the same way. SISTRIX measured ChatGPT rotating 74% of its citations weekly in April against 56% for Google AI Mode. Semrush analysed 126 million United States prompts and found only 36 of more than 1,200 tracked brands holding top-100 status on every platform every month. Paid tooling splits in two tiers: Peec AI starts at EUR 89 per month, Ahrefs Brand Radar runs 199 dollars per AI index or 699 for six, and Profound is contact-for-pricing. The category is expanding while the stability of what it measures has never been demonstrated.

What does that mean for a small business in Taiwan? If you are about to spend on schema, llms.txt and page speed in order to get into AI answers, that work sits in the 2.1% column. Below: what the 60 conversations measured, what three groups should do today, and why I think the first casualty is not technical SEO but the visibility subscription itself.

The Detail: 60 Conversations, 3,554 Pages, 110 Citations

The data is raw HTTP responses captured from a logged-in ChatGPT Plus account in Dubai between 24 and 25 July 2026: 57 conversations for citation measurements, 27 for the first-query test. The mechanism needs no privileged access — ChatGPT rewrites your prompt into search strings of its own, and those strings sit in the response the browser needs to render the page, under a key named search_queries.

The ordering test carries the weight. The author isolated the first user message and the first search query by timestamp, a point at which nothing has been fetched. In 21 of 27 conversations, that first query contained brands the user never typed. A request for the best AI note-taking app produced a string ending in Granola, Notion AI, Otter and four more products; the robot vacuum row named Roborock Saros and Dreame X50, current model numbers, unprompted.

The second filter is harsher. Of 3,554 retrieved pages, only 110 earned a citation, a rate of 3.1%. Position inside the domain group predicts nearly everything: 5.2% in first place, 2.4% in third and 0.3% from sixth onward. Piling pages from one domain into the same group backfires: two pages convert at 6.2%, six or more at 1.7%. One brand was fetched 66 times and never cited.

Immediate Actions for Three Kinds of Reader

Brand owners and small business operators

Ten-minute test: ask the exact "best [category]" question your buyers ask, five times, and read the string the model writes for itself.
Absent in all five runs means you are not in the category vocabulary and no page work fixes it; present in all five means budget belongs at the page layer.
Never judge from a single answer. The shortlist drifts.

Marketing and SEO practitioners

Split budget by column: absent from the query means spending on being written about, reviewed and listed, not on schema.
Run displacement prompts: alternatives to a rival, or wanting off one. Those name new brands directly, and your competitors are being recommended there for free.
Consolidate pages: one per purchase intent, answer sentence near the top, numbers in plain HTML text.

Developers and agencies

Turn the four-step DevTools recipe into a deliverable: filter for conversation, search for queries, record the first string, log one row per month.
Pull ChatGPT-User fetches from the server log and reconcile against pages that earned citations. Ship it as a conversion table.
Expect format drift: after the key rename in early August, captured fan-outs fell from twelve searches per answer to four.

Tool Comparison: Which Column Are You Paying For

Option	What it measures	Published price	Column
Peec AI	Cross-platform mention and citation tracking	From EUR 89/month, Pro EUR 199	2.1%
Ahrefs Brand Radar	Brand presence at AI index level	199 USD per index per month, 699 for six	2.1%
Profound	Enterprise visibility and source analysis	Contact for pricing	2.1%
Semrush AI visibility	Large prompt datasets, mention and citation counted separately	Enterprise add-on	2.1%, with a denominator
FanoutFox free extension	Reads fan-out strings and fetched-versus-cited from your own session	Free	Both columns
DIY: DevTools plus log scripts	String capture and fetch-to-citation conversion	Free, half a day of labour	Both columns

The last two rows are the point. The string that decides 68.9% against 2.1% is visible in your own browser in two minutes, with no subscription attached.

