AI shopping index: which brands ChatGPT and Perplexity recommend

September 2026 at a glance

Collected 25 September 2026. 36 questions across 12 categories went to each assistant; 72 of 72 answers came back complete. Counts below are the number of answers that named a brand or cited a site, out of the answers that category received. Read them as one month’s observations, not rankings.

The tables list the ten most-named brands per assistant.

Data and reuse

• brands.csv: every brand listed above, with its rank and answer count per assistant and category.
• sources.csv: the most-cited domains per assistant and category.

You can reuse the data; cite Citelift and the collection month. Raw answer text stays private because it is third-party content.

Method and limits

• Questions. Twelve DTC categories, two for each of beauty, supplements, apparel, home, pet and food. Three buyer questions per category, asked unchanged in each edition, once each per assistant.
• Collection. Through the DataForSEO API. ChatGPT answers come from DataForSEO’s scraper of the ChatGPT web interface, from a US location with web search on. Perplexity answers come from its Sonar model over the API, with US web search.
• API answers are not app answers. These are single-turn answers collected by API. They are not what a signed-in shopper sees in the ChatGPT or Perplexity apps, where memory, location, personalisation and model settings differ.
• Counting. A brand counts when an answer recommends it by name. Where an answer tags brands (a few ChatGPT answers do), we use those tags; otherwise we take the names the answer sets in bold, match them against brands tagged elsewhere that month, and ask a classification model whether each remaining name is a brand or a heading. Names it cannot judge are left out. A product line counts as its brand when the same category’s answers also name that brand on its own, so “Naturium Vitamin C Complex Serum” counts as Naturium; a product named without its brand stays as named, and a line that lists several brands is left out. Each answer counts a brand once. Cited sites are the domains of the links each answer showed.
• Minimum answers. An assistant with fewer than 3 successful answers in a category is shown as insufficient data, with no lists.
• Variation. Answers change between runs. Our repeatability pilot found no identical source sets across 48 repeated answers, so treat each month as one observation.
• No commercial relationship. Citelift does not sell the products named.

This index follows our 23 September 2026 study of ChatGPT and Claude in six DTC categories, which used consumer apps by hand rather than an API.