AI brand recommendations and citations are splitting apart

AI tools can recommend your brand and still send the visible citation to someone else. That is the main lesson from a recent Shero Commerce analysis of Google AI Mode, ChatGPT, and Perplexity. Across 1,851 cited sources, only 2.8% were brand-owned pages. Even when a brand was recommended by name, the brand's own site was cited only 31% of the time.

For SEO and GEO teams, this changes the brief. It is no longer enough to ask, "Did the model mention us?" You also need to ask, "Which page did it trust as evidence?" Those are related outcomes, but they are not the same outcome.

• AI recommendation and AI citation are now two separate visibility layers.
• In Shero Commerce's sample, publishers captured far more citations than brand sites.
• Thin or duplicated product copy may weaken a brand page's chance of being treated as the source.
• Brands need both strong owned pages and strong third-party validation.
• GEO measurement should track mentions, citations, sentiment, and source quality together.

• What the data showed
• Why AI cites other sites
• What this changes for SEO and GEO
• BotRank's Take
• How to improve both mention and citation visibility
• FAQ

What did the data actually show?

The short answer is simple: AI shopping answers often name a brand but credit another website as the supporting source. In the Shero Commerce analysis, publishers dominated the citation layer while brand-owned pages barely appeared.

The study looked at buying questions across 60 product categories and collected the sources surfaced by Google AI Mode, ChatGPT, and Perplexity. According to Shero Commerce, 59% of the citations came from publishers such as Good Housekeeping, Verywell Fit, and Reviewed.com. Brand-owned pages accounted for only 2.8% of all cited sources.

The split remained visible even when the AI answer clearly favored a brand. Shero Commerce counted 159 brand recommendations across the platforms it tested, but the recommended brand's own page was cited only 31% of the time. In other words, the answer might say "buy Brand X," while the clickable evidence points to a reviewer, publisher, forum, or marketplace instead.

A good example came from a query about squat-proof leggings. ChatGPT and Perplexity recommended brands such as Gymshark, Alo, and Beyond Yoga, yet the cited sources were third-party sites rather than the brands' own pages. That is the tension many ecommerce teams are starting to notice in practice: visibility in the answer does not guarantee attribution in the source list.

Google AI Mode also looked selective in the sample. Shero Commerce reported that brands were cited or recommended in 9.5% of relevant Google AI Mode store checks, and in around one-third of the tested categories no sampled brands appeared at all. That does not make Google uniquely bad or uniquely good, because Shero did not publish a directly comparable recommendation rate for ChatGPT and Perplexity. But it does reinforce one important point: model behavior differs, and broad assumptions are risky.

Why can AI recommend a brand but cite someone else?

Because recommendation and citation solve two different jobs. A recommendation is the model's final answer to the user. A citation is the evidence the model chooses to justify that answer. Those jobs often overlap, but they do not have to come from the same page.

When someone asks for the best leggings, coffee grinder, protein powder, or office chair, the model is not just looking for the brand's product page. It is trying to assemble enough external proof to sound credible. A review publisher, a comparison article, a buyer's guide, or a Reddit thread may be easier for the system to use because it already frames options, tradeoffs, and credibility signals in one place.

This is where the Shero Commerce dataset becomes more interesting than the headline number. The study did not just count citations. It also looked at the quality and uniqueness of product content on Shopify stores. In a stratified sample, Shero found that 20% of product descriptions were identical to or very similar to descriptions on other websites. It also reported that, among 173 stores it could measure cleanly from raw HTML, 27 had fewer than 50 words of product-specific content.

That does not prove duplication caused the citation gap. Shero Commerce was careful not to claim that. But it does suggest a plausible mechanism. If the same or nearly the same description appears across retailers, marketplaces, and resellers, the model has less reason to treat the brand page as the original authority. And if a page is thin, generic, or written like catalog filler, it gives the model less material to quote in the first place.

