Agentic checkout is live. Yet Almost nobody is using it

19 min read

Editor’s note: The numbers in this piece come from Adobe Analytics, Checkout.com, Shopify, Salesforce, court filings, and public announcements. We’ve flagged vendor data throughout so readers know when a company is quoting its own numbers.

TL;DR

• Merchants in the UK and US report that AI agents are involved in roughly 3% of transactions today, while 89% say they are actively preparing for agentic commerce.
• Nearly a quarter of consumers say they will never delegate a purchase to AI, and 27% trust no organisation to operate a shopping agent for them.
• OpenAI started moving Instant Checkout toward merchant apps in March 2026, about six months after launching it.
• The referral side went the opposite way. Adobe recorded AI referred retail visitors converting 54% better than non AI traffic in May 2026, after converting at roughly half the rate a year earlier.
• Shopify reported AI referred orders growing close to 13x year on year in Q1 2026, with more than half of those sessions landing straight on a product page.
• Roughly a third of the content on an average retail product page is unreadable to large language models.

Let’s start with the buy button

In September 2025, OpenAI put a purchase flow inside a chat window. Instant Checkout launched with US Etsy sellers, and OpenAI named a million Shopify merchants including Glossier, SKIMS, Spanx and Vuori as next in line. The plumbing was the Agentic Commerce Protocol, an open standard co-developed with Stripe.

Perplexity followed in November, opening its PayPal powered checkout to free US users after a year of restricting it to Pro subscribers. Microsoft arrived at NRF in January with Copilot Checkout, auto enrolling Shopify merchants. Google announced the Universal Commerce Protocol the same week, co-developed with Shopify, bringing purchases into AI Mode and Gemini.

So, four surfaces with four payment stacks. An enormous amount of conference stage time and a great many slides with the word agentic on them.

Then somebody did a study. Checkout.com commissioned Censuswide to survey 12,005 consumers across the UK, US, Brazil, China, France and the UAE, along with 400 heads of payment at consumer facing merchants, with fieldwork in early March 2026. Those merchants put AI agents at 3% of transactions. Eighty-nine per cent said they were preparing anyway.

Hold those two numbers next to each other and you have the shape of the market as it currently stands: an expensive readiness programme aimed at a sliver of volume.

The consumer hesitation is specific rather than vague. 24% of respondents said they will never delegate a purchase to AI. 27% said there is no organisation they would trust to run a shopping agent on their behalf. It’s important here to point out that people did not refuse the assistant itself, only the idea of handing it the card.

What happened in March 2026

On 5 March 2026, Digital Commerce 360 reported that OpenAI was sidelining Instant Checkout in favour of purchases through retailer apps inside ChatGPT. A spokesperson said the company was evolving how it approaches commerce in ChatGPT. Reporting on the shift put the number of live merchants at around a dozen, and narrowed the future scope to a small set of large integrated retailers such as Target, Instacart, Expedia and Booking.com.

Checkout.com, which sells into this category and has no obvious reason to talk it down, read the move as structural rather than directional: AI handles discovery and intent, merchants keep control of checkout. This is exactly the opposite of the story most vendors told in late 2025.

Five days later, a federal judge added a second constraint. In Amazon’s suit against Perplexity, Judge Maxine Chesney granted a preliminary injunction blocking the Comet browser agent from password protected areas of Amazon’s site. The finding that will get cited for years: Comet accessed accounts with the user’s permission, but without authorisation from Amazon.

Your customer wanting an agent to shop for them, in other words, does not by itself give that agent the right to operate on a merchant’s systems. Agent access is now something to be negotiated instead of taken for granted.

Four rails, one shelf

This is why betting on a single checkout integration is a poor use of a quarter. There are at least four live standards, backed by companies that compete with each other, and none has won.

Mastercard’s position, stated in January 2026, is that it participates in all of them. That is the tell. When the networks refuse to pick, the merchants underneath them do not need to either.

Every one of these standards, though, sits on top of the same requirement: a structured, accurate, machine readable product feed. OpenAI’s own documentation says merchants must supply a regularly refreshed CSV or JSON feed covering identifiers, descriptions, pricing, inventory, media and fulfilment before anything else works. Google is asking for the same thing through Merchant Center. Microsoft through Microsoft Merchant Center. So even though the checkout buttons differ, the shelf underneath is shared.

