MEIKAICited is not recommended: measuring AI visibility in retail
A shopping assistant naming your product is not the same as a shopping assistant recommending it, and neither is the same as a sale. Most retail AI-visibility reporting blurs those three things together, which is how teams end up celebrating a number that has no commercial meaning.
This is a guide to what the observable signals in AI-assisted shopping actually license you to conclude, and what to fix on your product data once you know.
What each signal does and does not prove
Four things can be measured today. They are routinely treated as interchangeable and they are not.
| Observed outcome | What it means | What it does not prove |
|---|---|---|
| Product cited | A source URL for the product appeared in the response. | That the product was recommended, or that the shopper saw the citation at all. |
| Product mentioned | The assistant named the product or brand in its answer. | That the information was fetched live, or that it is current. |
| Included in a shortlist | The product was among the options offered for a defined prompt. | That it will appear for a different shopper, phrasing or market. |
| AI referral visit | A shopper arrived from an identifiable AI source. | That a specific citation or mention caused the visit. |
Mention volume is therefore a poor primary KPI. A product can be mentioned constantly and described wrongly. It can be cited from a page that no longer reflects the assortment. It can be shortlisted in one market and absent in the next.
How big is this channel, really?
Adobe's July 2025 reading of U.S. retail traffic remains the most-quoted figure in this space: generative AI referrals were up 4,700% year over year. In the same measurement, Adobe described the channel as small relative to established sources and found that AI-referred visits converted 23% less often than non-AI traffic.
That is rapid growth from a very low base, measured a year ago. Both qualifiers usually disappear when the number gets recycled. So the case for acting has little to do with the size of AI referral traffic today. You act because the cost of being described inaccurately is already being paid, silently, in answers that never generate a click at all.
Your old product pages may still be selling for you
Between March and July 2026 we tracked what happens when cited pages disappear. A publisher removed a large directory of articles on 27 March, briefly restored them from 3 to 8 June, then removed them again. We kept measuring the same URLs across five AI platforms throughout. Comparing citations per 1,000 responses for a fixed cohort of 126 brands, the 15-21 July window put Perplexity at 109.5% of its March baseline, ChatGPT at 12.2% and Google AI Mode at zero. The full citation-decay analysis has the method.
For a retailer, that finding lands directly on the assortment lifecycle. A discontinued line, a superseded model, a regional product that was delisted, a page pulled during a replatform. Any of these can keep being cited and described for months on some platforms while vanishing immediately from others. Removing a page does not remove the claim.
The mechanism in that study is worth copying down as an anti-pattern, because retail sites do it constantly. The removed pages returned HTTP 200 with a generic error page of about 3.3 KB where an 84 KB article used to be. To a crawler that is a live page, so nothing signals that the content is gone. If your out-of-stock, discontinued and delisted URLs behave the same way, you are asking assistants to keep repeating stale commercial facts.
What to fix on your product data
Assistants can only describe a product accurately if the facts are retrievable and consistent. Five things carry most of the weight.
- Put the deciding attributes in visible text. Names, variants, identifiers, materials or ingredients, dimensions and compatibility should be readable on the page itself, not hidden in a script tag or behind a tab that renders client-side.
- Keep commercial facts current. Price, availability, delivery, returns and regional restrictions are the claims most likely to be repeated after they stop being true.
- Make structured data match what a shopper sees. Google's own guidance for AI features asks for exactly this, and markup that overstates the page is a liability, not an advantage.
- Give each product one stable URL. Fragmenting a product across seasonal, campaign or temporary pages splits the evidence and multiplies the number of things that can go stale.
- Retire pages deliberately. Return a real status code, or keep the URL alive with an accurate discontinued state and a pointer to the successor product. Do not serve a soft error at HTTP 200.
What Meikai measures, and what it will not tell you
Meikai measures how monitored AI experiences mention and cite brands and products across a reviewed prompt set, per platform and per market. That surfaces missing products, wrong attributes, thin source coverage and the gaps between models.
It is visibility measurement, not transaction attribution. Where referral and conversion data exist, treat them as separate downstream outcomes and resist the temptation to draw a straight line from a citation to a basket. Our prompt-modeling framework covers how to build the prompt set so that the measurement is repeatable in the first place.