For twenty-five years, online retail was built around a human doing the looking. A person searched, compared a few tabs, landed on a product page, read the reviews and checked out on your domain. Every metric that matters to an ecommerce team assumes that sequence. Sessions, bounce rate, add-to-cart, cart abandonment, even the layout of the page itself all exist because a human eye was the thing being persuaded.

A machine is doing the looking now.

In July 2026, traffic arriving at US retail sites from AI assistants rose 62% year on year, according to Adobe Analytics. Measured against October 2024, when Adobe started tracking the category, it is up 1,219%. Those visitors convert at a rate 60% higher than everyone else and generate 53% more revenue per visit. It is the eleventh month in a row that AI traffic has beaten non-AI traffic on conversion. Adobe also found something less flattering for retailers: product pages are among the hardest page types for a large language model to read.

What the Adobe data actually shows

Adobe draws these figures from more than a trillion visits to US retail sites, so the sample is not the problem. The behaviour underneath the headline number is consistent and worth listing plainly.

Shoppers who arrive from an AI assistant spend 59% more time on the site than other visitors. They bounce 33% less. They engage 14% more, and they add items to the cart 28% more often. Then they convert 60% better and spend more when they do.

The direction of travel is just as striking as the level. In May 2026, AI-referred retail traffic was up 138% year on year and hit the highest share of total retail visits Adobe had recorded. A year before that, the same measurement showed AI visitors converting at roughly half the rate of everyone else. The metric did not improve. It inverted.

Why the 60% conversion lift is smaller than it looks

This is where most coverage stops and the sales deck begins. It should not.

Start with the denominator. Even at a record high, AI referrals are a small slice of total retail visits. A 60% conversion advantage calculated on a small, self-selected pool tells you who those people are. It does not tell you what the channel does at scale.

Then think about who they are. Someone who asks an assistant to find a waterproof running jacket under a certain price, reads the comparison, and clicks through to one retailer has already done the shortlisting. They arrive at the bottom of the funnel, pre-qualified, with the comparison work finished. Of course they convert. Paid search looked exactly this good in 2004, for exactly this reason, and the advantage compressed as volume arrived and the mix shifted from high intent to everything else.

The inversion should also make you cautious rather than excited. A channel metric that flips from half the conversion rate to well above it inside twelve months is not measuring a stable property of the channel. It is measuring two moving things at once: consumers learning what to ask, and the assistants themselves getting better at handing over qualified traffic instead of curious browsers.

The honest read is this. Agentic commerce is not yet a channel you can plan a budget around. It is a small stream of unusually good traffic that costs very little to prepare for, and the preparation work is the same work that improves your site for answer engines and zero-click search anyway. That is the actual argument for doing it now. Not the revolution. The overlap.

Three protocols decide whether an agent can buy from you

The buying layer split into three architectures over the past year, and they reward different things.

OpenAI launched Instant Checkout and the Agentic Commerce Protocol in September 2025, built with Stripe and open-sourced. The agent completes the purchase inside the chat, then hands the order to the merchant’s own backend. The merchant stays the merchant of record and keeps fulfilment, returns and support. OpenAI states that enabling Instant Checkout does not give a product preferential ranking.

Google announced its Universal Commerce Protocol in January 2026, supporting checkout from eligible listings inside AI Mode in Search and the Gemini app, plumbed through Merchant Center.

Amazon’s Rufus is the outlier. It is a closed loop, trained on Amazon’s own reviews, purchase histories and product questions rather than the open web. Nothing you do to your own site reaches it.

Dimension	OpenAI ACP	Google UCP	Amazon Rufus
Architecture	Open protocol	Open protocol	Closed loop
Where checkout happens	Inside ChatGPT	AI Mode and Gemini	Inside Amazon
Data source it reads	Open web and merchant feeds	Merchant Center feeds	Amazon catalogue and reviews
Merchant of record	The merchant	The merchant	Amazon
Who owns customer data	The merchant	The merchant	Amazon
What it rewards	Feed quality and structured product pages	Merchant Center feed health	Listing quality and review velocity
Payment plumbing	Stripe tokens or delegated payments	Merchant Center	Amazon Pay
Effort to enable	Low if you already use Stripe	Low if your feed is current	No new work, existing Amazon SEO
Best fit	Considered purchases, DTC	Replenishment and consumables	Anything already selling on Amazon

The practical order for most brands is unglamorous. Fix the feed first, because both open protocols read it and Merchant Center hygiene is probably already someone’s job. Enable one protocol, measure for a quarter, then decide about the second.

