The transaction hasn't moved. The decision already has.

The agent is already in your category

Last month, Meta launched a personal AI agent for everyone. The launch pitch was that it would switch your car insurance in five minutes. That's live today, in a category where the sale is worth hundreds of pounds and the comparison is so boring that people are delighted to hand it over.

If you sell anything people compare, whether it's insurance, cameras, supplements or trainers, an agent is already doing some of the comparing for your customers. The question is whether it's choosing you.

Meanwhile, the first checkouts failed

At the same time, the first attempts to let AI complete the purchase have struggled. OpenAI launched Instant Checkout in September 2025 and pulled it back by March 2026, saying it didn't offer the flexibility it wanted. Walmart reported conversion three times lower inside the assistant than on its own site. Stripe's own annual letter said agentic commerce had been "overhyped too early in some corners".

So two things are true at once, and most of the commentary only tells you one of them. The transaction hasn't moved. People still mostly buy on your site, your app or your retailer's shelf. But the decision about what to buy has already moved. It's being shaped upstream, inside an assistant, before the customer ever reaches you.

We've been here before

This isn't the first time customers have handed the comparison to a machine. In the UK, the first car insurance comparison site launched in 2002. Comparison sites trained a whole generation to let a machine produce the shortlist while a human makes the final choice.

AI agents inherit that habit. The difference is reach and method. A comparison site only knew what insurers fed it. An agent reads the whole web, decides which sources it trusts, and writes its own answer. Most brands have no idea what it's reading.

Four engines, one question

So I tested it. I asked the same buying question to Google AI Mode, Perplexity, ChatGPT and Claude, the way a real customer would phrase it. Four engines produced eleven answers between them, and they didn't agree.

More interesting than the answers were the sources. The recommendations were built on an affiliate blog, trade reviews, a comparison site and, in one case, a competitor's own website. Not one of the recommended brands had written the evidence that got it recommended. One affiliate blog effectively wrote most of two engines' answers.

You are being recommended on evidence you didn't write and can't edit. That's uncomfortable. But you can influence what the machines read, and most brands haven't started.

Six things you control

1. Open the door. Check your robots.txt and your firewall settings. Many large brand sites block the crawlers that AI companies use, often by accident, through a security default nobody revisited. If the agent can't read you, it can't recommend you.
2. Be readable. Agents read text. If your key information sits behind JavaScript, inside images or in a sixty-page PDF, it may as well not exist. Put the facts that matter, prices, specifications, terms, comparisons, in plain, structured text on a page.
3. Win the badges. Agents lean on the sources humans already trust: independent reviews, consumer tests, trade awards, ratings sites. Those badges are the evidence. Earn them, and make sure they're visible and machine-readable on your own site.
4. Feed the sources. Affiliate sites, review blogs, comparison sites and forums are writing your recommendation for you. Make sure they have accurate, current information. Measure where you appear and where you don't.
5. Connect the family. If you own several brands or products, make sure the machines know they're connected. Consistent names, structured data, a clear corporate page. Entity data matters more than ad spend here.
6. Get inside the assistant. Amazon's Buy for Me, Google's Universal Cart, Microsoft's Copilot checkout. Pilot at least one. Learn what it asks for, what data it needs and how it presents you, before your competitor does.

Measure it, monthly

None of this works as a one-off project. Pick your twenty most important buying questions, the ones real customers ask before they choose. Ask them of the main AI engines every month. Record who gets recommended, which sources get cited, and whether you appear. It takes an hour and it'll tell you more about your future sales than most dashboards you currently look at.

Five things I'd bet on

• Through 2027, agents decide. They mostly don't buy.
• Agent payments reach consumer scale in the UK by 2028. Visa and Mastercard are already transacting.
• The rules get written in the next year. Regulators are moving faster than the market expected.
• Comparison sites are defending, not transforming. The layer above them is still open.
• Machine readability is the fight. Checkout isn't, yet.

What to do this week

None of this needs a budget. It needs a name against it. Put one person in your team in charge of what the machines read about you. Give them a month to run the twenty questions, fix the obvious blockers and report back with a before and after. You'll know more than most of your competitors by Christmas.