7 min readEnglish and Spanish
Will Agents Learn Every Website, or Will Websites Learn Agents?
Two bets on AI agents that shop for you: make the agent good enough to use any site as it is, or make every site easy for agents. I went through the 2026 numbers, including the ones that argue against my own product.
Yesterday I wrote about Bubblio's door for AI agents, and how nobody walked through it. It left me with a bigger question I can't stop thinking about.
There are two ways AI agents end up buying things for people.
Bet 1: the agent learns every site. Make the agent so good that it can use any website as it is. It reads the page, closes the cookie popup, picks the size and gets through checkout. The site owner does nothing. That's OpenAI's agent, Perplexity's Comet, Meta's Muse, and the browser agents behind them.
Bet 2: the site learns agents. Give agents a proper entrance: a protocol, an API, a product feed, a checkout built for software. That's the Agentic Commerce Protocol from OpenAI and Stripe, the Universal Commerce Protocol from Google and Shopify, the agent programs at Visa and Mastercard, and, at my much smaller scale, Bubblio.
I build for bet 2. So I went looking for the numbers that could tell me I'm wrong.
First, is there even a market?
Everyone agrees there will be one. Nobody agrees on its size.
| Who | US shopping by agents, 2030 | What they count |
|---|---|---|
| Morgan Stanley | $190B to $385B (10 to 20% of online sales) | Purchases agents complete |
| Bain | $300B to $500B (15 to 25%) | Purchases agents start, influence or complete |
| McKinsey | $900B to $1T | Retail revenue agents "orchestrate" |
Forecasts are cheap. What's happening today is more convincing. Adobe tracks visits to US retail sites that come from AI assistants. In late 2024 those visitors bought less often than everyone else. By May 2026 they bought 54% more often, and spent 53% more per visit. On November 3, Shopify said AI traffic to its stores grew 2.6x in a year, and orders from that traffic grew 3.3x.
So people already shop through AI. The open question is who does the last step: the part where money moves.
The case for bet 1: agents got good, fast
Online-Mind2Web is a test of 300 real tasks on 136 live websites: find this, book that, fill this cart. The paper that introduced it in April 2025 was called An Illusion of Progress?, because agents were doing far worse than the demos suggested. A browser agent on GPT-4o scored 30% in late 2024.
By March 2026, two separate teams that published every single run scored 90%.
The traffic followed. Cloudflare said in September that daily requests from AI agents on its network grew more than 1,700% in a year, and that more than half of the traffic reaching its customers' sites is no longer human.
The money is starting to follow too, and it's going to agents that browse. Stripe's wallet for agents now powers personal agents like Meta's Muse, Grok Bot and Instinct, and Stripe says agent purchases through it grew 38x in a single month this fall. Look at what it ships: help when the agent hits a 3D Secure check, help when a card is declined, a way to raise the approved amount when shipping gets added at the end. That's tooling for agents getting through ordinary checkouts. It's infrastructure for bet 1.
Even the courts moved. Amazon sued Perplexity over Comet in November 2025 and won an injunction in March. In August an appeals court overturned it. As reported, the court's view was that an agent acting only when its user tells it to is an extension of that user, not a trespasser.
The case for bet 2 is thin, so far
In September 2025 OpenAI launched Instant Checkout: buy without leaving ChatGPT, built on a new open protocol with Stripe. That was the strongest version of bet 2 anyone could build. The biggest AI app, the biggest online payments company, Etsy on day one, and more than a million Shopify stores promised next.
By early 2026, about a dozen Shopify merchants had actually gone live. In March, OpenAI ended it and went back to sending shoppers to the store's own site. Walmart said products bought inside the chat converted three times worse than ones where people clicked out to walmart.com.
Visa has more than 100 partners in its agent program. The last real-world count it shared publicly, in December 2025, was "hundreds" of agent transactions.
That's the lesson my door taught me, at a much bigger scale: sites don't build entrances for traffic they can't see yet.
Google and Shopify launched the Universal Commerce Protocol in January, with more than 20 big retailers and payment companies behind it. That one has a real chance, for a reason I'll get to.
The player I forgot: the walled store
My two bets had a gap. The biggest store in the US wants neither.
I opened amazon.com/robots.txt today. ChatGPT-User, Claude-User, Perplexity-User, GoogleAgent-Shopping, Manus-User and Meta's agent are all barred from the whole site. In September Amazon also cut Meta's Muse off from buying there. Meanwhile its own assistant, Rufus, had more than 300 million users in 2025, and Amazon credits it with about $12 billion in extra yearly sales. Its "Buy for Me" feature even buys from other brands' websites on your behalf.
So Amazon's answer to "agents learn sites, or sites learn agents" is: my agent learns my site, and everyone else waits outside.
What shoppers actually want
The people with the money are less sure than the people building for them. Worldpay asked shoppers in seven markets what they'd let an agent do on its own. Most were fine with research. Very few were fine with paying.
In the UK, 6 in 10 people told YouGov they'd stop using an AI shopping agent after a single mistake.
Put that next to Walmart's number and Adobe's number and a pattern shows up. People are happy to let AI find the thing. They want to buy it themselves, on the store's page, with a moment to say yes.
So who wins?
Here's my read of the numbers.
- Reach goes to bet 1. Integrations only cover the top of the market. The long tail will never write one: OpenAI couldn't get more than a dozen stores to turn on a checkout that cost them nothing to join. Meanwhile, agents that use sites as they are already handle 90% of a hard benchmark, and get better every few months without asking anyone's permission.
- Bet 2 only wins when it's free for the owner. UCP matters because Shopify can switch it on for millions of stores at once, and the owner writes nothing. A protocol that every site has to build by hand loses to an agent that just reads the page.
- The fight is over the last step. Finding things is close to solved. Paying isn't. DataDome found that 80% of AI agents don't properly identify themselves, so a site can't tell your agent from a scraper holding a stolen card. Whoever owns identity and approval at checkout gets the money: card networks, agent wallets, or the platform.
- The biggest stores keep the gate shut and run their own agent, at least until outside agents bring them more customers than they cost.
So if I have to pick one: the agents that learn every site win the browsing. The sites still win the checkout, but only the ones that make it cheap. Nobody is going to sit down and write a list of actions for agents.
What this means for what I'm building
It's a bit uncomfortable to publish this a day after shipping a site-side product. But the numbers mostly agree with where the desk already went: don't ask the owner to write anything, learn from the agents that already visit, and save the owner's attention for the one step where money moves.
My August door asked the owner to do the work up front. The data says that's the losing shape, for me and for OpenAI.
If the desk turns out wrong too, it'll be because agents got so good at ordinary checkout pages that even the last step didn't need one. I'll keep counting.
Sources
- Forecasts: Morgan Stanley, Bain, McKinsey
- AI-referred retail traffic: Adobe Q3 2026 AI traffic report (PDF)
- Shopify Q3 2026: AI orders grew 3.3x
- Agent benchmark: Online-Mind2Web paper, leaderboard with sources
- Agent traffic: Cloudflare, "The Internet has a second audience", DataDome on agent identity
- Stripe: Link's wallet for agents
- Amazon and Perplexity: the injunction, the appeal
- Instant Checkout: what went wrong (Modern Retail), about a dozen merchants live (Rye)
- Universal Commerce Protocol: Shopify engineering
- Amazon: Rufus and $12B, Muse blocked, robots.txt (read 18 Nov 2026)
- Shoppers: Worldpay Agentic Commerce Report 2026, ACI and YouGov
- Visa: agent payments in 2026, by the numbers
- My own door, and why nobody walked through it: the previous post