I build selling agents for webshops, so you know where I stand before you read this. Most writing on agentic commerce is either a forecast about 2030 or a checklist about product feeds, and neither tells a merchant what to do differently on Monday.
Agentic commerce is shopping where an AI agent such as ChatGPT, Gemini, or Perplexity's Comet researches, compares, and shortlists products for someone and increasingly buys on their behalf. The person still decides what they want. The agent does the shopping.
In 2020, the COVID-19 pandemic hit. Lockdowns began, and physical stores were cut off overnight. We remember it as the year shops went online or died, but the real test was much narrower than that. A shop survived if its products could be understood by someone who couldn't touch them, smell them, or ask the person behind the counter anything. That meant photos, descriptions, sizing tables, and honest delivery promises. The product had to survive being translated into a webpage, and plenty of good products didn't.
The same test is running now against a different reader. A machine opens your page, pulls it apart, sets it against forty others, and tells its user about three of them in roughly three sentences.
2020 caught us off guard, but we can see this shift coming. Every time an AI agent researches your industry, it learns which stores give the best answers, and those early favorites become its go-to choices. You won't get locked out all at once. Instead, recommendations will quietly form without you, and your standard reports won't show that you're missing.
The checkout hasn't moved yet. The decision has, and the decision was always the part that mattered.
Agents research, compare and shortlist at volume right now. They don't buy autonomously at any real scale yet. But a shopper who asks an agent for three good options and gets three good options buys one of those three. The click still happens on your site with a human doing the clicking, and in your dashboard it looks like any other sale. The choosing happened twenty minutes earlier, in a conversation you weren't part of and can't see.
The Dutch numbers are enough to take that seriously. Half of Dutch consumers now use generative AI while shopping, up from a third eighteen months earlier (Q&A Retail, May 2026), and among 18- to 34-year-olds it is 83%. The traffic those conversations produce behaves unlike anything else in your analytics: Adobe measured AI-referred visitors converting 42% better by March 2026, with 37% more revenue per visit.
What interests me more is what happens when one of those conversations goes badly, because you see nothing at all. Roughly 70% of visits arriving from AI tools land in analytics as direct traffic, since no referrer gets passed. That's the visible half of the problem. The invisible half is a conversation that ends with the agent telling someone to skip your shop because the return policy was unclear. There is no visit to misclassify. Nothing bounces, nothing gets abandoned, no internal search comes back empty. Every problem a merchant has been trained to spot leaves a mark somewhere in a dashboard, and this one doesn't leave anything.
An agent has to tell its user why it picked what it picked, and I think that single requirement changes selling more than all the feeds, plugins and integrations everyone is busy building.
Online selling has spent twenty years optimising for attention: the image that stops the scroll, the urgency banner, the ad that finds an insecurity at eleven at night. None of it lands on a reader with no urgency and no insecurities. A team from Columbia Business School, Yale and MyCustomAI audited shopping agents in a controlled storefront and found their sensitivity to price, ratings and reviews varies sharply from one model to the next. What I'd add from watching them work is that they explain themselves. Ask ChatGPT for a recommendation and it tells you why, every time. If the reason your product is better can't fit in a sentence the agent would repeat out loud, you've already lost the sale.
Current evidence is limited. Agents can explain their product choices, but different models carry distinct biases and aren't purely rational. More importantly, no one has actually tested whether emotional marketing still sways them.
Yes, but on one condition.
Start with the case against me. These models carry a bias toward big familiar brands, and a 2026 study tested exactly that on skincare across three of them. Where two products looked the same on paper, the known brand won every time. The authors called it a conditional monopoly. Multiply that across everyone asking the same question and the same few names come back every time, so agent shopping funnels buyers toward the brands that were already winning.
The condition sits in the same study. A small documented quality advantage was enough to break the monopoly. The brand name only decides when nothing else separates the products. Give the model something concrete to compare and it compares.
For two decades, whoever paid got seen. Ad budget, retargeting, placement fees. A ten-product shop could hold the best answer in the country for fragrance-free, pregnancy-safe, under forty euros, delivered tomorrow in the Netherlands, and never sell one of them, because a human shopper reads the first handful of results and stops.
An agent reads far more than a handful. It can work through hundreds of products and check each one against the whole constraint list, which makes sitting at the top of the page worth less than having an answer that survives the checklist. It is also unimpressed by paying. The same study found agents consistently marking down products tagged as sponsored while rewarding platform endorsements. Placement still counts for something, since position bias is real, but it stops being the thing money reliably buys.
So the gate moves from money to work, and the work is making a real advantage legible to a machine, limits included, since a limitation you leave out doesn't stay hidden. It turns up in a review, a spec sheet, or the returned parcel. A ten-person shop can do that work. None of this makes the competition easy, and shops that do nothing will lose to the defaults faster than they ever lost to Google, but effort is a fairer gate than money.
If you want to see what agents currently make of your store, we built a free snapshot for WooCommerce shops. Give us your URL and we'll show you how AI buying agents read your catalog today.
No, and here I disagree with most of the industry. The standard advice says your product feed is your storefront now, so fix the structured text, ship the JSON, and you're done. Structured data is a must-have, not the finish line. Emotion doesn't leave when an agent arrives, it turns into data.
Google says more than one in six queries in AI Mode aren't text, and image-started searches have grown over 40% month on month since launch. People are showing the machine what they mean because typing it was never how they held the preference in the first place. Taste in a dress or a bottle has structure even when the person can't state it, and a model that sees what someone saves, compares and sends back can learn that structure.
