Mentioned, Linked, Cited: The Three Ways an AI Answer Treats Your Store

Someone screenshots an AI answer with your brand name in it and drops it in Slack. Everyone is pleased. Then you look for the traffic and there is none.

That is not a tracking failure. A brand mention and a citation are different events, and in retail the gap between them is enormous.

The gap, measured

BuzzStream ran a study across roughly 12,000 AI responses, about 200 brands, ten industries and 2,967 prompts, on Google AI Mode, AI Overviews, Gemini and a GPT-5 class model, last updated on July 14, 2026. They resolved 221,946 cited URLs against tracked brand domains.

Across all industries, only 23.1% of brand mentions came with a citation to that brand. The reverse ran at 69.9%: when a brand's domain was cited, the brand name usually appeared in the text too.

Retail was worse than average on the first number. 16.5% of retail brand mentions carried a citation.

Five out of six times an AI answer names a store, it points nowhere.

Prompt type moves it a lot. Single-brand queries backed 39% of mentions with a citation. Head-to-head comparisons, 35.8%. Open-ended category prompts, the ones that read "best waterproof duffel for cycling," managed 7.2%.

Worth sitting with that one. The query where the shopper has no brand in mind yet, which is the query you have the most to gain from, is the query least likely to hand them a link.

Three rungs, not two

It helps to stop thinking in binary. There are three distinct outcomes and they are worth different amounts.

• Mentioned. Your name is in the text. The shopper now has to go and search for you. Some will. Most will keep reading the answer.
• Domain cited. A link to your site, usually the homepage or a collection page. Better. The shopper lands somewhere generic and has to find the thing that was recommended.
• Product cited. The answer links the exact product page for the exact product it just named. The shopper arrives on a page with the spec, the price and a buy button.

Only the third one removes friction entirely. Everything else asks the shopper to do work at the precise moment their attention is cheapest to lose.

Anyone measuring AI visibility with a single blended score is averaging across three things that behave differently. Split them.

Why product pages lose

Two datasets point the same direction, and neither is flattering.

Siege Media analysed roughly 1,000 bottom-of-funnel prompts and 57,095 citations across ChatGPT, Perplexity, Gemini and Google AI Overviews between January 29 and February 4, 2026. Comparison pages and listicles dominated the citations. Product pages and homepages showed up at single-digit rates.

Then there is the supply side. Adobe scored retail pages on machine readability in its Q3 2026 AI traffic report and found homepages averaging 75%, category pages 74%, and product pages 66%. Returns, contact and FAQ pages all scored above 80%.

Read that again. Your returns policy is more legible to an AI than the page you want it to recommend.

That is not a coincidence. Returns pages are plain text with clear headings and no theming. Product pages are the most heavily merchandised surface in the store: tabbed content, review widgets, image carousels, specs rendered as icons, size guides behind a modal, key facts written into marketing prose instead of fields.

Every one of those decisions was made for a human. Every one of them costs you legibility.

Moving from mentioned to cited

The model already knows your brand exists. It declined to attach a URL. Usually one of three reasons.

There is nothing worth attaching. The answer says "several brands make good merino base layers" and names you among them. No page anywhere, yours or anyone else's, specifically supports the claim being made. A page that answers the actual question gives the model something to point at.

Or the claim is not corroborated anywhere but your own site. Assistants lean hard on third-party sources for anything that sounds like an opinion. If the only place your durability claim exists is your own product description, the model will name you and cite someone else. That is an off-site problem, not an on-site one.

Or your page cannot be read. Check the basics: is the AI crawler you care about allowed in robots.txt, does the page render its content without JavaScript, does the important text exist as text.

Moving from domain cited to product cited

This is where structured data stops being a checkbox and starts being the whole game.

The model needs to distinguish one of your products from the other four hundred. It does that on fields, not prose.

• Product and Offer markup on every product page, with price, currency, availability and a stable SKU or GTIN. Not just on the ones you remembered.
• The deciding spec as an attribute, not a sentence. "Waterproof to 10 metres" in a structured field beats the same phrase in paragraph three. A model filtering a hundred candidates for someone who asked for waterproof will use the field.
• Category set properly. A product filed under "Accessories" competes with phone cases. Filed correctly, it competes with what it actually is.
• Review data marked up if you genuinely have it. AggregateRating with real counts, not decoration.
• Unique copy per product. Supplier-default descriptions give a model nothing to tell you apart with, which is the whole problem with a dropshipped catalogue and a good place to start a readiness check.

On llms.txt: it costs almost nothing to publish and no major assistant has publicly confirmed using it as an input. Publish it if you like. Do not publish it instead of fixing your product data.

The number to watch

Once you are tracking all three states, the ratio worth putting on a dashboard is product citations divided by total mentions.

Most stores start near zero. That is normal and it is not a reason to panic. It is a reason to stop reporting mentions as if they were wins.

Watch the ratio month over month on a frozen prompt set. If mentions climb while product citations stay flat, your brand awareness inside the models is improving and your pages are still unreadable. That is a specific, fixable diagnosis, and it is invisible to any tool that reports one blended score.

If mentions and product citations climb together, whatever you did to your product data worked. Do more of it.

Seokai's AI Visibility tracker separates brand mentions, site citations and product-page citations on tracked queries, and the site audit flags the product pages that are not readable yet, so you can see both halves of the problem in one place.