AI Traffic Is Up 393%. Most Retail Sites Can't Read It.

There is a specific kind of frustration in watching high-quality traffic arrive and then watching the page fail to help it.

Adobe's Q1 2026 AI traffic data has that quality to it. Traffic from AI sources to US retail sites grew 393% year-on-year in the first three months of 2026. That figure is built on a real base: the 693% growth during the 2025 holiday season was not a spike that corrected. It continued into the new year and settled into a sustained new level of activity. AI assistants, agentic shoppers, and consumers using ChatGPT, Perplexity, and Google's AI Mode to research and discover products are sending traffic to retail sites at a scale that would have looked implausible in early 2024.

And then Adobe dropped a second finding, buried slightly in the same report, which is the more important one: the average product page across the retail sector scores 66% for machine readability.

What Machine Readability Actually Means

Machine readability in this context is whether the content on your pages is structured in a way that AI systems can parse, interpret, and use accurately. It is not the same as general accessibility, though there is overlap. It is not purely a technical SEO consideration, though there is overlap there too.

Adobe's benchmarking covered different page types across retail sites. Homepages averaged 75%. Category pages 74%. Customer service and FAQ pages scored in the high 70s to low 80s, probably because those pages tend to have structured Q&A content that is inherently parseable. Product pages, at 66%, were the laggards. Which is particularly painful given that product pages are the pages where the detail that drives purchasing decisions lives.

A 66% average means roughly a third of the content on a typical product page (material specifications, size guides, ingredient lists, availability information, product descriptions) is structured in a way that makes it difficult for an AI system to confidently surface and cite. If an AI assistant is trying to help a consumer decide between two jackets, and your competitor's product page scores 85% machine readability while yours is at 60%, the AI is more likely to confidently recommend your competitor's product. Not because it is better, but because it can understand it better.

This is a problem I wrote about as an emerging risk back in mid-2024, when GEO was still a fairly theoretical concept. It is considerably less theoretical now.

The Quality Gap Is Getting Expensive

The conversion data from the same report makes the stakes concrete. In March 2026, AI-referred traffic converted 42% better than non-AI traffic, including paid search and email, historically the highest-quality channels a retailer runs. AI visitors spent 48% longer on site and browsed 13% more pages per visit.

This is high-intent, high-quality traffic converting at significantly elevated rates. And a meaningful portion of it is arriving at product pages that cannot be read properly by the systems that decided to send that traffic in the first place.

There is a compounding effect here worth understanding. AI systems learn from engagement signals. When they send traffic to a product page and that page fails the visitor because the AI could not accurately summarise the product to begin with, that is a negative signal. Over time, sites that score poorly on machine readability get surfaced less, which means less AI traffic, which means less opportunity to build the engagement history that would improve their surfacing.

The rich getting richer, in a new dimension.

The Practical Problem

Why do product pages score lower than every other page type? A few reasons.

One is legacy content architecture. Many retail sites have product content that lives in PIM systems, ERP exports, or vendor data feeds that were built for a keyword-and-crawler SEO world. The structured data markup is minimal or absent. Specifications are buried in unstructured text or loaded dynamically in ways that do not render reliably for crawlers or AI parsing systems.

Another is scale. A retailer with 50,000 SKUs cannot manually audit machine readability on every product page. The sites doing well here are mostly the ones that have built machine readability into their product data pipeline from source — structured data schemas, clean spec attributes, proper schema.org markup on price, availability, and reviews — rather than retrofitting it onto existing content.

A third is organisational. The people who own product content (usually trading or merchandising teams) are rarely the same people who understand how AI systems read web pages (usually technical SEO, architecture, or front-end teams). The problem sits in the gap between them and tends not to get prioritised by either side.

What Actually Moves the Score

The practical interventions that move machine readability scores in meaningful ways tend to cluster around a few things.

Proper schema.org Product markup, covering price, availability, brand, description, and aggregate ratings. Not as afterthought structured data — as core markup present on page load, not added by JavaScript after the fact.

Clean, factual specification content in clearly labelled attributes. "Material: 100% cotton" in a dedicated field beats "this gorgeous 100% cotton shirt" in a long marketing paragraph. Both are human-readable. Only the former is reliably machine-readable.

Consistent naming and categorisation across the catalogue. AI systems comparing similar products across a catalogue need consistent attribute names. If "colour" is sometimes "Color", sometimes "Shade", sometimes absent, that inconsistency degrades readability.

FAQ or Q&A content that directly addresses the questions AI assistants tend to ask about products. This is the GEO angle: structured content that mirrors the natural language questions your products actually get asked is increasingly valuable not just for voice search but for AI-mediated discovery.

None of this is particularly exotic. It is mostly discipline around data quality that the industry has been nominally committed to for years. The difference now is that the consequences of not having it are arriving in the form of 393% traffic growth you are not fully capturing.

The traffic is there. The question is whether your product pages can meet it.