Which Brands Get Recommended? 540 Answers From OpenAI, Perplexity and Claude, Counted

TL;DR: On November 9, 2026, we asked OpenAI, Perplexity and Claude 20 buyer questions in three categories (water filters, sheets and bedding, and standing desks and office chairs), three times each, through their APIs with web search on. That's 180 answers per category, 540 in all. Four things stood out. The engines disagree: Saatva was named in 31 Claude answers and no OpenAI answers. Claude named two to three times as many brands per answer as the other two. Each category leans on a short list of sources, led by nsf.org for water filters, sleepfoundation.org for bedding and techradar.com for office furniture. And in two of the three categories, the most-named brands' own sites were among the ten most-cited.

Most brands check whether AI recommends them the same way: type one question into one app and read one reply. That's one sample of an answer that changes every time you ask.

So we counted instead. I built EcomIndex to do this: it asks a fixed set of buyer questions, several times, on several engines, and counts which brands come back. On November 9 it ran three categories for its AI Shelf Index. Here's what the counts show, and what a Shopify operator can do with them this week.

How we counted

• 20 buyer questions per category, written the way shoppers ask: best-for questions, use cases, comparisons, price limits and alternatives.
• Each question asked 3 times on each engine: OpenAI (gpt-6.1-sol), Perplexity (perplexity/sonar on its Agent API) and Claude (claude-sonnet-5-5), all through their APIs with web search on, searching from the US. Every one of the 540 answers completed.
• A brand counts as named when the answer recommends it or lists it as an option to buy. A brand mentioned only negatively doesn't count. Stores such as Target and Amazon are counted separately, as places the answers send shoppers to buy.
• Citations come from each engine's own citation list, counted once per answer by site.

Two limits. These are API answers from one day, not a record of what a shopper sees in the ChatGPT app, and answers change from run to run. And the order is a count, not a rating: nothing here says a product is good, and no brand paid to be listed.

1. The three engines don't agree

The same 20 questions, the same day, and some brands live on one engine only. Each engine gave 60 answers per category.

If L.L.Bean checked only Claude, it would conclude AI barely knows its bedding. If Saatva checked only OpenAI, it would conclude the same thing. Both would be reading one engine and calling it "AI".

Part of the gap is how much each engine says. Claude named 6.9 to 9.1 brands per answer across the three categories. OpenAI named 2.6 to 3.0, and Perplexity 2.3 to 3.1. A brand that makes Claude's long lists can still miss the short ones.

2. A few brands take most of the mentions

The five most-named brands in each category, out of 180 answers:

Behind them is a long tail: 94 brands were named at least once for water filters, 123 for bedding and 80 for office furniture. Plenty of brands get one mention. All fifteen brands above were named on all three engines, dozens of times each.

3. Each category has its own short list of sources

The answers cited a lot of sites, from 168 to 209 per category. A handful came back again and again.

Water filter answers lean on the certification body and government sites. Bedding answers lean on a sleep foundation and big publishers. Office answers lean on tech review sites and osha.gov.

Those sites are where the engines look before they decide who to name. For a water filter brand, that makes your NSF listing, and the certification stated plainly on your product pages, worth checking before your next blog post. For a bedding brand, it points at the review and testing coverage those publishers do.

4. Brand sites get cited too

In water filters, aquasana.com was cited in 23 answers, the fifth most-cited site, and Aquasana was the most-named brand. In office furniture, upliftdesk.com (26 answers), branchfurniture.com (24) and flexispot.com (18) were all in the top ten sources, and Branch, FlexiSpot and UPLIFT were three of the four most-named brands. No brand's own site made bedding's top ten.

That doesn't prove the citations caused the mentions. It does show the engines read brand sites and cite them, which makes your product and collection pages one source you control.

5. Where the answers send shoppers to buy

Retailers were counted separately from brands. Target was named as a place to buy in 36 bedding answers, and Threshold, Target's house brand, was named in 37. IKEA came up in 6 office answers. Amazon was named as a place to buy in 5 of 180 bedding answers and 1 of 180 water filter answers.

What to do with this

1. Ask on all three engines, more than once. Write 10 to 20 questions a buyer in your category would ask. Ask each a few times on ChatGPT, Perplexity and Claude, and tally who gets named. One reply from one app tells you very little.
2. List the sites the answers cite. The five or so that keep coming back are the ones to work on: get listed, get reviewed, or get your certification on record there.
3. Make your product pages state what those sources check. Certifications, materials, dimensions, warranty terms. Engines repeat facts they can find on pages they read.
4. Check that AI crawlers can reach your store. EcomIndex's free AI crawler check reads your robots.txt the way each crawler does and shows the line that lets it in or keeps it out.
5. If you own the DTC Stack, /geo-engine covers the feed, schema and content changes. Point it at the sources from step 2.

For more on how AI shopping agents read your pages, see AI Shopping Agents Are Reading Your Product Pages and the Shopify agentic commerce field guide.

The full lists

Every brand, per-engine count and cited site for these three categories is on the AI Shelf Index: water filters, sheets and bedding and standing desks and office chairs. Each page lists the 20 questions and the method.

If you want the same count for your own store and the competitors you name, that's the EcomIndex AI Shelf Report. Disclosure: EcomIndex is mine.

FAQ

Do these counts show what shoppers see in the ChatGPT app?

No. They come from the engines' APIs with web search on, on November 9, 2026. The ChatGPT app and the API can give different answers, and both change from run to run. The counts show how often each brand was named across 180 answers per category, which says more than any single reply.

Why do OpenAI, Perplexity and Claude name different brands?

Each engine searches the web its own way, cites different pages and writes answers of different lengths. In these runs Claude named 6.9 to 9.1 brands per answer, against 2.3 to 3.1 for OpenAI and Perplexity, so a brand can show up often in Claude's answers and rarely in the others.

Did any brand pay to be listed?

No. The lists are counts of completed answers. No brand paid to be included, and the order doesn't rate any product.

How often should a brand check?

Answers change as models and the web change, so one check goes stale. Re-running the same questions every month or quarter shows what moved. Small changes between runs are often noise, so look for moves that hold across several questions and engines.