Websites in AI shopping answers: OpenAI's links vs Perplexity's sources

If you want to know which websites appear in the sources of an AI shopping answer, the first problem is that there is no single list. Each engine hands back its sources in its own way, and a table that adds them together mostly describes whichever engine hands back more. Below are the two October 2026 inventories, one per engine, each with its own denominator: 800 answers from OpenAI's API with the web search tool switched on (model gpt-6-luna), and 787 from Perplexity's Agent API with the Sonar model (perplexity/sonar).

What exactly is in each list?

For OpenAI, the list is the pages its answer links. For Perplexity, it is the pages its search retrieved, and the API does not say which of them the answer relies on. So an OpenAI domain is one the answer pointed a reader to, and a Perplexity domain is one the search pulled in. We reduce every URL to its domain and count answers that contain the domain, once per answer, however many of its pages appear.

The shares below are the percentage of an engine's answers with at least one source (99.6% of OpenAI's answers and 100% of Perplexity's). The lists come from prompts we wrote across 160 categories in the United States, Mexico and Canada, so they include country domains such as amazon.ca and liverpool.com.mx. This post is the inventory; the methodology explains how the recording works.

Which domains does OpenAI link most often?

OpenAI links about two pages per answer (a mean of 2.2 links and 2.0 distinct domains), so its lists are short and its most frequent domains sit at low shares.

Of OpenAI's top 30 domains in this edition, only 2 reach a 5% share, and the 30th place is a tie at 7 answers (0.9%), nike.com among the tied domains. Its list also holds domains that match brand names, among them ikea.com, samsung.com, apple.com, breville.com and levi.com, which is the kind of match the strict own-site count below picks up. Nothing in these 800 answers links nytimes.com.

Which domains does Perplexity retrieve most often?

Perplexity's lists are about twelve times longer: a mean of 27.3 URLs and 15.2 distinct domains per answer. A domain therefore has far more chances to appear, and the shares are much higher.

All 558 nytimes.com URLs in these lists were on Wirecutter pages; the Wirecutter post looks at that one domain in detail. Among Perplexity's top 30 domains, 15 reach a 10% share, and the 30th place is a tie at 46 answers (5.8%), walmart.com.mx among the tied domains. OpenAI's 30th place is a tie at 7 answers, so the two curves have very different shapes: OpenAI's thins out almost at once, Perplexity's stays broad.

Which domains appear in both lists?

Eleven domains are in both engines' top 30: goodhousekeeping.com, rtings.com, target.com, rei.com, wired.com, tomsguide.com, nbcnews.com, outdoorgearlab.com, liverpool.com.mx, techradar.com and walmart.com.mx. Because the 30th place is a tie in both lists, the count of eleven depends on where the cut falls: walmart.com.mx is one of several Perplexity domains tied at 46 answers. The shares are not comparable face to face.

For all eleven shared domains, Perplexity's share is roughly 2 to 10 times OpenAI's. This may reflect list length as much as the domains; we did not test which. So compare a domain's place within each engine's list rather than its raw share across engines. A domain that is first in one list and ninth in the other, as rtings.com is, is a regular presence in both.

What kinds of websites are they?

Our data classifies source lists three ways. These are our definitions, and an answer can fall into more than one group, so the bars do not add up. This edition does not classify editorial review sites as a type: no verified, hand-made list of them exists in our data, so we make no claim about their share beyond the named domains in the tables above.

• Retailer: the list holds a domain from a short hand-made list of 49 retailers and marketplaces, so stores not on it (shopping.yahoo.com, for example, which is in 10.4% of Perplexity's answers) are not counted and the share undercounts.
• Forum or video: the list holds one of six hand-picked forum, video and social domains (reddit.com, youtube.com, quora.com, tiktok.com, instagram.com, facebook.com), so it undercounts too.
• Own site (strict match): the list holds a domain whose first label equals a recommended brand's name with spaces and punctuation removed, such as samsung.com for Samsung. It misses country domains, parent companies and other store names, so it undercounts.

The order flips between engines. A retailer domain is in Perplexity's retrieved list in 76.7% of answers and in OpenAI's links in 17.6%. A strict own-site match is in OpenAI's links in 46.6% of answers and in Perplexity's list in 62.0%. A longer list is more likely to contain any given kind of domain, which may explain part of the gap; we did not test how much. The retail and own-site cases get their own posts later in the series; this one only shows where they sit among all domains.

How should you read an inventory like this?

• Presence is not influence. A domain in a list means a page was linked or retrieved. It does not show that the answer used it or that it changed which brands were recommended.
• Categories drive the lists. rei.com appears in the lists for 23 categories in OpenAI's answers and 39 in Perplexity's, not across all 160. Which domains show up depends on what you sell.
• The lists are one month and one prompt set. In the October 2026 edition each prompt was asked once per engine, so small gaps between low-share domains are not meaningful.
• API is not app. The consumer apps can show different sources, and the sample is not a copy of them.

A preprint by Chen and Koudas (arXiv:2610.15033) queried OpenAI's API (gpt-5.4 nano) with web search turned on, using other questions (651 queries written for a US fiber internet provider). It reported a similarly small number of linked pages: 2.0 distinct pages per answer. Different questions and a different model, similar order of magnitude.

How can you build the same inventory for your category?

1. List the questions a shopper in your category asks: the best product overall, the best under a price limit, a premium pick, a need-specific question, and a plain buying-advice question.
2. Ask each engine each question, with the same location and a fresh session, several times on different days. Record the sources the way the engine reports them.
3. Reduce each URL to its domain and count the answers that contain it, not the URLs.
4. Keep a table per engine with its own number of answers. Do not merge them.
5. Mark the domains where a page could be checked for accuracy: your own site, retailers that stock you, and publishers that review your category. Appearing in a list does not show a page shaped the answer. Our free manual method covers the wording and sample size.

Each US category page in the index lists the domains for that category engine by engine; air purifiers is a category that keeps the October 2026 edition. The monthly figures behind this post are in the October report.