AI-referred traffic now converts better

A year ago, a visitor who reached a retail site from an AI assistant was, on average, a worse customer than one who came from search or email. That has flipped. According to Adobe's Q1 2026 figures, as reported by Decrypt, AI-referred traffic to US retail sites rose 393% year over year, and in March 2026 those visitors converted 42% better than other traffic. Twelve months earlier they converted 38% worse. Revenue per visit was 37% higher.

The holiday forecast points the same way. Adobe projects US online holiday sales of $275.1 billion, up 6.7%, with AI-referred traffic to retail sites up 130% over the season and 159% on Thanksgiving itself. Mobile is expected to account for 57.4% of online sales.

Two caveats before anyone rebuilds a roadmap around these numbers. First, "AI-referred" today mostly means a person asked an assistant for a recommendation and then clicked through. A human still sees your page. Fully delegated purchases, where an agent compares, selects and pays, are earlier in adoption. Second, the same Adobe release includes a survey of 5,000 shoppers in which 77% of those who used AI for shopping said they felt more confident and 69% said they were less likely to return items. That is self-reported sentiment. We would not put it in a business case on its own.

What the conversion figure does tell you is that shoppers arriving through an assistant have already done their comparison. They land ready to buy. If the product page contradicts what the assistant told them, or checkout breaks, you lose a customer who was further down the funnel than almost anyone else on the site that day.

UCP, ACP and AP2: the standards behind agentic checkout

Three protocols are competing to define how an agent buys something on a shopper's behalf. Manchester Digital's mid-2026 roundup is a useful map:

  • Universal Commerce Protocol (UCP), from Google and Shopify, public since January 2026.
  • Agentic Commerce Protocol (ACP), run by OpenAI and Stripe.
  • Agent Payments Protocol (AP2), from Google, which uses signed "mandates" to record what a user authorized an agent to do.

Our advice to clients is to avoid betting on one of them this quarter unless your commerce platform has already chosen for you. They differ in how they handle payment authorization, and that will matter. All three depend on the same thing underneath, though: product data that is clean, structured, machine-readable and accurate in real time, especially price and availability. Manchester Digital reaches the same conclusion, and it is the part of the work you control today.

What agents see when they visit your store

An agent does not admire your hero video. It reads. It pulls the title, the variant options, the price, the stock status, the shipping terms and the return policy, and it compares them with what it found elsewhere. Adobe's AI Content Visibility Checker gives a sense of how far most stores are from serving that well: product pages averaged 66% on it in the Q1 data.

The failures we look for first are rarely exotic:

  • A price in structured data that differs from the rendered price, usually because a promotion was applied in the front end but not in the feed or the JSON-LD.
  • Variant availability that only appears after a JavaScript interaction, so the default state says "in stock" for a size that sold out on Tuesday.
  • Key specs (materials, dimensions, compatibility) baked into images instead of text.
  • Shipping and returns rules hidden in a modal or a PDF.
  • Consent walls or interstitials that block the content entirely for a non-browser client.

None of this is new. A human shopper squints and works around it. An agent either takes away the wrong information or gives up and recommends a competitor whose page parsed cleanly.

The hidden cost of broken flows when the shopper is a bot

When a person hits a broken coupon field, there is a chance they email support, and at least you hear about it. When an agent hits it, you get nothing. No complaint, no session replay anyone will watch, often not even an analytics event, because the agent may never have fired your tag manager. The sale goes elsewhere and the failure stays invisible.

That changes the economics of QA. Manual regression testing before a release made sense when the main risk was a bad deploy. Now the bigger risk is drift. A merchandiser edits 400 products for a promotion, a third-party reviews widget updates itself, a payment provider renames a field, and none of it passes through your release process. A test suite written in August knows nothing about the bundle you launched last week.

Agent-grade QA: continuous, evidence-backed audits

This is the problem we built Zenyt to solve. The platform runs autonomous agents that browse stores the way shoppers do, with specialized agents for product information, visual content, functionality, checkout, SEO and broken links. Brands including Quip, Caudalie, LG and New York & Company use it to audit their stores continuously.

Two design decisions from that build are worth copying, whatever tool you end up using.

The first is verification before reporting. An agent that flags every oddity buries the team in noise, and within two weeks nobody reads the alerts. Zenyt uses two-stage detection: when a scanning agent finds a potential issue, a second pass re-investigates it and captures screenshot proof before anything reaches a human. That reflection loop is what cut false positives down to a level teams actually act on. If you are building agent-based checks in-house, budget for the second pass from day one. It costs more per page and saves far more in ignored tickets.

The second is one architecture for every catalog size. Zenyt runs the same agent design on 50-product boutiques and on million-page enterprise catalogs. With a big store the temptation is to sample a few hundred "representative" pages. Agents shopping for customers do not sample. They land on the long-tail SKU nobody has opened since 2023, so coverage counts for more than it used to. Under the hood we drive real browsers with Playwright and route work across several LLM providers, so the cost of scanning everything stays manageable.

Performance belongs in the same audit. On Earth.com we reviewed the whole ad stack, from header bidding and floor prices to page performance, because ads that slow a page down cost more than they earn. That work was done for human readers, but the lesson carries over: a page that takes several seconds to render its price is a page an automated client may never finish reading.

A pre-Black Friday readiness checklist

Thanksgiving falls on November 26 this year, which leaves about five weeks. Here is what we would do with them, roughly in this order.

  1. Diff structured data against the rendered page for every product, not a sample. Price, currency, availability and variants must match. Fix the pipeline that produces mismatches rather than patching pages one by one.
  2. Rehearse the promotion data path. Decide how Black Friday prices reach the page, the product feed and the structured data at the same moment, then test it with a dummy promotion this week.
  3. Run checkout end to end every day on mobile viewports. With mobile forecast at 57.4% of sales, cover every payment method, guest checkout, discount codes and the address form for each country you ship to.
  4. Move critical text out of images and modals. Specs, sizing, shipping and returns should be readable in the HTML.
  5. Load a product page with JavaScript disabled and a non-browser user agent. What survives is close to what a lightweight agent sees.
  6. Segment AI referrals in analytics so you can see that cohort's conversion during the season instead of guessing in January.
  7. Name an owner for the protocols. Someone should track UCP, ACP and AP2 and know what your platform supports. Nobody needs to implement them this month.

If you want an outside view of what agents see on your store before the peak arrives, get in touch.

Our guess is that agents will still drive a small share of most stores' orders this season. We would do the list anyway, because every item on it also fixes something human shoppers run into. What nobody knows yet is how an agent treats a store whose data contradicts itself: does it quietly move on to a competitor, or tell the shopper why it did?

Sources