Open ChatGPT and ask it to pick the best running shoes, the best CRM, or the best pair of glasses under 50 dollars. Read the answer closely and notice what it is built from. There will be a brand mentioned, often several. But the substance of the recommendation, the reasons one option beats another, almost never comes from the brand's own website. It comes from a review, a comparison table, a buying guide written by someone the brand does not employ and does not pay.
That is not a quirk of one query. It is the default behavior of AI product discovery, and it is the opposite of what most brands assumed would happen. The intuitive expectation was that a brand with a polished site, strong product pages, and good SEO would be the source an assistant quotes. Instead, the assistant treats the brand's own description of itself as one of the least useful things in the room.
This post is about why that happens. Not how to fix it, that is a separate craft with its own playbook, but the underlying mechanics: what AI assistants actually read when they recommend a product, how small a slice of that reading is brand-owned, and what it means that the affiliate and editorial review layer has quietly become the raw material of the most important new discovery surface since Google.
How product discovery got rerouted
For two decades, finding a product to buy followed a stable shape. You searched, you got ten blue links, you opened a few, you read a couple of reviews, you decided. The brand's site sat in that list. So did review sites, comparison sites, and the affiliate publishers whose entire business was ranking a "best of" page and earning a commission when you clicked through and bought.
That shape is collapsing into a single answer. McKinsey's research on AI search found that across major consumer categories, electronics, grocery, travel, wellness, apparel, beauty, and financial services, somewhere between 40 and 55 percent of consumers now use AI-powered search to make purchase decisions. The same work found that 44 percent of AI search users call it their primary and preferred source of buying insight, ahead of traditional search at 31 percent, ahead of retailer and brand websites at 9 percent, and ahead of review sites at 6 percent. The brand's own site, as a place a buyer goes to decide, has fallen behind the assistant that summarizes everything else.
When the answer arrives as a synthesis instead of a list, one question becomes everything: what did the model read to write it. Because whatever the model read is what now shapes the purchase. And the answer to that question is the uncomfortable part for brands.
What the assistant is actually reading
Several independent studies, using different methods and different data, converge on the same finding. AI assistants recommend products mostly from third-party content, and brand-owned domains are a small minority of the sources.
The cleanest single illustration comes from eyewear. Max Willens, a senior analyst at EMARKETER, shared a study finding in an October 2025 webinar: when large language models discussed the eyewear brand Zenni, roughly 70 percent of the content the models drew on was affiliate content. Not Zenni's website. Not Zenni's product pages. The reviews, roundups, and comparison articles that publishers wrote about Zenni in order to earn affiliate commissions.
Zenni is one brand, but the pattern holds at scale. McKinsey's analysis, cited by Adam Weiss, president of North America at the affiliate network Awin, put a number on it: brand site content makes up only about 5 to 10 percent of the sources AI search references. The other 90 percent or so is the distributed web of reviews, comparison guides, buying recommendations, expert analysis, and user discussion that brands do not own. Semrush, studying how AI tools cite brands, found the same shape from the other direction: AI systems lean on independent third-party pages over a brand's own domain, because a self-description is not independent evidence and an outside review is.
Look at the most-cited content type and the reason starts to show. Evertune analyzed roughly 25,000 unique URLs, the most-cited pages across ChatGPT, Copilot, Gemini, Google AI Mode, Google AI Overviews, and Perplexity. Half of those URLs were listicles. Across close to 400 million citations, 63 percent pointed to listicles, and the great majority of those were ranked lists: "best 7 of something," scored and ordered. Corporate sites, earned media, and affiliate domains were the top sources of those lists. The structured "best of" article, the native format of affiliate publishing, is the single thing AI search reaches for most.
The review-site layer shows up just as clearly in software. An SE Ranking study of 30,000 commercial keywords, analyzing nearly 23,000 Google AI Overviews in December 2025, found that three of the five most-cited domains in that whole set were review platforms: Gartner Peer Insights, G2, and Capterra. Review platforms were only 8.5 percent of all links by volume, but they punched far above that weight in visibility, and on explicit review queries nearly half of all AI Overviews cited at least one of them. The brand being evaluated was rarely the source of the evaluation.
One caveat keeps this honest. The numbers move depending on how a study defines a source and what it measures. Some vendor research, counting a brand's website plus its directory listings plus its review-platform profiles all as "brand-influenced," reports that a large majority of citations sit within a brand's sphere of influence. That is a different question. A brand can influence a G2 profile; it does not write the comparison article that quotes that profile. Strip the definition back to what a brand actually authors and controls, its own site, and the figure lands in McKinsey's single-digit range. The disagreement is about labels. The underlying fact, that AI leans on outside content to recommend, is not in dispute.
Why models extract review content and skip the brand
This is not the assistants picking favorites. It falls out of how language models work and what kind of text answers a shopper's question.
A model generating a recommendation is doing extraction and synthesis. It needs text that already contains a comparison, a judgment, a ranking, a reason one option suits a need better than another. A brand's product page, by its nature, does not contain that. It describes one product in the most flattering available terms. It has no second option to weigh it against, no scenario where the product is the wrong choice, no independent verdict. It is an advertisement, and the model treats it accordingly: useful for a specification, close to useless for a recommendation.
A good affiliate review or comparison article is the inverse. It holds several products side by side. It states who each one is for. It assigns scores, names trade-offs, and reaches a conclusion. That is exactly the shape of the answer the assistant is trying to produce, which makes it the easiest, highest-value text to lift from. The publisher did the comparative reasoning already; the model just has to compress it.
