Ask ChatGPT for the best wireless earbuds under 150 dollars and watch what comes back. You get a shortlist, a sentence of reasoning for each pick, maybe a small carousel. What you almost never get is a recommendation built from the product pages of the brands themselves. The assistant read reviews, a few roundup articles, a Reddit thread, a couple of retailer listings, and assembled an answer. The brand's own marketing copy was, at best, a minor witness.
That is the uncomfortable starting point for generative engine optimization in commerce. Part 1 of this series covered the technical foundation: a clean product feed, accurate Schema.org markup, complete attributes, and crawler access, the work that makes a single SKU machine-readable. That work is necessary but not sufficient. A perfectly structured feed gets you eligible to be recommended. It does not get you recommended. What earns the recommendation is something a merchant mostly cannot edit directly: a reputation, spread across the web, that an AI model reads as evidence.
This post is about that second layer: how an LLM decides which product to name, why your own site is a weak source, what off-site presence moves the needle, and what a merchant can and cannot control.
Where the recommendation actually comes from
Start with the data, because the instinct here is wrong.
Omniscient Digital analyzed 23,387 unique citations across 240 branded prompts, spread over four industries and queried through ChatGPT, Perplexity, Gemini, Google AI Mode, and AI Overviews. When someone asks an LLM about a brand, only 23 percent of what it cites comes from that brand's own website. The other 77 percent is external: editorial media, forums and social platforms, review sites, and directories. Narrow it to the queries that matter most for a purchase, the customer-experience and "is this any good" questions, and earned media climbs to 82 percent of citations. The brand's own pages all but vanish from those answers.
The same body of research, sorted by content type, found that 57 percent of citations for branded queries land on reviews, listicles, forums, social media, and case studies. Product and commercial pages account for around 12 percent. The model is not being difficult. It is doing what a careful human shopper does: trusting the description written by someone with nothing to sell over the description written by the seller.
Shopping queries add a second mechanism. A January 2026 study by Semrush, testing 100 shopping prompts five times each, found that 75 percent of the time ChatGPT's top product recommendation also appeared in Google Shopping's first three results, with retailer, title, and price matching exactly. ChatGPT's shopping answers lean on the Google Merchant Center ecosystem and on retailer listings, not on a brand's storefront. Google's own AI shopping runs on the Shopping Graph, a product database it describes as holding more than 50 billion listings and refreshing over 2 billion times an hour. The structured retail web is the substrate, and your homepage is not.
So the recommendation is assembled from three pools of evidence: third-party editorial and review content, community discussion, and structured retail listings. A merchant's owned site touches all three only indirectly. GEO has to be built around that fact.
Why your own site is a weak source
This is not a flaw in the models but a sensible design choice.
An LLM generating a product recommendation is trying to predict what a knowledgeable, neutral person would say. Brand-owned copy is structurally suspect for that job, because every brand claims its product is excellent. The signal in "best-in-class sound quality" on a manufacturer's page is close to zero, because the page would say that whether or not it were true. The same sentence in a review from a publication that also panned three competitors carries real information. Models are trained, directly and indirectly, to weight the second kind higher.
Reviews matter because they are adversarial in the right direction. Comparison and roundup content matters because it has already done the ranking work the assistant is being asked to do. Forum threads matter because they capture unscripted experience, the failure modes and quiet praise that no brand publishes about itself. Retailer listings matter because they are standardized, cross-checkable, and carry aggregate ratings the model can read as a crowd verdict.
A brand's own product page is useful for one narrow thing: ground truth on specifications. When the question is "does this support a particular codec" or "what are the exact dimensions," the manufacturer's page is authoritative, and the Omniscient research bears this out, with owned content performing best on product-functionality queries. But the buying decision, the which-one question, is rarely a specification lookup. It is a judgment call, and for judgment the model goes outside.
The practical reading: to an AI assistant, your storefront is a fact sheet, not a salesperson. Treat it as the place that keeps specs, price, and availability correct and machine-readable, the Part 1 work, and stop expecting persuasive copy on it to do persuasive work. The persuasion happens elsewhere, and that is where GEO spends its effort.
What actually moves an AI recommendation
If 77 percent of the evidence is off-site, the work is off-site. Five levers carry most of the weight.
Third-party reviews, with volume and recency. Review platforms are now among the most cited sources in AI answers, and the effect size is large. A March 2026 study commissioned by Trustpilot and run by Seer Interactive, examining more than 800,000 AI responses across ChatGPT, Gemini, Perplexity, and Google AI Mode, found that a brand with no Trustpilot profile was cited in roughly 1 percent of relevant answers. Simply having a profile lifted that to 53.5 percent. Brands that collected more than 80 reviews and responded regularly reached 75.3 percent. Separate analysis by Passionfruit found that domains with profiles on G2, Capterra, or Trustpilot were about three times more likely to be cited. The lesson is not "collect reviews," which every merchant knows. It is that reviews on independent, well-known platforms are a direct input to AI visibility, and an active, recent, responded-to presence beats a larger but stale one.
Comparison and roundup content, including content you do not own. When a shopper asks for the best option in a category, the assistant looks for sources that have already answered that question. A "best X for Y" article from a credible publication is the highest-value real estate in GEO, because the model can adopt its ranking almost wholesale. You cannot write those articles yourself for the outlets that carry weight. You can earn placement: sampling products to reviewers, pitching the editors who run category roundups, making sure the publications your buyers trust have actually tested your product. This is digital PR pointed at a new target, and the Omniscient analysis found listicles and comparison pages among the most cited content types for branded queries.
