Agentic shopping: the shops I no longer see

November 4, 2026 · 7 min read

The white front of De Drie Dorstige Herten, an old house with dark blue window frames and a dark blue door under the word Proeflokaal, two green lanterns beside it, and a sign with the bar's crest hanging on the left.

Over the past few weeks I bought four things online with the help of Claude, my AI assistant, and hardly saw any of the shops.

The first was a pair of T-shirts from Labfresh. I was only doing research, but it went so smoothly that I bought them along the way. I sent Claude a screenshot of my previous order:

Part of my earlier order confirmation from Labfresh, in Dutch: a Relaxed T-shirt in light blue and one in light grey, both in size M, at 59 euros each.

Claude found the new version of that T-shirt. I picked two other colours, and it sent me one link with both in the basket. I pressed it, signed in with Face ID to Shop, Shopify’s app for shoppers, and paid with iDEAL.

A few days later I needed new blades for my razor. I asked Claude for the best deal, and it came back with a link to a drugstore where I have no account, with the blades in the basket. I checked out as a guest on my iPhone, and its autofill filled in my email address and delivery address.

Since then, I have also bought diapers at bol.com and coffee beans at Peeze, a roaster I often buy from.

In my first post on agentic shopping, Claude only added one product to a basket. My trust in AI is growing faster than I expected.

All four purchases have something in common: they are things I buy again and again. I know the products, so for most of them the only question is where and for how much. For a handful of things I buy again, an agent makes the web shop almost invisible.

So my question is: if an agent does the searching for repeat purchases, what does a web shop still need to be good at?

How Claude shops for me

Behind that one chat, Claude takes three routes:

MCP is a standard way for an AI agent to work with a system directly, through a fixed set of tools: actions the system offers, such as “search the catalog” or “create a basket”, that the agent can use.

The shirts came through that last route. Labfresh runs on Shopify, and on 13 October 2026 Shopify opened one MCP server for all the shops it runs: Shop MCP. Put together, it looked like this:

How I bought two T-shirts through Shop MCP. In the Claude app on my phone I send Claude a screenshot of my earlier order, and get one basket link back. Claude, in the cloud, sends a search to Shop MCP, which I use signed out, and gets products with links back. I press the link myself, land on Labfresh's checkout, sign in to Shop with Face ID and pay with iDEAL. The tools I allowed in Claude: searching the catalog, looking up a product and getting its details, always; everything else, such as the basket, checkout and my account, never. No sign-in, so Claude never saw my account, my orders or a payment method.

For other shops, Claude follows the instructions in my shopping project, including a short list of shops I trust.

What the shop no longer gets to show me

What comes back from Shop MCP is not a web page but plain information: the product, its price, whether it is in stock, the seller and a link. No banners, no offers of the week. When I open a shop’s app myself, those come first.

The banner at the top of the bol.com app, in Dutch: Aanvuldeals, top up your supplies at an extra discount, with products such as dishwasher tablets, coffee and toilet paper, and the labels Topdeal and Slim inslaan.

I understand why they are there, but they are not why I am there. The promotions only came back when I pressed the link and landed on Labfresh’s site: before I paid, it suggested socks and another shirt.

It reminds me of the sweets at the supermarket checkout. By the time you get there, you are tired from shopping, and with kids even more so. That is exactly when you, or your kid, fall for a chocolate bar. Online shops use the same moment. Claude is not tempted the way I am, at least for now: here the offers did not even reach it.

What shopping through Claude gives me

Three things stand out:

  1. Searching and comparing in one place. No Google search first, and no list of shops to click through one by one.
  2. No promotions I did not ask for. I skip the shop’s app and its offers, which takes less of my attention.
  3. One interface, so life gets simpler. I share a photo or the name of a product, look once at the best options Claude finds, and say go.

Is it faster? Only when there is something to compare. For the diapers, I asked which shop had the best offer, Claude added the pack to my basket at bol.com, and within 30 seconds of opening the app I had paid.

For now, the checkout is the one moment I still see the shop. I do not think it will take long before I feel confident enough to let Claude pay as well, when I have a payment card specifically for agentic shopping. Then even the checkout no longer matters.