What to fix on your product pages first

Adobe’s finding that product pages are among the least readable page types for language models is the most actionable thing in the whole dataset. These pages are usually built for a human scanning a hero image and a price, with the real detail loaded by JavaScript, hidden behind tabs, or drawn into an image.

Work through these in order.

• Put the commercial facts in the HTML, as text. Price, currency, availability, size and colour variants, shipping cost, delivery window and returns window. If any of these render only after a script runs, or live inside an image, assume the agent cannot see them.
• Mark up every product with schema.org Product and Offer. This is the single highest-return change, because it converts your page from prose an agent must interpret into fields it can read directly.
• Fill the boring feed attributes. GTIN, brand, condition, material, dimensions and category. Agents use these to decide whether your product is even a candidate for the question being asked, long before ranking matters.
• Answer the comparison question on the page. Agents are asked to compare, so pages that state what a product is for, who it suits and what it does not do give the model something to quote. Marketing adjectives give it nothing.
• Check what your robots policy allows. If you block AI crawlers, none of the above matters. This is a genuine decision with a real trade-off, not an oversight to fix by default.
• Keep stock and price accurate to the hour. An agent that recommends an out-of-stock item once will discount your catalogue afterwards.

None of this is new discipline. It is structured-data work that has existed for a decade, applied to a reader that cannot guess.

The attribution hole agentic sales create

Here is the part that will cause the most internal trouble, and almost nobody is planning for it.

When an agent completes a purchase, the order lands in your system without the things your reporting is built on. No campaign. No channel. Often no session on your own site at all, because the discovery and the decision happened somewhere you cannot instrument. The customer’s email arrives with the order rather than before it, which means your usual sequence of capture, nurture and convert runs backwards.

The immediate consequence is that agentic revenue will look like direct traffic, and direct traffic is where credit goes to die inside most marketing organisations. The team that did the feed and schema work will get no attribution for the revenue it generated.

The second consequence is about the relationship. The shopper who bought inside a chat window will not go back to that chat window to ask where the parcel is. They come to you, with no prior contact and no account. Whether that becomes a repeat customer depends entirely on a post-purchase experience that was never designed to be a first impression.

Both of these are arguments for taking first-party data collection seriously at the point of order, rather than assuming you captured it earlier in a journey that no longer exists.

What this means for your team

Do the feed and schema work this quarter. It is small, it is owned by people you already employ, and it pays off in AI search visibility whether or not agentic checkout ever becomes large. Then enable one protocol, pick the one that matches how your customers buy, and give it a quarter of clean measurement before you argue about the second.

Do not build a budget line for agentic commerce yet. The traffic is excellent and the volume is small, and treating a high-conversion trickle as a scalable channel is how teams end up over-investing in a surface that has not settled.

For Indian brands there is a further wrinkle worth naming. Much of India’s product discovery already happens inside apps rather than on the open web, and quick commerce platforms are closed environments that open-web agents cannot read at all. If most of your volume comes through retail media and quick commerce, agent-readiness is mainly about your export and direct web business, not your domestic app sales. Do the work, but size the prize honestly.

The friction you should expect. The feed and schema work is unglamorous and invisible, so it competes badly against a campaign for the same engineering time, and it will slip unless someone owns it by name. Your analytics will not show you the return, because agentic orders arrive as direct traffic, which means the team doing the work cannot prove it worked. Expect to argue for it on reasoning rather than dashboards for at least two quarters. And when the volume does arrive, expect finance to ask why revenue is appearing in a channel with no cost and no campaign attached to it, which is a pleasant problem that still takes a week to explain.

People also ask

What is agentic commerce? It is a purchase where an AI assistant handles discovery, comparison and checkout on the shopper’s behalf, without the person visiting the retailer’s site to buy. OpenAI’s Agentic Commerce Protocol and Google’s Universal Commerce Protocol are the two open standards behind it in 2026.

Do AI shopping agents actually convert better? Yes, but read the number carefully. Adobe recorded AI-referred visitors converting 60% higher than other traffic in July 2026. That reflects unusually high purchase intent among a small group of visitors, not a property of the channel that will survive at scale.

What is the single most useful change to make? Add schema.org Product and Offer markup and make sure price, availability and shipping appear as text in the HTML rather than being drawn by a script.

Should I block AI crawlers instead? That is a real choice, not a mistake. Blocking protects content from being summarised without a visit, and it also removes you from the results agents shop from. Decide it deliberately, and decide it once, for the whole site.