Ask ChatGPT for the best cleanser for oily skin and the answer mentions products. It also cites pages as its sources. In one Zumi category run, none of the ten most-cited sites belonged to a beauty brand.
Most guides on getting products recommended by ChatGPT cover the same setup work. That work matters, but it only makes a product eligible. This guide covers the step most of them skip: reading which pages the answers actually cite, and deciding which of those a brand can change.
Key takeaways
- Crawler access, a consistent product feed, and matching page markup make a product eligible to be mentioned. No setup step buys a mention, and no tactic guarantees one.
- In Zumi's India beauty question run, 8,740 answers carried 40,304 citations across 2,443 domains, and none of the ten most-cited domains was a beauty brand's own site.
- Those ten were social and video platforms, retailers and marketplaces, and editorial or health publishers.
- A working method fits in an afternoon: list the pages cited for real buyer questions, sort them by owner, and split them into pages a brand can influence and pages it can only measure.
- A product that is mentioned is not necessarily the one the answer picks, so position gets read alongside mentions.
What does the setup work cover?
Four things come up in nearly every published guide. OpenAI's search crawler must not be blocked in robots.txt, and the product feed, the product page, and its markup (the code that labels a page's price, stock, and product details for machines) must agree.
Titles and descriptions should use the words buyers use. Written reviews on and off the site give the answer something to verify against.
All of it is worth doing once and checking occasionally. None of it shows where a recommendation comes from, and for that the cited pages are the evidence.
How are the cited pages listed?
The list starts from questions a buyer would type, not the brand's product names. "Best cleanser for oily skin under 500 rupees" is a shopping question, and "Brand X cleanser" is not. How to choose prompts to track covers building that set.
Each question is run in ChatGPT, then in the other engines that matter to the category, with every mentioned product and every cited page recorded. Answers change from run to run, so each question is repeated before any pattern is trusted. How to track AI citations explains the repetition and the logging.
The result is a list of pages. A product list is a by-product, and the pages are the finding.
Which of the cited pages can a brand move?
Sorting each cited page by who owns it turns the list into a plan. Six groups cover almost everything.
The brand's own site. The brand can change these pages directly: accurate product pages, a feed and markup that agree, and plain descriptions.
Retailers and marketplaces. The brand can correct its own listings, titles, images, and review prompts. Ranking on the retailer stays outside its control.
Review sites and comparison lists. Inclusion is earned with accurate facts, sample products, and honest review requests. This is influence, not control.
Forums and communities. These are measured. Real customers and staff who disclose their role can take part, but a thread cannot be bought or edited.
Editorial and health publishers. Accurate facts can be supplied to journalists and editors. Coverage cannot be requested.
Social and video platforms. The brand can work with creators who actually use the product, and measure what the answers pick up.
Only the first group is fully in the brand's hands. The other five are influenced or measured, so a plan limited to the brand's own site leaves five of the six groups untouched.
Retailer pages appear in this list as cited sources. Zumi tracks a brand's own products and domains, not retailer product pages, so the retailer group is about what the answers cite, not what the brand tracks.
What does one category run show?
Zumi asked 304 questions about beauty and personal care in India every day from 23 July to 2 August 2026. Nearly all of the 8,740 answers came from ChatGPT, Perplexity, and Google AI Overviews, and they carried 40,304 citations across 2,443 domains. The ten most-cited domains took 28.9% of all citations.
None of those ten was a beauty brand's own site. They were Instagram, YouTube, Nykaa, Reddit, NBC News, the American Academy of Dermatology, Amazon India, Allure, Smytten, and Healthline.
That is one category, one market, and one run, spread across three engines rather than ChatGPT alone. A different category will split differently, and a brand should not read these figures as its own. The run shows what a source list can look like when buyer questions are asked at scale.
It also means a brand's own domain should not be assumed to be among the pages cited. The full run shows how concentrated the citations were.
What gets read next to the cited pages?
Two checks go beside the list. The first is mention versus pick.
An answer that mentions a product in the fourth sentence of a long list has not recommended it, so average position is read next to mention rate, the share of answers that mention the product. The four signals Zumi reports are described in geo metrics beyond mention rate.
The second is the count itself. One product that appears under four names reads as four weak products, which hides the real mention rate. Product visibility in AI answers shows how those counts break and how to merge them.
Neither check connects to sales. What the answers say and what shoppers later do are separate measurements, and AI search attribution explains why no figure should be calculated across them.
What should be done this week?
The first step is ten shopping questions in buyer language, each run several times, with the cited pages sorted into the six groups above. The group holding the most cited pages, and the pages in it that a brand can actually change, is where effort goes first.
Zumi tracks product mentions and the pages cited for the same shopping questions across engines. Book a demo.
Sources
- Zumi, India beauty and personal care question run: 304 questions asked daily on ChatGPT, Perplexity, and Google AI Overviews, 23 July to 2 August 2026. An industry run with no client data. Figures recorded in Zumi's claims register.