What Nobody Will Tell You

The sample is one account. The author states every percentage is directional rather than measured. Personalisation shows up in the data: a meal kit query returned a local Dubai company, and he never stated his location.
The author has a commercial interest, and discloses it. The recommended path routes toward Keyword Insights, which he co-founded, and FanoutFox, which he built. Disclosure is not neutrality.
The 68.9 against 2.1 comparison carries contamination risk. It uses whole conversations, so a brand learned in turn one and queried in turn two counts as pre-known. The author says the clean number is the 21 of 27.
Two more steps sit between a mention and a visitor. Similarweb found AI-recommended brands 2.5 times more likely to be visited within seven days, but 56% of that traffic arrives through branded search, which GA4 barely sees.

The No-Subscription Alternative

Five-run category test. Open ChatGPT in Chrome with DevTools, filter Network for conversation, submit your "best [category] 2026" question, search the response for queries and read the first string. Names appearing in all five runs are your real competitive set inside the model.
Run displacement prompts too. Alternatives to a rival, or wanting off one. These force the model to supply names itself, and one pass tells you whether yours is among them.
Build a fetch-to-citation table. Count monthly ChatGPT-User hits per URL from the access log, then reconcile against pages that earned citations. Fetched often and never cited means the content needs work.
De-duplicate by intent and move the answer sentence up. Keep the tightest match per purchase question and 301 the rest into it; put the answer sentence near the top with numbers as plain HTML text.

Checklist before you start:

☐ Is the category question phrased the way customers speak, not with your product name?
☐ Have you run the same question five times?
☐ Are you logging the string the model wrote, not the answer it gave?
☐ Have you run displacement prompts as well?
☐ Is there exactly one page per purchase intent?
☐ Is a monthly re-test scheduled?

FAQ

I asked once and did not see my brand. Is that the verdict?

No. Two of three repeat runs moved substantially: accounting collapsed from six vendors to a single probe at QuickBooks, and web hosting dropped every vendor for a review site. SISTRIX separately measured 74% weekly citation rotation. A single result carries no judgement. Run it five times.

Are llms.txt and schema wasted effort?

Two separate questions. They cannot influence the shortlist step, because the model writes the names before contacting your server; but in the citation contest after retrieval, clean HTML and an explicit answer sentence still matter.

Is this test worth running for a small business in Taiwan?

Yes, and the result is usually more extreme. Training-data density for Chinese-language category questions is far lower than for English, so many local categories have no stable shortlist inside the model, and the first string tips toward large platforms or price-comparison sites.

So where should the budget go?

Split it by column. Absent from the string: spend on reviews, comparisons and roundups, which map to the 68.9% column. Already in the string: page consolidation and answer-sentence work, mapping to the 3.1% column. Buying them as one bundle is the expensive mistake.

One account only. Is that credible?

The mechanism is credible and the percentages are directional. The query strings are the artefact rather than an inference about it, and you can reproduce them in two minutes. But a single account cannot separate model-level knowledge from personalisation, which the author lists as the study that would settle it. Use it as a diagnostic method, not as a benchmark to cite.

My Take

The marketing conversation is collapsing this research into a slogan: technical GEO is dead, go do digital PR. I expect that reading to look wrong within 90 days, and the direction is backwards.

What it kills first is not technical optimisation. It is the pricing basis of AI visibility subscriptions. The core asset those dashboards sell, starting around 199 dollars a month, is telling you whether AI mentioned you. What this article demonstrates is that the string deciding exactly that sits in your own browser's response payload, four steps and two minutes away. Once the raw signal is free and self-serve, a subscription can only sell scale and labour — a reporting business rather than a data business. Vendors read the IAB figure of 16% as headroom. I read it as most of the remaining 84% needing one diagnosis done properly.

For a studio like ScriptWalker the opportunity is concrete: a one-off AI category vocabulary audit. Four deliverables — the competitive set from the five-run test, the displacement-prompt list, a fetch-to-citation table built from the client's own access log, and a page consolidation and rewrite plan. No subscription, half a day of labour, and the client can decide between content spend and media placement. Attach a quarterly re-test, because the shortlist drifts, which is a natural renewal reason.

To outsource the audit, or to discuss wiring it into an existing site:

Sources

Primary sources

Third-party sources