Buyers now ask ChatGPT what to buy. If you are not in the answer, you are not being discovered.

Discovery has moved, and most brands have not followed it

People are asking ChatGPT what to buy. On 27 August 2026 OpenAI announced that ads are coming to ChatGPT’s free and Go tiers in India, after a US rollout in February. A company does not sell ads on a page nobody shops from. If a buyer asks ChatGPT which sunscreen to buy and your brand is not in the answer, that buyer never hears your name. That is the discovery problem now, and no dashboard you own reports it.

So we measured it the way we measure a shelf. We asked the ten questions a sunscreen buyer in India types, five times each, in fresh sessions with no account and no memory. Fifty answers. We counted which brands were named, how often, and which pages the model read to decide. This measures what the model says on one day. It is not demand, not traffic and not sales.

Then we set it beside the one shelf where sales can be measured. Amazon India publishes a sales signal on each product, and we read its sunscreen shelf on 19 August: 64 brands, a revenue floor of 197 crore rupees a year. So for every brand we hold two numbers: the share of ChatGPT answers that name it, and its share of Amazon revenue. They are different units, placed side by side and never added.

Who ChatGPT names

Re'equil was named in 32 of 50 answers and Minimalist in 30. Together they hold 12.7 percent of Amazon revenue. Dot & Key holds 9.1 percent of Amazon revenue, more than any other brand, and was named in 12 answers. The model names a different set of brands from the set that sells.

Shaded rows are Indian D2C brands. Amazon revenue share is from our 19 August read of the Amazon India sunscreen shelf, a floor built from the bought-in-past-month badge.

By owner group the answers track the shelf. By brand they do not.

Indian D2C brands take 56.1 percent of all brand mentions and hold 51.5 percent of Amazon revenue. Pharma brands take 27.1 percent of mentions on 19.7 percent of revenue. Multinationals take 8.9 percent on 15.4 percent. The group totals are close. Inside the groups they are not. Re'equil and Minimalist carry most of the D2C mentions on 12.7 percent of the revenue. Fixderma carries the largest share of the pharma mentions, named in 21 answers, on 2.7 percent of revenue.

The ten largest brands on Amazon, and how often ChatGPT names them

Which brand answers which question

Each question was asked five times. A brand is listed when it was named in three or more of the five answers. Where the list says none, the model gave a different set of brands each time.

The same question asked five times returned 4.3 brands per answer on average, between 6 and 16 distinct brands across the five runs, and a core of 0, 1 or 2 brands named every time. One answer is one sample. Any claim that ChatGPT recommends a brand needs a count and a date.

Where ChatGPT reads

48 of 50 answers searched the web and cited pages, 132 citations in all. nykaa.com is the most cited single site, in 20 answers. But as a class, brand-owned sites are cited 40 times, marketplaces and pharmacies 42, and publishers 29. The model reads the brand’s own product page. The two most-named brands are also the two most-cited brand sites.

43 of 50 answers also showed product cards with a price, a rating and a store. 158 cards in all, median price 395 rupees. 33 of them linked to Amazon.in. The rest linked to Purplle, Nykaa, Myntra and, in 15 cases, to a brand’s own store.

What we predicted before the read, and what happened

We wrote six predictions down before the first question was asked. Two were right, two were wrong, two were half right. We predicted Indian D2C would take the majority of mentions and over-index against its Amazon share: partly right, 56.1 percent of mentions on 51.5 percent of revenue. We predicted pharma brands would be under-named: wrong, they over-indexed, and Fixderma was named in 21 answers on 2.7 percent of revenue. We predicted a login wall inside ten sessions: wrong, there was none in fifty. We predicted the sources would be marketplaces and blogs with no manufacturer site near the top: half right, because brand-owned sites turned out to be the second largest class.

Method, and what this is not

Ten buyer questions define this read. They are listed in the table above. None names a brand. Each was asked five times in a fresh browser session on 11 September 2026, logged out, on the free product, from a residential connection in India, with web search left at its default. Change the questions and you have a different read.

A brand is named when its name appears anywhere in the answer, once per answer. Names were matched against the 126 brands on the four Indian sunscreen shelves we hold plus an alias list, and every name the list missed was reviewed by hand. UV Doux is counted as Brinton, its maker, because the model names the product line and Amazon names the company.

This is not the answer a signed-in user sees. A signed-in answer is shaped by that person’s history and has no single value to record. It is not a quality ranking. A brand named often is a brand the model found on pages it could read. And model answers change between runs and between days, so every figure here is dated, and a re-run is a new measurement.

Find out where your brand stands

Buyers are moving part of their product search into chatbots, and the chatbots are now selling ads against those searches. If you are not in the answer, you are not being discovered there. This read covers one category, one assistant, one day. The same read runs on your category, on ChatGPT and the other assistants, and beside the shelves you sell on: who is named, how often, from which pages, and where you sit.

If you want to know how visible your brand is in your product segment, across the assistants and the platforms you sell on, get in touch through the enterprise page. Tell us the category. We reply with what a read would cover.

Related measurements

The full report, with every table and the prediction record, is free to download as a PDF. The Amazon shelf it is set against is measured here. The full library is at impuls8.com/reports.