AI agents are shopping. Can they buy from you?
Shopping assistants read your store the way a crawler does: no JavaScript, no clicks, no patience. If they can’t find your products, read your prices, or reach your checkout, they recommend someone who can. This checks all three, free.
Whether your robots.txt turns shopping agents away, whether a sitemap exists to enumerate products, and whether the page answers a non-browser client at all.
Whether price and availability survive without JavaScript, and whether Product structured data is present and valid. This is where most stores fail: the price renders for a human and does not exist for an agent.
Whether a login wall or consent banner stops an agent before checkout, and whether your primary call-to-action is reachable as a plain link.
Checked against your landing page, the way an agent first arrives. This catches site-wide blockers: a turned-away crawler, prices that only exist after JavaScript, a login wall, missing product markup. It does not open an individual product page or attempt a real checkout.
Why this matters more than it did last year
6 figures, each linked to whoever published it. The last two come from academic studies of agents attempting real tasks on live websites.
In March 2026, visitors arriving from AI sources converted 42% better than visitors from other channels. In March 2025 the same traffic converted 38% worse.
In twelve months AI referrals went from the worst-converting traffic a retailer received to the best. Someone who arrives after an assistant has already answered their question arrives further along.
Adobe Analytics, April 2026
Across Shopify storefronts in Q1 2026, more than half of AI-referred sessions started on a product detail page, compared with about 20% for organic search. AI-referred orders grew nearly 13x year over year.
AI skips your homepage and your category pages. It sends people, and agents, directly to the product page, which means that page is doing work the rest of your site used to do first.
Shopify, May 2026
On Adobe’s AI content visibility assessment, retail product pages averaged 66%, the lowest of the page types measured. Homepages averaged 75% and category pages 74%.
The page carrying the price, the variants, and the buy button is the least machine-readable page on the average retail site, and it is now the page AI traffic lands on first.
Adobe Analytics, April 2026
Walmart’s EVP of product and design said purchases completed inside ChatGPT converted at one third the rate of shoppers who clicked out to Walmart’s own site. Walmart ended the Instant Checkout integration and now routes shoppers to its own platform.
The largest retailer to try buying inside the assistant measured it converting worse and moved the transaction back to its own site. That makes your checkout, not a protocol handoff, the thing an agent has to operate.
Search Engine Land, March 2026
Evaluated on 300 tasks across 136 live websites, the best web agents succeeded 61.3% (OpenAI Operator) and 56.3% (Claude Computer Use 3.7) of the time. The researchers found earlier benchmarks overstated performance: a baseline that only used web search scored 51% on the older WebVoyager benchmark and 22% here.
On real sites, roughly four attempts in ten fail. The gap between a site an agent can finish on and one it abandons is not a rounding error, and it is decided by your pages.
Ohio State University and UC Berkeley (Online-Mind2Web), 2025
When a call to action was baked into a banner image, every model tested clicked it 0 times out of 10. Expressed as a real page element instead, the same call to action was clicked up to 7 times out of 10. The study also found alt text and speculative data-llm-* attributes produced no interaction at all, 0 out of 30.
An agent builds a list of things it can click and acts on that list. A button that exists only as artwork is not on it, and the alt text most people assume covers this does nothing.
We test for these. The failure modes above are scored checks in our public rubric (2026.10.4), which means a scan tells you whether your pages have them, and the fix pages are free to read whether or not you buy anything.
The bordered line under each figure is our reading of it, not the source’s. Figures are replaced rather than accumulated when a publisher updates them, so this page carries one vintage of each number.