"Reviews" means three different things to an AI shopping tool, and mixing them up is the main way people get this question wrong. There are customer reviews on retailer and brand pages, editorial reviews by publishers that test products, and rating markup that a store adds to its own pages. Our data speaks only to the second kind: which publisher domains, and which retailer domains where customer reviews live, show up in each engine's source lists. It says nothing about rating markup, and nothing about whether any of these sources changes which brand gets named.
Which review sources show up in each engine's lists?
We sampled 800 answers from OpenAI's API with web search (gpt-6-luna) and 787 from Perplexity's Agent API with the Sonar model (perplexity/sonar), across 160 categories in the US, Mexico and Canada. OpenAI's source list is the pages its answer links. Perplexity's is the pages its search retrieved, which the answer may or may not use. Because the two lists are recorded differently, each engine is read against its own answers.
We sorted the 30 most frequent domains of each engine's list by hand into four groups: publishers (review sites plus news and lifestyle publishers), retailers, brand sites, and organizations that publish guidance. A hand sort is a judgment call, and a domain such as rei.com could sit in more than one group.
The rest of the Perplexity list is retailers: 11 of its 30 domains are Amazon in three countries, Target, Best Buy in two, REI, Chewy, Klarna, and the Mexican retailers Liverpool and Walmart Mexico. OpenAI's other 17 are 6 retailers, 6 brand sites and 5 organizations or government sites. The lists mix three markets, so Canadian and Mexican domains are in them.
Seven publisher domains are in both lists. The shares differ a lot, and neither engine's number is a statement about the other.
| Domain | OpenAI answers linking it | Share of 797 answers with a link | Perplexity answers retrieving it | Share of 787 answers |
|---|---|---|---|---|
| rtings.com | 47 | 5.9% | 109 | 13.9% |
| tomsguide.com | 35 | 4.4% | 80 | 10.2% |
| wired.com | 26 | 3.3% | 90 | 11.4% |
| outdoorgearlab.com | 23 | 2.9% | 77 | 9.8% |
| goodhousekeeping.com | 20 | 2.5% | 133 | 16.9% |
| techradar.com | 20 | 2.5% | 50 | 6.4% |
| nbcnews.com | 8 | 1.0% | 78 | 9.9% |
October 2026 edition. OpenAI: pages linked in the answer (gpt-6-luna), 797 answers with at least one link. Perplexity: pages its search retrieved (perplexity/sonar), 787 answers. Each engine on its own denominator.
Two other facts bear on customer reviews, not editorial ones. First, retailer pages are one place customer reviews live, and the retail listings post goes deeper on them. Perplexity's retrieved lists held a retailer domain in 76.7% of its answers against 17.6% for OpenAI's linked pages, on a hand-made list of 49 retailer domains. The amazon.com domain alone was in the retrieved sources of 265 of 787 Perplexity answers (33.7%), across the US, Mexico and Canada runs. Second, no site whose main business is hosting customer reviews appears among either engine's 30 most frequent domains. Reddit appeared in the links of 6 of 800 OpenAI answers and in the retrieved sources of 26 of 787 Perplexity answers; the Reddit and YouTube post covers that in detail.
Presence is not influence. A review site in a list shows that the engine linked or retrieved it, not that its score or text decided the recommendation.
What do the platforms say about reviews?
Three public statements, as of October 2026:
- OpenAI says that when ChatGPT shops for you it may consider options, price, reviews and ease of use, and that it uses structured metadata from first-party and third-party providers. It also says product results are selected independently and are not ads. That page describes the ChatGPT app; our numbers come from the API, which can differ.
- Amazon, in its announcement of its shopping assistant Rufus, said answers draw on product listing details, customer reviews and community questions and answers. So on Amazon's own assistant, customer reviews are a stated input. Our data does not measure Rufus.
- Google does not ask for special markup for its AI features, and its review snippet rules say not to aggregate reviews or ratings from other websites and not to include fake or undisclosed incentivized reviews in the page or its markup.
None of the three gives a weight. We also found no published statement from Perplexity, or from any assistant, that ties a star rating or a review count to a recommendation. That is the honest summary: reviews are named as an input in places, and nobody publishes how much they count.
What do we not know?
- Whether answers use rating values. Our pack records source domains and brand names, not whether an answer quoted a score or a review count. We cannot say how often any answer did.
