How does ChatGPT decide which brands to name?

We logged every URL ChatGPT fetched before naming a skincare brand, registered one number before looking, and found that 88.6% of what the model read was third-party pages — not brand sites. One word in the question swings the own-site-fetch rate by a factor of ten.

Most times ChatGPT named a skincare brand, it never opened that brand’s own site: 61.8% of name-checks included no own-site fetch at all, and only 38.2% did (95% CI 35.8%–40.8%). It builds the answer mostly from third-party editorial, clinical, and retail pages — 88.6% of what it read. The lever is not your own site; it is whether the question demands a fact only your site holds.

20 July 2026Results registered before data was seen

Most brands believe that getting recommended by ChatGPT is a problem of visibility — that if your site is well-optimized, authoritative, and crawlable, the model will find you, cite you, and name you when a shopper asks. This assumption drives content budgets, technical SEO sprints, and structured-data rollouts aimed squarely at making brand pages legible to machines. The wound it leaves is quiet but real: brands invest in the wrong surface, measure the wrong outcome, and cannot explain why they never appear in answers no matter how much they spend.

The data breaks the assumption that recommendation is a visibility contest. ChatGPT does not scan the open web for your brand and then decide whether to recommend it. It assembles a reading list on the fly from editorial, clinical, and retail pages, and it names brands from that list. Your site is a minor contributor to the list — most of the time, it is not on the list at all. The sources it fetches before it generates a name are overwhelmingly third-party.

One number, locked before we looked

Named without a visit

How often ChatGPT named a skincare brand without ever opening that brand’s own site.

61.8%

A name-check is one (call, brand) pair in which ChatGPT named the brand. An own-site fetch is any case where the named brand’s own domain appeared among the pages the model fetched for that same call. We ran 1,444 name-checks in the skincare category — skincare because it passed the pre-registered pilot gate for this study window. Before looking at any results, we locked in one registered number: the share of name-checks that included an own-site fetch. Every other figure in this piece is descriptive. We also introduced two invented brands as controls; both were never recommended and never fetched, confirming the instrument behaved as designed.

Named without a visit

Three buckets, one name-check

For every brand ChatGPT named, the pages it had fetched fell into one of these.

38.2%
Own-site fetched
15.8%
Competitor site, not named brand
46.0%
No brand site at all

ChatGPT names most brands from a reading list it assembles on the fly from editorial, clinical, and retail pages — not your own site. The registered own-site-fetch rate is 38.2% (Wilson 95% CI: 35.8%–40.8%; n=1,444; skincare; day-0). The majority path, 61.8%, included no own-site fetch at all. In 46.0% of name-checks, no brand site appeared among fetched pages — not the named brand’s, not any competitor’s. A further 15.8% fetched a competitor’s site but not the named brand’s. Whether your site enters the reading list depends on one variable: whether the question demands a fact your pages uniquely hold.

One word changes the fetch

One word changes the fetch

Same brands, eight ways of asking. Own-site fetch runs from 8.5% to 84.9%.

best skincare brands under $30
84.9%
recommend 5 skincare brands and say why
73.3%
best affordable drugstore skincare brands
67.1%
rank the 5 best skincare brands
37.9%
top skincare brands
22.9%
skincare brands worth buying
18.6%
best skincare brands for sensitive skin
10.2%
best skincare brands
8.5%
Bottom · open-ended wording pulls brand pages least (8.5%).
Top · a price or use-case constraint pulls them most (84.9%).

The phrasing of the question is the lever. Own-site-fetch rates swing tenfold across eight phrasings, from 8.5% for ‘best skincare brands’ to 84.9% for ‘best skincare brands under $30’. The full ordering, low to high: ‘best skincare brands’ 8.5%, ‘best skincare brands for sensitive skin’ 10.2%, ‘skincare brands worth buying’ 18.6%, ‘top skincare brands’ 22.9%, ‘rank the 5 best skincare brands’ 37.9%, ‘best affordable drugstore skincare brands’ 67.1%, ‘recommend 5 skincare brands and say why’ 73.3%, ‘best skincare brands under $30’ 84.9%. Generic questions yield editorial and retail aggregators. Questions that embed a constraint — price, format, audience — pull pages that hold the specific fact, and brand sites are where those facts live.

Where the fetched pages came from

Share of the pages ChatGPT read across 1,444 name-checks. Third-party dominates; own-brand is one bar in ten.

Reference (Wikipedia etc.)
1.6%
UGC / social
1.7%
Recommended brand own site
8.2%
Third-party editorial, clinical, retail
88.6%
Sources seen
aad.orgnationaleczema.orgallure.comgoodhousekeeping.comsephora.comulta.com

The source mix confirms why. Across the 1,444 name-checks, 88.6% of fetched pages were third-party editorial, clinical, or retail; own-brand pages accounted for 8.2%; UGC and social (Reddit, Quora, YouTube, TikTok, and others) for 1.7%; and reference sites (Wikipedia, Britannica) for 1.6%. The top fetched sources were aad.org (dermatology), nationaleczema.org, allure.com, goodhousekeeping.com, sephora.com, and ulta.com. The model is mostly reading pages the brand does not control.