What I tell Claude also stays in one place. If I want to spend less for a while, I say so once and it counts for every shop, instead of sharing that budget with every shop separately.

This is moving fast

For now, I expect only a few people let an AI assistant shop for them. But there are many signs that this will change, and quickly. Cloudflare wrote this summer that, for the first time, more than half of internet traffic is no longer human, ordinary crawlers included. In September, Meta launched Muse in the United States, an AI agent that can also shop.

Shops are adapting too. Some rewrite their product descriptions so that AI agents find and recommend them, which is called generative engine optimisation (GEO).

What changes for a web shop

From the web shop’s side, the drugstore stands out most. Its user experience did not matter to me, and the site gave me no feeling about the brand. It was a transaction: the product was there, the price was right, and I could pay as a guest. It could just as well have been another drugstore: the name was just text on my screen.

The coffee beans showed the other side. Claude put 2 kilos of Amora beans in my basket at Peeze, where the browser on my laptop is signed in. It could not send me a link, since the basket was in my account, so I paid on Peeze’s website on my phone. Claude added nothing, and I can see why: only Peeze sells these beans, and I know exactly which ones I want. There was nothing to compare, unlike the diapers and the blades, which several shops sell.

So perhaps the product matters more than the shop. A product nobody else sells is a reason to go to that shop, with or without an agent. Supermarkets have known this for a long time: Albert Heijn puts a lot of work into its own brands, such as AH Terra and AH Excellent.

I worked at Albert Heijn online from 2017 to 2019, and I saw up close how much work goes into a shop’s website and app. If agents do more of the searching for repeat purchases, I wonder which of that work still pays off, and what becomes more important instead. From my purchases: a product nobody else sells, product information an agent can read, the price, for now a checkout that works without an account, and trust.

That raises a question I cannot answer yet: what does this mean for small shops, or for someone starting a web shop today? I trust the shops on my list because I already know them, and most of them are large, with good customer service. My own instructions make it harder for a shop I do not know: it has to be at least 10% or €5 cheaper, and pass a trust check, before Claude even suggests it. I am curious how people in retail see this.

Next: the weekly groceries

I expect groceries to be different. A weekly basket at Albert Heijn holds 30 to 50 products, and part of it changes all the time: I do not want the same dinner every week. Context could matter much more there too, such as friends with an allergy coming for dinner on Thursday.

To find out, I asked Claude to go through my last ten orders, with two hypotheses:

  1. About 80% of what I buy, I buy again within three weeks.
  2. Most of those repeat purchases are not specific to Albert Heijn: any supermarket sells them.

The first results are interesting. I am looking forward to continuing the research, and to sharing it here.

▸Behind this post: 54 corrections, 49 of them caught by me

This post went through many rounds between me and Claude, over two weeks in one conversation. We kept a log of every correction: what went wrong, and who caught it.

  • Clarity. 26 of the 54. A first draft that was a bit of everything, a turn to the shop's promotions that was never announced, and "it started again" without saying what.
  • Not landing with my readers. 10. A first sentence that assumed readers had read the first post, an explanation of Shopify's layers that took more attention than it deserved, and a post that ran to nine minutes before it was cut to seven.
  • Wrong, but plausible. 9. The wrong years at Albert Heijn online, a coffee purchase described before I had made it, a launch date taken from the news instead of the source, selling points "written for agents" that are on the website too, and a preview Claude called live while it was still building.
  • Not sounding like me. 5. "Really", "so fast", being tempted by socks I did not want, and more about my family than I wanted to share.
  • The rest. Three diagram fixes, two of them caught by Claude in its own screenshots, and one about the process: we did not agree on the focus before writing.

Most corrections were about what the post was trying to say, not about sentences. It went through three storylines before one sentence held it together, so we now agree on that sentence first.

The log is kept by a Claude Code skill: a set of instructions Claude follows at the end of a session. It sorts every correction by cause and records who caught it. If you write with AI as well, you are welcome to read and reuse the meta-learnings skill.

The numbers for every post are on Behind the posts.

No sponsored or affiliate links. If something here is new to you, the link is an easy way in.