One in Seven Shopping Citations Came From Brand Websites

In OptimizeGEO's study of 40,000 unbranded shopping queries, brand websites supplied only about one in seven citations. Editorial publishers supplied 72%, and three-quarters of shopping answers directed people towards retailers. Brands are increasingly judged through product information published somewhere other than their own websites.¹

A shopper can now arrive at a retailer ready to buy after an assistant has compared products on their behalf. The marketer sees a promising referral, while the customer sees a recommendation assembled from information the brand may never have checked. This week's retail findings suggest those two views can be uncomfortably far apart.

Better visitors, higher stakes

Adobe's latest US retail analysis gives marketers a strong reason to take referrals from AI assistants seriously. In September 2026, traffic from generative AI services grew 71% year on year, and those visits converted 90% better than non-AI traffic. Revenue per visit was 83% higher, based on Adobe's analysis of transactions across more than one trillion visits, 100 million products and 18 retail categories.² Those are unusually attractive numbers for a channel that still feels experimental to many retail teams.

The explanation is plausible even without attributing the improvement to the assistant itself. Someone who has already spent ten minutes asking about size, suitability, alternatives and price may arrive with fewer unanswered questions. They are not browsing a product category from scratch; they have been through part of the consideration process elsewhere. That makes the quality of the earlier answer commercially significant, even when the final purchase happens on a conventional retail website.

It is tempting to read the 90% figure as a promise that adding an AI discovery strategy will raise conversions by the same amount. Adobe is comparing groups of visitors, not reporting a controlled test in which the same shoppers were randomly assigned to an assistant. The traffic could differ in intent, product mix or other ways. What the findings establish is a valuable pattern in observed behaviour, not a guaranteed return from buying another visibility tool.

The comparison also misses an awkward category of shopper: someone who was given the wrong information and never visited the site at all. A recommendation might exclude a suitable product because the assistant thinks it lacks a feature. It could include an unsuitable one because a retailer has an old specification. Neither event appears as a failed conversion on the brand's own website, so the marketing team may have no obvious sign that the product was misrepresented upstream.

Adobe's report contains a second number worth putting beside the conversion result. Its average US retail homepage achieved a 63% Citation Readability Score, Adobe's measure of how much page content an AI system can understand and cite. That is a proprietary metric rather than a universal industry standard, but it highlights a practical issue: a page can look polished in a browser while concealing useful information inside navigation, promotional modules or poorly organised copy.² An attractive storefront and a useful information source are no longer necessarily the same thing.

Whose description gets quoted

OptimizeGEO's research provides a different view of that journey. The company analysed 40,000 unbranded shopping queries across 12 consumer categories and reported that editorial publishers accounted for 72% of cited sources. Brand websites represented roughly one citation in seven, and more than 15% of products appearing in the answers were described inaccurately.¹ These are vendor-reported findings, released alongside a product launch, so they should be treated as evidence from one study rather than a settled error rate for all shopping assistants.

Even with that qualification, the commercial problem is specific. Product information no longer follows a neat path from manufacturer to product page to customer. It can pass through review articles, reseller descriptions, comparison sites and older versions of a catalogue before being assembled into an answer. Each stop offers another opportunity for a dimension to be rounded, a compatibility claim to be copied incorrectly or a discontinued variation to be mistaken for the current model.

Consider a hypothetical small skincare brand selling a fragrance-free cream in three sizes. Its own website may carefully distinguish the ingredients in each formulation, while a marketplace listing combines reviews from an older version. A shopping assistant summarising several sources could present a claim that the current product does not make. The danger is not only a disappointed sale; customers may believe the brand made a promise it never actually published.

This is where marketing teams need to challenge a habit borrowed from SEO. They often treat a mention, citation or recommendation as an unqualified gain. But being mentioned correctly and being mentioned often are different outcomes. A tool that counts the brand's presence in shopping answers without recording the claim made about the product can make the visibility chart look encouraging while leaving the more expensive problem untouched.

There is a further asymmetry in the figures. If publishers dominate citations, brands cannot edit much of the material that shapes an answer. They can correct their own catalogue, provide product evidence, brief retail partners and respond to inaccurate coverage, but they cannot order every independent publication to use a preferred description. The useful ambition is therefore consistency of verifiable facts across sources, rather than perfect control over what an assistant says.

The retailer owns the landing

The same OptimizeGEO analysis found that around three-quarters of the shopping destinations in its sample led to retailers rather than directly to brand websites.¹ For a brand that distributes through several channels, this is not inherently bad news. Retailers provide inventory, delivery options, comparison and checkout convenience, and a sale through a partner is still a sale. The tension appears when the brand is accountable for the product experience but does not control the last description the buyer reads.

Retailers have incentives of their own. A marketplace wants the shopper to find an available item and complete a purchase, even when several similar products would fit the request. A brand wants its own product represented accurately and its differentiating features understood. Those interests overlap frequently, but they are not identical, particularly when a retailer carries old stock, groups product variants or shortens descriptions to fit a standard template.

Nosto's research makes the retailer's position clearer. In a survey of 2,000 consumers across the US and UK, 40% said they were likely to leave a retailer's website and ask a service such as ChatGPT or Claude when onsite search produced irrelevant results. Its new AI Search combines keyword matching with vector search, allowing shoppers to describe a need in ordinary language instead of guessing the retailer's catalogue terms.³ That is a direct attempt to keep the conversation, and the eventual transaction, inside the shop.