There is also a trust issue. Brand pages are inherently self-interested. That does not make them useless, but it does mean many models will lean on third-party validation when answering a commercial query. If Good Housekeeping compares five products and names one as the best for a specific use case, that page may look more like neutral evidence than the product page itself.

This is why the old SEO instinct of asking only, "How do we rank our page?" is too narrow for AI search. In GEO, a page competes not just for clicks. It competes to become the sentence an answer engine is willing to borrow, summarize, or trust.

What does this change for SEO and GEO teams?

It changes the unit of analysis. The page is still important, but the brand, the source type, and the answer pattern now matter just as much. Generative engine optimization, or GEO, is the work of increasing how often AI systems mention, compare, and cite your brand. In practice, that means measuring more than rankings or sessions.

First, teams need to separate answer visibility from source visibility. A brand can win one and lose the other. If your name appears in the answer but a publisher gets the citation, you may still influence the buyer, but you lose attribution, traffic opportunity, and some control over how the evidence is framed.

Second, off-site reputation is no longer a side effect of SEO. It is part of the input layer for AI answers. Review sites, publishers, marketplaces, discussion forums, and expert commentary are often the surfaces where a model learns how your brand is described. That is one reason BotRank recently argued in AI visibility starts before the prompt and ends with citations that citation readiness starts far upstream of the answer itself.

Third, model-by-model testing is mandatory. Shero's data already shows different behavior across Google AI Mode, ChatGPT, and Perplexity. And this pattern aligns with a broader point BotRank made in its analysis of ghost citations in AI search: some systems are more willing to link out, some are more willing to name brands, and some do both inconsistently depending on query type.

Finally, teams need better success criteria. If you only track organic traffic, you may miss a real brand win inside AI answers. But if you only celebrate brand mentions, you may overlook that someone else owns the evidence layer. Strong reporting should treat mention share, citation share, competitor presence, and source quality as separate but connected signals.

BotRank's Take

This study confirms one of the easiest mistakes in AI search: treating visibility as a single metric. It is not. A brand mention, a citation, and a favorable description each solve a different part of the buying journey. If you compress them into one score, you hide the exact problem you need to fix.

That is why Source Analysis matters in this context. It helps teams inspect which pages and domains are actually cited behind LLM answers, then review whether those pages mention the brand clearly enough to be useful. Paired with AI Visibility tracking, you can see where your brand is recommended, where competitors replace you, and where the answer layer relies on third-party evidence instead of your own pages. The practical goal is not to force every citation back to your domain. It is to understand when third-party proof is helping you, when it is displacing you, and where your owned assets are too weak to compete for attribution.

How can brands improve the odds of being both named and cited?

The answer is to build pages and signals that are useful as evidence, not just useful as sales copy. That usually means combining stronger owned content with better off-site validation.

1. Replace thin product copy with original, product-specific information

If multiple sellers use near-identical descriptions, the brand page becomes easier to ignore. A better page includes specific use cases, fit notes, material details, comparison points, FAQs, and explicit claims the model can extract. For a skincare brand, that might mean a clear explanation of ingredients, skin types, and usage steps instead of a generic two-line description.

This is also a technical publishing problem, not just a writing problem. Pages need to be crawlable, structurally clear, and easy for systems to parse. BotRank's GEO Page Analysis is useful here because it checks the technical readiness of the pages you actually want AI systems to understand.

2. Put attribution inside the claim

If you publish a claim worth citing, tie the brand to it directly. A sentence such as "Brand X tested this fabric across 500 wash cycles" is more portable than a vague paragraph where the brand name appears far away from the evidence. Models often extract short passages, so the attribution has to travel with the fact.

This works especially well for original studies, sizing frameworks, material guides, benchmark data, and category explainers. If you make the page useful to a shopper and quotable to a model, you improve your odds on both fronts.