So where did the money move?

While the checkout story stalled, something less photogenic happened to the traffic.

In March 2025, visitors arriving at US retail sites from AI assistants converted 38% worse than everyone else. By March 2026 they converted 42% better. By May 2026, per Adobe’s most recent read, 54% better, with 53% higher revenue per visit and 23% more pages viewed. Adobe bases this on more than a trillion visits to US retail sites.

A sign flip that complete, in twelve months, in a channel that large, does not happen often.

Two buts before you quote this internally. Adobe sells an LLM Optimizer product, and this data was published alongside it. The comparison is against non AI traffic in aggregate, which blends paid search, email, affiliate and organic, so it is not a clean head to head with Google.

Which is why the second dataset matters more than the first. Shopify’s Q1 2026 commerce data, gathered independently across a different merchant base, found AI referred shoppers converting close to 50% higher than organic search visitors with 14% higher average order values. Referral sessions from AI chatbots grew more than 8x year on year and, more importantly, orders from them grew nearly 13x.

The mechanism sits in one number from that same report: more than half of AI referred sessions start directly on a product detail page, against roughly 20% for organic search. The assistant already did the comparison shopping. The visitor lands pre-qualified, at the bottom of a funnel you did not have to build.

Salesforce, measuring a third population of 1.5 billion shoppers, put AI and agent influence at 20% of global retail sales over holiday 2025, worth around $262 billion, and found retailers running their own shopper agents grew sales 59% faster than those that did not.

Three vendors, three methodologies, three commercial interests in the answer, and they agree on the same direction and magnitude.

The visibility gap that needs fixing

Adobe also ran its Content Visibility Checker across the US retail sector, scoring how much of a page an LLM can actually read. Homepages averaged 75%. Category pages, 74%. Individual product pages came in at 66%.

So the page most likely to convert is the page least likely to be readable by the channel converting best. The best performing retailers scored 82.5% on homepages. The worst, 54.2%.

The cause is usually architecture rather than laziness. A decade of headless and composable frontend work produced sites where the base HTML is a shell and everything meaningful (reviews, specs, FAQs, social proof) hydrates afterwards through JavaScript. Human browsers wait for it. Most generative engines take the base HTML and leave.

By category, Adobe’s May data put cosmetics at 63% readable, electronics at 56%, sporting goods and apparel at 51% each.

The Surface Audit Framework: five steps

Call this the Surface Audit because it treats every assistant as a shelf you either appear on or do not, and because it deliberately avoids committing to any one checkout rail. Five steps, in order. None of them requires a protocol decision.

1. Score what the machines can actually read

Why: a third of your product page content may be invisible to the systems doing the recommending. You cannot optimise what you have not measured.

How: take twenty product pages spanning your best and worst sellers. Fetch each with JavaScript disabled and read what comes back. Anything missing from that raw HTML (specs trapped in images, reviews loaded by a third party widget, sizing tables rendered client side) is missing from the assistant’s view of you. Score each page, record the baseline, and hand engineering a ranked list.

What to check on each page, and what it costs you when it is missing:

2. Fix the feed before the frontend

Why: every standard on the table (ACP, UCP, Copilot) reads from a structured product feed. Feed work is the only investment here that pays regardless of which rail wins.

How: get identifiers, pricing, live inventory, fulfilment options and media complete and refreshed daily. Then add the attributes assistants use to disqualify you: what the product is not for, compatible accessories, materials, care requirements, return window. Google has started accepting answers to common product questions as feed attributes in Merchant Center. Treat that as the direction of travel.

Sequence it so the cheap work happens first:

3. Instrument the channel honestly

Why: if AI traffic is buried in Direct and Referral, no one in your finance meeting believes the conversion numbers above.

How: GA4 added a native AI Assistant channel on 13 May 2026, with no configuration required. Use it, then note its three limits out loud in your reporting. It counts forward only, so your earlier trend stays buried. Google’s own AI Overviews and AI Mode clicks land in Organic Search. And a large share of assistant traffic arrives with no referrer at all, which means Direct. Keep a parallel custom channel group running, and treat every AI number as a floor.