- Whether review volume changes who is named. We ran no experiment. The data is one October sample of about 10 answers per category, described in the methodology.
- How the apps behave. The API and the apps can differ in personalization, location and product features.
- Which reviews count. Nothing public separates a retailer review, a publisher's test and a forum thread for the purpose of a recommendation.
How do you earn reviews within the rules?
In the US, the FTC's rule on consumer reviews and testimonials has been in effect since October 21, 2024, and the FTC says it authorizes courts to impose civil penalties for knowing violations. The FTC's own question-and-answer guidance gives practical limits. This is information, not legal advice, and it covers the US only.
- Ask every buyer. The FTC's guidance says the rule exempts generalized solicitations to purchasers to post reviews about their experiences. It also says asking for reviews only from customers you think are happy is not specifically prohibited by the rule, but could violate the FTC Act. Send the same request to everyone who bought.
- Do not tie a reward to a sentiment. The rule covers compensation or incentives conditioned on reviews expressing a particular sentiment, and the FTC says that can be implied. Its example is an offer of a coupon for telling the company how much you loved a visit. The FTC says the rule does not prohibit incentives for reviews, but failing to disclose them could violate the FTC Act, so make sure any reward is disclosed and is offered for a review of any kind. Marketplaces may ban incentives outright (see item 6).
- Disclose free products. If a reviewer got your product free or at a discount, tell them to disclose that wherever they post about it, as the FTC's guidance advises; put that in your reviewer brief.
- Do not suppress. The FTC says review suppression can create liability under section 465.7. Its Q&A says organizing reviews is not suppressing them, but organizing them so negative reviews are hard to find could be unfair or deceptive. The Q&A says a review-exclusion policy can be fine if it treats positive and negative reviews equally. It does not mention spam, so read section 465.7 before you write a policy.
- Do not run a fake review site. Do not set up a site you control and present it as an independent reviewer of products, your own included; the rule bars that misrepresentation.
- Mind each marketplace. In an October 2016 announcement, Amazon said it updated its guidelines to prohibit incentivized reviews unless facilitated through its Vine program. That is an old announcement, so read Amazon's current Community Guidelines before you act; other retailers have their own rules.
- Keep rating markup honest. Google's review snippet rules say the review content you mark up must be readily available to users on the marked-up page. Mark up only the reviews that appear on that page, from real customers.
Where to put effort first is unknown, so spread it across the places a shopper reads reviews: your product pages, the retailers that carry you, and the publishers that test products. A follow-up post on getting into the publishers' product tests is planned for late November.
How would you test whether reviews move your answers?
You need before and after samples, a comparison and enough answers. A worked example, computed with 95% Wilson intervals:
- Baseline. Ask fixed prompts for your category in each engine, in clean sessions, and record how many answers name your brand. Say 4 of 40: 10.0%, with an interval of 4.0% to 23.1%.
- Change one thing. For example, a sustained effort to collect reviews on the retailer page, with nothing else changing.
- Resample later. Say 8 of 40: 20.0%, with an interval of 10.5% to 34.8%. The share doubled, but the intervals overlap, so this alone cannot tell you the change was real.
- Add a comparison. Track a similar product you did not touch. If its share moved the same way, something shared by both products, such as a model or retrieval change, is a more likely explanation than your reviews.
- Account for everything else. Models, retrieval and your competitors' work all change between samples.
The share of answer post explains the arithmetic, and how to measure AI visibility gives the manual routine.
Frequently asked questions
Should reviews live on my site or on retailer pages?
The data cannot say. Perplexity's retrieved sources held a retailer domain in 76.7% of answers, and OpenAI's links held one in 17.6%, so retailer pages are a regular source in one engine and an occasional one in the other. Amazon names customer reviews as an input to its own assistant. Collect reviews where your customers buy, and keep the same facts on your own site.
Do negative reviews keep a brand out of AI answers?
We do not know. Nothing in our data or in the platform statements we found links negative reviews to being left out of an answer. The FTC says review suppression can create liability under section 465.7, though.
How we write: drafted with AI tools from Dopbot Index data and public sources; every number and outside fact is traced to its source in a claims record and checked before publishing. More about how we write; corrections to hello@dopbot.com.