This also changes the value of a brand's own content. A product video showing how a coat's hood adjusts or how a kitchen tool fits inside a drawer can make a claim easier to understand than another adjective in a listing. The clip may never be cited by an assistant, yet the observation it documents can be translated into a clear, checkable product fact across channels. Original material becomes useful twice: first as a reason for people to trust the brand, then as evidence that helps others describe it correctly.

For a small ecommerce business, the practical priority is not to choose between Instagram content strategy and product data maintenance. They support different parts of the same decision. A social post can show why a product deserves attention, while a consistent product record prevents that attention from turning into a misleading recommendation. There is little value in making a product desirable on Instagram if a shopping assistant or retailer then gives a potential buyer the wrong size, material or use case.

Product data has to travel

NielsenIQ and Google announced a collaboration on 29 October that shows how seriously larger companies are treating this underlying information. NIQ plans to bring intelligence on more than 250 million products, together with store information covering 90 countries, into Google's AI commerce efforts. The stated aim is to improve product discovery and local availability, although the partnership announcement describes planned work rather than a fully deployed shopping service.⁴ The scale is striking, but the basic ambition is familiar to anyone who has managed a product catalogue.

Customers need to know what an item is, whether it meets their requirements, where it can be bought and whether it is available. Assistants need those answers in formats that can be retrieved and compared reliably. That makes details such as model identifiers, variant names, dimensions, materials, stock status and revision dates more important than they may appear in a conventional marketing plan. Each detail has the potential to determine whether a product is included in a recommendation.

An announcement from a different industry underlines the same principle. People Inc., the publisher behind titles including People, is building custom interfaces that supply AI partners with article text, metadata and freshness information without forcing them to process an entire HTML page. Its innovation chief told Digiday that a typical webpage contains large amounts of material an agent does not need.⁵ Publishers such as Forbes and Time are also experimenting with Markdown-based content access, though their approaches and commercial arrangements differ.

The lesson for product businesses is narrower than 'convert your entire website to Markdown'. Structured access is useful only if the underlying facts are dependable, and a retailer with clear product pages may already be readable without adding another file format. What matters is being able to identify the authoritative description, update it when something changes and distribute the right version to each destination. A machine-friendly page full of stale specifications is still a bad source.

Smaller teams can act on this without an expensive new platform. Give each product a dependable record that distinguishes enduring facts from seasonal marketing copy. Keep the identifiers and variants consistent across the website, Shopify or other commerce software, marketplaces and retailer feeds. When a product changes, update the places where old specifications will continue to circulate, and record which partners have received the revision.

Fix the handoff before the copy

A predictable industry response to all of this will be another set of tools promising to improve citations in ChatGPT, Gemini and other assistants. Some will be useful, particularly when they identify inaccurate descriptions or missing information across a large catalogue. Yet the task should start with a product question, not a platform ranking. Ask an assistant how it would compare two of your best sellers and inspect the facts it uses to make that distinction.

A workable audit is deliberately unglamorous. Pick five important products and write down the claims a buyer must get right: size, ingredients or materials, compatibility, price range, availability and any meaningful exclusions. Test common unbranded questions as well as searches that mention the product by name, and compare the answers with the current product records. Note whether the error originated on the brand site, a marketplace, an independent article or an answer that no cited source appears to support.

Run the test across more than one assistant and repeat it after meaningful catalogue changes. Answers vary by wording, timing and the sources available, so a single prompt is a poor basis for a sweeping conclusion. The goal is to find patterns that can be corrected, such as an obsolete dimension appearing repeatedly or a retailer's description missing an important limitation. Track resolved errors alongside visibility, referral traffic and sales, rather than pretending every correction will immediately produce a measurable revenue lift.

The editorial risk sits beside the technical one. MarTech warned on 29 October that generative engine optimisation could revive the least useful habits of SEO: publishing repetitive pages designed primarily for machine attention, with little to offer a reader.⁶ A brand might make all its descriptions uniform and still fail to explain why the product is worth buying. Accuracy is necessary for a credible recommendation, but it cannot manufacture a compelling product or a persuasive point of view.

Bodycare offers a useful contrast. As the UK beauty retailer rebuilds its High Street presence, it is using a custom platform to connect store activity, inventory information and material produced by local micro-influencers. The company says it has onboarded 700 creators and that their content has generated 7.5 million views.⁷ Those are company-reported figures, but the approach illustrates an important division of labour: people capture experiences and local relevance, while software helps connect that material to what the stores actually carry.

The catalogue is part of the campaign

Marketing teams have spent years treating product feeds as a technical dependency, often maintained far from the people writing campaigns. Shopping assistants are making that separation harder to defend. A beautifully produced launch can attract attention while an out-of-date retailer listing misstates the product in the very answer that sends someone to buy it. The error happens before the campaign receives a click, so it is easily mistaken for weak demand rather than poor information.

There is room for more useful marketing here than a new round of machine-written landing pages. Product teams can provide the specifications and version history, customer support can identify recurring points of confusion, and marketers can turn both into clearer explanations and original demonstrations. The payoff is a brand whose products are easier to recommend accurately, wherever the customer starts. That will not guarantee a citation, but it gives publishers, retailers, assistants and shoppers something better to work with.

The next time a report celebrates an increase in AI shopping referrals, check the recommendation that preceded the visit. Find out which source supplied the decisive claim and whether a customer would recognise the product from the description they were given. A valuable new traffic channel deserves that level of attention before its numbers become another reason to publish more content.

Sources