Agent traffic distorts paid media reporting in much the same way, which we looked at separately in advertising in the age of AI agents.

4. Publish the content to get cited instead of ranked

Why: assistants cite comparison and buying guide content at higher rates than product pages, because that is the format that answers the question being asked.

How: write the comparisons you have been avoiding, including the ones where a competitor wins on a specific dimension. Open each section with a declarative answer to the question in the heading rather than a wind up. Add the specificity that lets a model match you to a narrow query, price bands, use cases, who the product suits badly. Vague superlatives give a model nothing to work with.

Four formats worth building first, in order of how often assistants reach for them:

• Head to head comparisons, including against competitors. The version where you lose on one dimension is the version that gets quoted, because it reads as evidence rather than marketing.
• Buying guides organised by constraint, not by product. Budget, room size, skin type, skill level, whatever your category’s real filter is.
• Compatibility and fit pages. Explicit lists of what works with what. These answer a query type nobody writes content for and everybody searches.
• Who this is wrong for. The single most citable paragraph you can write, and almost nobody publishes it.

5. Take the cheapest checkout integration available, and no more

Why: at 3% of transactions, a bespoke integration cannot earn back a quarter of engineering time. But refusing to appear at all costs you the option.

How: if you are on Shopify, most of this is a toggle in Admin rather than a project. Microsoft auto enrols Shopify merchants in Copilot Checkout with an opt out. If you are not on Shopify, route through PayPal or Stripe rather than building against a protocol directly. Check the economics before you switch anything on: trade press has put OpenAI’s merchant fee at 4% of completed purchases, on top of normal processing, though OpenAI itself has only ever described it publicly as a small fee. Perplexity’s merchant programme has run at zero commission.

Before switching anything on, get answers to these four. They decide whether the integration is worth having at all:

1. Who is merchant of record? This determines who owns the customer data, who handles the refund, and whose name appears on the statement.
2. What is the total take rate? Platform fee plus your normal processing. Compare it against your contribution margin per order, not against your revenue.
3. What data comes back with the order? Email consent, marketing permission and returning customer identity are the three that determine whether this is a sale or a relationship.
4. How do you turn it off? Auto enrolment is common. Know the exit before the entrance.

Where this thesis could be wrong

Microsoft’s numbers say in-chat completion works. Microsoft reported that shopping journeys involving Copilot are 33% shorter than traditional search paths, produce 53% more purchases within thirty minutes, and are 194% more likely to end in a purchase when intent is present. Copilot apps have more than 100 million monthly active users. If those figures hold across a wider merchant base, the 3% ceiling rises fast.

Consumer intent is running ahead of consumer behaviour. A third of respondents in the Checkout.com study expect at least 10% of their purchases to be AI driven within a year, and 72% of merchants think consumers will adopt faster than merchants are ready for, though stated intent has never been a reliable predictor of what people actually do.

Google has not moved yet, and Google is the one that matters. Universal Cart is rolling out across Nike, Sephora, Target, Ulta, Walmart, Wayfair and Shopify merchants including Fenty and Steve Madden, announced at Google Marketing Live 2026. Google can absorb years of unprofitable agentic commerce in a way OpenAI cannot. If checkout inside AI Mode gets genuinely good, the distribution behind it is a different order of magnitude to anything discussed above.

None of which changes step one. Every one of those scenarios still requires that a model can read your product page.

Your first 90 days

1. Weeks 1 to 2. Run the readability score on twenty product pages. Confirm the GA4 AI Assistant channel is reporting and annotate the date. Establish your baseline before anyone changes anything.
2. Weeks 3 to 6. Fix feed completeness first, server rendered product content second. Ship the two comparison pieces your category most obviously lacks.
3. Weeks 7 to 12. Switch on whichever checkout integrations are a configuration change rather than a build. Tag those orders separately. Re-score the same twenty pages and report the delta, not the absolute.

Then wait ninety days before drawing any strategic conclusion from the numbers. The channel is moving too fast for a four week read to mean anything.

FAQ