3. Create citation-worthy assets beyond product pages

Many brands expect the product page to do everything. In AI search, that is often unrealistic. A brand may earn more citations from a comparison page, a buyer's guide, a returns explainer, a sizing hub, or a methodology page that explains how the product should be evaluated.

A strong example would be a footwear brand publishing a guide on arch support, fit by activity, and material durability with concrete decision rules. That page can still support sales, but it also gives answer engines something clearer to cite than a sales-first PDP.

4. Treat third-party coverage as part of your AI search strategy

The Shero Commerce dataset makes this impossible to ignore. Publishers captured most citations. That means digital PR, review seeding, expert commentary, creator partnerships, affiliate coverage, and community discussion all influence how the answer layer gets built.

This does not mean every brand should chase the same kind of coverage. A consumer electronics brand may benefit from lab-style reviewers. A fashion brand may benefit more from trusted lifestyle publications and creator comparisons. The point is to invest where the model is already looking for validation.

That is also the logic behind BotRank's post on why AI recommends your competitor and how to change that. The winner is often the brand with clearer third-party evidence, not simply the brand with the louder homepage.

5. Test prompts, then turn findings into actions

Most teams still discover AI visibility problems by accident. A better process is to define the prompts that matter, run them regularly across models, and inspect the answers at source level. Which competitors are named? Which domains are cited? Does the answer tone help or hurt your positioning? Are category pages, help pages, or review sites shaping the narrative?

That is where a workflow matters as much as a dashboard. BotRank's Prompts Studio helps teams create reusable prompt sets, and Recommendations turns repeated weaknesses into concrete GEO actions. That matters because diagnosis alone does not improve visibility. Execution does.

What should teams measure now if clicks are no longer the whole story?

They should measure the answer layer directly. That means tracking where the brand is mentioned, how it is described, who gets cited, and which competitors appear in the same answer. If your reporting still collapses all of that into "AI traffic," you are missing the structure of the problem.

A practical dashboard should include at least five views: mention frequency, citation frequency, source mix, sentiment or perception patterns, and model-by-model differences. For example, if Perplexity cites your brand page but ChatGPT cites review publishers, you have a different problem than a brand that is not named anywhere at all.

It also helps to review whether the cited pages truly reinforce your positioning. A citation is not automatically a win. If the cited page barely mentions your brand, contains outdated pricing, or frames a competitor as the safer option, the visible link can work against you. That is why BotRank's post on recognition-first SEO matters here: in AI search, being understood clearly often matters as much as being present.

The takeaway is blunt. In classic SEO, you could often treat the click as the conversion proxy. In AI search, the path is more fragmented. A buyer may see your brand in one answer, validate it on a publisher's site, compare it in another model, and only later visit your website. Measurement has to follow that reality.

FAQ

Is a brand mention in AI answers enough?

No. A mention can build awareness, but it does not guarantee traffic, attribution, or message control. If another site owns the citation, that site may shape the evidence layer around your brand.

Why do AI tools prefer publishers over brand sites?

Publishers often package comparisons, pros and cons, and buyer guidance in a format that is easy to reuse as evidence. Brand pages can still win citations, but thin or duplicated copy makes that much harder.

Does rewriting product pages automatically fix the problem?

Not automatically. Better owned content helps, but Shero Commerce's analysis did not prove that content changes alone cause citation gains. Off-site validation and model behavior still matter.

What is the best KPI for AI search visibility?

There is no single best KPI. The strongest setup tracks mentions, citations, source quality, competitor presence, and sentiment together so teams can see whether the problem is discovery, recommendation, or attribution.

What should a brand do first?

Start by auditing the prompts that matter most to your category across multiple models. Then strengthen the pages and third-party signals that repeatedly appear in the answers, instead of guessing where the visibility gap is.

If your brand is being recommended but not cited, you do not have a small reporting issue. You have a visibility architecture issue. The fastest next step is to audit real prompts, inspect the source layer, and decide whether your gap is owned content, third-party proof, or both. That is exactly the kind of work BotRank is built to make measurable.