Reddit and forum presence. Community discussion is disproportionately powerful. Reddit accounts for a large share of citations across the major assistants, and on Perplexity it can reach close to half of top sources for some query types. Research from 5W Public Relations found Wikipedia and Reddit together drive more than a quarter of ChatGPT's US citations. A brand cannot fabricate this and should not try; astroturfed threads are detectable, against platform rules, and a reputational landmine. What a brand can do is be genuinely present: a real, identified account answering questions in the subreddits where its category is discussed, a product good enough that unprompted threads say so, support that resolves complaints before they become the top-voted warning. The goal is to be discussed accurately, not ignored or panned.
Retailer and marketplace listings. Because shopping answers lean on the structured retail web, being listed accurately, with complete attributes, on the marketplaces an assistant trusts is itself a GEO move. A product carried by several reputable retailers, each with consistent pricing and aggregate ratings, reads as established and verifiable. A product that exists only on its own storefront has a thinner, harder-to-corroborate footprint.
A consistent, factual entity across the web. Models build an internal sense of what a brand is from everything they see: Wikipedia where it qualifies, directory entries, the brand's description repeated across listings. When those sources agree, the model is confident. When the category, price band, founding facts, or product line conflict from source to source, the model hedges or omits. Consistency is unglamorous and it is real work.
How this differs from classic SEO
GEO and SEO overlap enough to be confused, and treating them as one will misfire. Classic SEO optimizes a page you own to rank in a list a user then scans and clicks; the unit of success is your URL in position three. GEO optimizes your presence across sources you mostly do not own, so a model selects you when it composes an answer; the unit of success is your product named in the recommendation, often with no click at all.
Three differences matter in practice. First, locus of control. SEO is largely on-page, levers you hold directly. GEO is mostly off-site, so it runs more like PR and reputation management than technical optimization. Second, the win condition. SEO wins a ranking; GEO wins a citation, and you can be the consensus pick of the web and still get no visit, because the answer resolved in the chat. Third, the feedback loop. SEO has mature tooling and a position you can watch. GEO is noisier: the same prompt can yield different answers across sessions and models, so you measure with sampled prompts and trend lines rather than a single rank. Profound launched a Shopping Analysis product to show retailers how their goods are surfaced, described, and ranked inside AI answers, and competitors are building similar visibility.
Strong SEO fundamentals still help: the same crawlable, well-structured content that ranks also gets read by the models, and a brand invisible to classic search tends to be invisible to AI too. GEO is a layer on top of SEO, not a replacement. The mistake is assuming the old on-site playbook is the whole job when most of the new evidence sits beyond your domain.
What a merchant can and cannot control
Honesty matters here, because GEO is sold with more certainty than it deserves.
You can control your feed, schema, attributes, pricing accuracy, inventory freshness, and crawler access, the Part 1 foundation. You can control whether you are listed on the review platforms and retailers that assistants trust, and whether that data is complete. Through real digital PR and a genuine community presence, you can influence how often credible third parties write about your product. You can control the consistency of your own facts everywhere they appear. That is a substantial surface, and most merchants have barely begun on it.
What you cannot control is the model's final output. An assistant can still omit you, rank a competitor above you, or state something about your product that is simply wrong. Attribute and pricing hallucinations are a known failure mode that no amount of optimization fully removes, because they are a property of how these systems generate text. Coverage is uneven across categories and regions. You cannot edit ChatGPT's answer the way you edit a title tag. The realistic goal is to shape the ground truth the model retrieves so that the most available evidence points your way. You are stacking the odds, not setting the answer.
A second limit: this is not a placement you can buy. Product recommendations inside ChatGPT shopping research and the equivalents in Gemini and Perplexity are organic, drawn from public retail data and the open web. Advertising units are appearing alongside them and are a separate matter. The recommendation itself is earned, slowly, through reputation. Any vendor promising a shortcut is selling something the platforms have not built.
Where this is heading
The direction is set. AI-driven traffic to US retailers rose 393 percent year over year in the first quarter of 2026, on Adobe Analytics data covering more than a trillion retail site visits, and that traffic converts better than non-AI traffic, because a shopper arriving from an assistant has already been advised. The behavior behind the numbers is shifting fast: a 2026 study found 37 percent of consumers now begin a search with an AI tool rather than Google or Bing, and a separate survey put 58 percent of consumers using AI tools somewhere in product research. As those shares grow, the recommendation becomes the shelf, and the brands on it are the ones the web has vouched for. The feed and schema work in Part 1 makes you legible; the off-site work in this post makes you credible. Brands that treat reviews, comparison placement, community presence, and entity consistency as a measured program will be the ones AI assistants name. The rest will be eligible, structured, machine-readable, and quietly skipped.
GEO for ecommerce is not a trick of phrasing or a schema tag. It is the slow accumulation of third-party evidence that a product is worth recommending, made visible to systems that now read all of it. You cannot write that evidence yourself. You can earn it, and you can make sure the parts you do control never contradict it.
Council summary
This post argues that GEO for ecommerce is won off-site: an AI assistant builds a product recommendation mostly from reviews, comparison articles, forum threads, and retailer listings, not from the brand's own pages. The council verified the headline figures against primary sources, including Omniscient Digital's 23 percent owned versus 77 percent external citation split, the Trustpilot and Seer Interactive study showing review profiles lifting AI citation rates from roughly 1 percent to 75.3 percent, and the 393 percent year over year jump in AI retail traffic from Adobe Analytics. Two claims were corrected: an unverifiable 239 percent earned-media lift attributed to Passionfruit was cut, and a vague consumer-adoption stat was replaced with sourced figures of 37 percent starting searches with AI and 58 percent using AI tools in product research. The reader takeaway is concrete: treat reviews, comparison placement, community presence, and entity consistency as a measured program, because that is what gets a product named when the answer resolves inside the chat.
Comments