Structure compounds the effect. A ranked listicle is tightly scoped to one question, "the best X," with a parseable format: a heading, a verdict, a short rationale, repeated. That is close to ideal input for a system that has to read fast and reproduce reliably. The Evertune finding that listicles draw the majority of citations is not a mystery once you see it from the model's side. The format that affiliate publishers refined over fifteen years to win Google rankings turns out to be the format that machines extract most cleanly.
There is a trust dimension too, and it predates AI. Shoppers have always discounted what a brand says about itself and trusted what an apparent third party says. Models trained on human text, and tuned with human feedback, inherit that prior. Independent-looking review and comparison content reads as more credible, so it is weighted as a better source. The assistant is reproducing a bias the open web already had. It is just doing it inside a single answer where the brand can no longer buy its way to the top of the list.
What this means for brands, publishers, and the channel
Three groups should read this finding differently, because it lands on each of them in a different way.
For brands, the hard message is a loss of narrative control that is structural, not temporary. The pitch, the positioning, the careful product story on the homepage: an AI assistant will largely route around all of it when a buyer asks what to buy. EY's May 2026 research on AI and consumer-product selection framed the stakes plainly, warning that brands now compete not just to be seen but to be selected, and that discovery and evaluation increasingly happen on third-party AI surfaces rather than channels the brand owns. A brand cannot fix this by improving its own website alone, because its own website was never going to be more than a tenth of the input. The recommendation is being written elsewhere, by other people, and the brand's only real lever is to shape that wider body of content honestly: earn genuine reviews, support credible publishers, get its products into real comparisons, and keep the factual record about itself accurate across the sources models actually read.
For publishers, the finding is a strange mix of vindication and warning. Vindication, because it confirms that independent review and comparison content is not obsolete in the AI era. It is the substrate. Assistants cannot recommend without reading, and what they read is disproportionately what publishers produce. But the warning is sharp. Being the source an AI quotes is not the same as being a page a human visits. Review platforms in the SE Ranking study were heavily cited and still lost most of their traffic, with several down 80 to 90 percent from early 2024 to the end of 2025, because the assistant satisfied the user without sending a click. Affiliate publishing has historically been paid for the click. When the content still gets read but the click does not happen, the value is real and the existing way of capturing it is broken. That gap, influence without a click, is the defining problem of the channel right now, and a separate question from this one.
For the affiliate channel as a whole, the strategic reading is the most interesting. Affiliate content has accidentally become critical infrastructure for a discovery surface its creators did not build and do not control. CJ, one of the oldest affiliate networks, has already repositioned around this, launching an AI visibility service in January 2026 built on the premise that getting a brand's products into the third-party content models read is now a core job of the affiliate channel. That is a meaningful shift in what the channel is for. It was a sales channel, a way to pay partners for traffic that converted. It is becoming, in addition, a visibility channel, a way to populate the corpus that AI assistants reason over. A brand that has invested for years in affiliate relationships turns out to have been seeding the exact content that now decides whether an assistant recommends it. A brand that treated affiliate as a low-priority discount channel finds itself underrepresented in the place purchase decisions are increasingly made.
There is a transitional risk worth naming. If the affiliate corpus becomes the thing that steers AI recommendations, the incentive to flood it with AI-generated, lightly-checked review content gets stronger, and that pollutes the well for everyone. The content that will keep its weight is the content models have reason to trust: first-hand testing, real comparisons, original data, clear sourcing. The thin roundup that AI both produces and learns to ignore is on the losing side of this. The channel's value in an AI discovery world depends on the content staying worth reading.
What to take from it
The counterintuitive fact is simple once stated. When an AI assistant recommends a product, it is mostly reciting other people's reviews, not the brand's. Brand-owned sites are a single-digit to low-double-digit share of what models read, by McKinsey's measure. The bulk is affiliate and editorial comparison content, because that content already contains the comparative judgment a recommendation needs, and because its ranked, structured format is what a model extracts most cleanly.
For a brand, that means the homepage is not the battleground anymore; the wider body of independent content about the brand is, and most of it is affiliate content. For a publisher, it means the work still matters but the old payment mechanism does not fully capture it. For the affiliate channel, it means a quiet promotion: from a place where sales are closed to a place where AI-era discovery is decided. None of this tells you the tactics, how to structure a page, what to test, how to earn a citation. That is the next conversation. This one is just the map: AI product discovery runs on affiliate content, and the brands and publishers who understand that are reading from a more accurate picture of where influence now lives.
Council summary
This post argues that AI assistants build product recommendations almost entirely from third-party reviews, comparisons, and listicles, leaving a brand's own site as a single-digit share of what models read, which hands the affiliate and editorial layer a new role as the raw material of AI discovery. The council verified every load-bearing figure against its source: McKinsey's AI-search adoption and source-preference numbers, the 5 to 10 percent brand-site share cited by Awin's Adam Weiss, the EMARKETER and Max Willens finding on Zenni, Evertune's 25,000-URL listicle study, SE Ranking's December 2025 AI Overviews analysis, and the EY and CJ developments from 2026. One claim was corrected: a Semrush statistic was overstated as brands being several times more likely to be cited via third parties, a figure the source does not quantify, so it was rewritten to match what Semrush actually shows. The reader takeaway is that the homepage is no longer where a purchase is influenced; the wider body of independent content is, and most of that is affiliate content.
Comments