Fetched often, recommended rarely

Fetch did not predict a name-check

Each dot is one brand: how often ChatGPT fetched its own site vs. how often it named it.

00252550507575100100if fetch → recommendOwn-site fetch rate (%)Recommendation rate (%)CeraVe · fetched 38%, named 99.8%L'Oreal · fetched 23%, named 3.2%

The finding does not travel evenly across brands. Anchor brands — CeraVe and L’Oréal — showed a 50.9% own-site-fetch rate when recommended, while mid-tier brands sat at 33.0%. L’Oréal was fetched in 22.8% of queries but recommended in only 3.2%; CeraVe was recommended in 99.8% of calls. Fetching a brand page does not mean the model recommends that brand — the path from fetch to recommendation is observed, not mechanistic. The same study run against Claude (Anthropic) in the same window and category produced an own-site-fetch rate of 5.9% versus ChatGPT’s 38.2% — a sixfold gap. The finding describes ChatGPT’s behavior in skincare on day-0; other categories, models, and dates may differ.

The fetch goes to the page that holds the fact

Stop optimizing brand pages for generic recommendation queries. The open-ended phrasings — ‘best skincare brands,’ ‘top skincare brands’ — are lost causes for own-site fetches; those answers are built from editorial roundups and retail aggregators. Instead, publish the specific facts that constraint-driven questions demand: price tiers, ingredient suitability for skin conditions, format comparisons. A page that uniquely answers ‘best skincare brands under $30’ earns a fetch that a generic brand page never will. The goal is not visibility for its own sake; it is being the only source that holds the fact the question requires.

How we ran this

What we measured
1,444 (call, brand) recommendation events from 408 skincare queries sent to ChatGPT (8 question phrasings × 17 repeats × 3 batches). The brand set was eight real skincare brands plus two invented brands as controls. The registered metric is the share of recommendation events in which at least one fetched URL matched the recommended brand’s own domain (hostname matching, www. stripped), reported with a Wilson 95% confidence interval.
When
Data collected 18 and 20 July 2026 (day-0 window). Analysis published 20 July 2026.
Which AI
ChatGPT (OpenAI API) with live web search, Claude (Anthropic API) — one labeled comparison note, same window, Gemini — absent; its API does not expose fetched URLs

The full numbers

Every number in this piece, at full precision. The prose rounds for reading; this table doesn't.

MeasureAs shownExactSource
How often ChatGPT opened the brand's own site38.2%38.2271%Measured
Likely range, low end35.8%35.7551%Measured
Likely range, high end40.8%40.7616%Measured
Named with no own-site fetch61.8%61.7729%Derived
Named with no brand site fetched at all46.0%46.0526%Derived
Named with only a competitor's site fetched15.8%15.7203%Derived
Phrasing “best skincare brands” — fetch rate8.5%8.5202%Measured
Phrasing “for sensitive skin” — fetch rate10.2%10.2190%Measured
Phrasing “skincare brands worth buying” — fetch rate18.6%18.5930%Measured
Phrasing “top skincare brands” — fetch rate22.9%22.8700%Measured
Phrasing “rank the 5 best” — fetch rate37.9%37.8698%Measured
Phrasing “best affordable drugstore” — fetch rate67.1%67.0659%Measured
Phrasing “recommend 5 and say why” — fetch rate73.3%73.2620%Measured
Phrasing “best skincare brands under $30” — fetch rate84.9%84.8921%Measured
Third-party pages in the reading list88.6%88.6182%Measured
Own-brand pages in the reading list8.2%8.1589%Measured
Forum/social pages in the reading list1.7%1.6495%Measured
Reference pages in the reading list1.6%1.5734%Measured
L'Oréal — its own site fetched22.8%22.7941%Measured
L'Oréal — recommended3.2%3.1863%Measured
CeraVe — recommended99.8%99.7549%Measured
Big brands — own-site fetch rate50.9%50.9524%Measured
Mid-tier brands — own-site fetch rate33.0%33.0078%Measured
Claude — own-site fetch rate5.9%5.9419%Measured
Brand recommendations analyzed1,4441,444Measured
Queries per brand408408Measured

What this doesn't settle

  • One category, one day, one leg — the registered rate is a finding for skincare on the OpenAI API on these dates, not a generalizable number across categories, engines, or time.
  • Correlation is not cause — the data shows which pages were fetched alongside a recommendation, not why the recommendation was made. The path from fetch to name-check is observed, not mechanistic.
  • Zero-retrieval name-checks have three live explanations (training memory, unlogged context, third-party authorship); this data cannot separate them, though the source mix favors the third.
  • These queries ran against the OpenAI API with live web search — not the consumer ChatGPT product. The API has no personalisation, no conversation memory, and no geo-targeting.

Is your brand on the pages ChatGPT reads before it names you?

A diagnosis maps which third-party sources — editorial, clinical, retail — are shaping recommendations in your category, and whether your brand appears on them when the model assembles its answer. Not a score: the evidence, run twice.

See what a diagnosis finds