Four numbers,each with its source.
Everything in this section is third-party measurement, cited. The argument that follows it is argument, and is labeled as such.
Better conversion from AI-referred traffic than from any other channel
Adobe Digital Insights, 2026Year-over-year growth in AI-referred traffic to U.S. retail sites in Q1 2026
Adobe Digital Insights, 2026Domains an answer engine typically cites in one response, against ten blue links
Profound, 2025Machine-readability score Adobe assigned to retail product pages — the lowest of any page type
Adobe Digital Insights, 2026What those numbers mean for a catalog.
Four readings of the same data. This is Pyxl's interpretation rather than measurement, and it should be discounted accordingly.
The traffic is small and it is the best traffic you have.
Two numbers have to be held at once. AI-referred traffic to U.S. retail sites grew 393 percent year over year in Q1 2026, and it converts 42 percent better than traffic from any other channel. Both figures come from the same Adobe dataset. The share of total sessions is still low enough to ignore on a dashboard, which is exactly why most brands are ignoring it. A channel that converts better than every other channel and is growing at that rate does not stay a rounding error, and the work to be visible in it takes months to compound rather than weeks.
The list got shorter.
A search results page shows ten links. An answer engine cites two to seven domains in a single response. That is the whole change compressed into one sentence: the same query that used to produce a page a shopper could scan now produces a recommendation with a handful of sources behind it. Ranking eleventh on Google was disappointing. Being the eighth domain an engine would have cited is nothing at all.
Product pages are the least machine-readable pages on the web.
Adobe scored retail product pages at 66 percent machine-readability, the lowest of any page type it measured. That is a structural finding, not a content one. Product pages are built for a human eye and a conversion funnel: imagery, badges, tabs, review widgets, recommendation carousels. What the model needs is the boring part — material, dimensions, fit, color, pattern, audience, availability, price, designer credit, and the relationships between products — exposed as structured data rather than implied by a photograph.
The gap between the two readers is where the work is.
Every product page now has two readers. The shopper, who is served by the imagery and the copy. And the model, recommending on their behalf before they arrive, which reads the structured layer underneath and nothing else. Most catalogs serve the first reader well and the second one by accident. Closing that gap is not a redesign. It is schema across the full catalog, FAQ content in the phrasing buyers actually use, a brand entity the engines can resolve, and the engineering to keep all of it accurate as the catalog changes.
Four moves,in this order.
The order matters more than the list. You cannot fix an entity you have not measured, and a catalog marked up underneath a brand the engines cannot resolve is structured invisibility.
Measure first
Run real buyer queries through ChatGPT, Claude, Perplexity, Gemini and Google AI Overviews. Log every citation, every mention and every miss, scored against three to five competitors. Without a baseline, every later decision is a guess.
Fix the entity
An engine will not recommend a brand it cannot identify with confidence. Knowledge panel, canonical bios, Organization schema, llms.txt, and consistent identity across every property the engines read.
Structure the catalog
Complete Product schema across one hundred percent of the active catalog, FAQ schema on the priority product and brand pages written in the brand voice, and the platform engineering to deploy it into the theme and metafields rather than bolt it on.
Then keep reading it
Citation share month over month, across all five engines, against a named competitive set. A foundation nobody monitors is a foundation that decays quietly.
Sources, in full.
Questions we getbefore the first call.
Does traffic from AI assistants actually convert?
Yes, better than any other channel. Adobe measured AI-referred traffic converting forty-two percent better than traffic from any other source in 2026. Volume is still a small share of total sessions, which is exactly why the brands acting now are doing so cheaply.
How fast is AI-referred shopping traffic growing?
Adobe measured 393 percent year-over-year growth in AI-referred traffic to U.S. retail sites in the first quarter of 2026. Compounding at that rate, a channel that is a rounding error on this year's dashboard is a material line on next year's.
How many sources does an AI answer cite?
Two to seven domains in a typical response, against ten links on a conventional results page. The list of brands a shopper is shown is dramatically shorter, which raises the cost of not being on it.
Why are product pages hard for AI to read?
Adobe scored retail product pages at sixty-six percent machine readability, the lowest of any page type it measured. Product pages are built for the human eye — imagery, badges, tabs, review widgets — while a model reads the structured layer underneath. Most catalogs serve the shopper well and the model by accident.
What should an eCommerce brand do first?
Measure before building. Run your real buyer queries through all five major engines and log where you appear, where a competitor appears instead, and where nobody is named. That baseline tells you whether your problem is the catalog structure, the brand entity or the content, and those three have very different price tags.