My guess is that this ends with someone sketching roughly what they have in mind and the agent finding the closest real product. Treating visual media as core product data instead of window dressing gives feeds a level of readability that text alone cannot match.
It collapses, and the warehouse notices before marketing does. What used to take a week of tabs, reviews and second thoughts now fits inside one conversation, and it happens somewhere you don't get to see.
That cuts both directions at once. A shop that delivers well grows faster than its own planning, because buyers who used to leak away over three days now arrive decided. Those same buyers arrive with expectations already fixed about stock and delivery, and if your page implied Thursday, the agent said Thursday, and Thursday is now a promise somebody in your warehouse has to keep.
Since 19 June 2026, every shop selling to EU consumers has to carry a withdrawal button that cancels an order inside the 14-day cooling-off period in one click, with no login and no phone call. The ACM enforces it here in the Netherlands, and a shop that gets the button wrong can find the withdrawal window stretched to twelve months.
Three things can go wrong once an agent is doing the choosing.
The first is a page that overstates. It still closes the sale, and it still comes back, because the person opening the box is human. An overstated claim completes the sale but hands back a refund, shipping in both directions, handling, and payment costs.
The second is a page that is accurate but incomplete. If your product page never mentions the ingredient a buyer told the agent to avoid, the agent recommends it in good faith and the return arrives anyway. Silence reads as absence to a machine, which makes stating your limits a commercial decision rather than a moral one. A page that says plainly who the product is wrong for gets recommended to fewer people, and the ones you lose are the ones who would have sent it back.
The third is the most painful. The page is correct, complete and clear, and the agent gets it wrong anyway. You made no mistake and you pay for it regardless: refund, return shipping, handling, payment costs, and a customer who files the disappointment under your name rather than under the tool that made the error.
Whichever of the three it is, the return comes from a human. Agents don't send things back. The person who does has never read your product page. They read a summary of it, written by something else, and they are holding what arrived against what they asked for rather than against anything you wrote. Every careful sentence on that page was aimed at a reader who never saw it.
No matter what goes wrong, you pay for it. The work is making sure the agent gets it right, which means answering its questions directly instead of publishing a page and hoping the answer is somewhere inside it.
I don't have evidence for this next part, only a gap that seems unlikely to stay open. Every protocol shipping now, Visa's Trusted Agent Protocol and Mastercard's Agent Pay among them, builds trust in one direction, proving to the merchant that the agent is genuine. None of them scores the merchant for the agent.
A bad fit costs you one return today. Under agents, the reason for the return becomes structured data: wrong size chart, misleading claim, delivery three days late. I can't see the engineering reason that a record would stay locked inside one conversation. Once it moves, agents will carry something like an opinion about your shop, shared the way review scores are shared now. Anything you do to game a shopping agent is being written down.
Nothing here is finished. The protocols are still being argued over, the tooling is young, ours included, and anyone selling you a settled standard is guessing. That's a fair reason to be sceptical about any particular product. It isn't a reason to wait, because the case for starting doesn't depend on the standards settling at all.
The usual pitch is to spend now for a payoff in 2028. I don't think that's the shape of it. Answer engines are shaping purchases this quarter, in the numbers above: half of Dutch consumers using generative AI while they shop, 83% among under-35s, and Adobe measuring those visitors converting 42% better with 37% more revenue per visit. Whatever you do to make your catalog readable by a machine gets paid against that traffic, not against a forecast. A shop that answers an agent's questions precisely gets picked more often in conversations happening this week.
The same work lands on your human visitors, who are still most of your traffic and still get stuck on the same unanswered questions. A shopper on your site wondering whether the serum is safe during pregnancy needs the same answer the agent needed, and neither of them will hunt for it. Answer it once and both readers are served, which is why this holds up even if autonomous checkout takes another three years.
Then there's the part that only accumulates. Every conversation an agent has with your shop produces something you have never had before: what a real buyer asked for, which constraint decided it, and the exact points where your catalog had no answer. Your order history only ever told you what sold. It never told you why, and it never told you about the people who left without buying because nothing you sell was described in a way that matched what they needed.
Ad spend stops working the day you stop paying for it. This doesn't. The software you run will be replaced. The record of what buyers asked your shop won't be, because it describes your catalog and your customers rather than any vendor's product. A shop that starts this year spends it collecting the questions that keep coming back, the constraints that decide the sale, and the gaps where it had no answer. None of that is training ChatGPT. It is your own agent getting better at answering on your behalf, and your catalog getting fixed where the record shows it falls short. A shop that starts in 2028 starts with none of it, and the gap widens instead of closing, because the shop with the record answers agents better, gets chosen more often, and learns more from being chosen.
None of that asks you to believe a forecast. The conversion numbers are from last March, and the record starts accumulating the first week you switch something on.
Four checks you can run today without spending anything, and one thing worth paying for.
Ask the agents what they say about you. Open ChatGPT and type the question your best customer would ask out loud: a fragrance-free moisturiser for sensitive skin, under thirty euros, delivered in the Netherlands. See whether your shop comes up, whether the product it names is the right one, and whether what it says about you is true. Then run the same question in Gemini, Copilot and Perplexity, because they retrieve differently and you can be present in one and missing from the other three. It takes a minute, costs nothing, and it is the only check here that shows you the whole path from question to recommendation. If you have someone who can read your server logs, GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot all identify themselves by name, and whether they reach your product pages or stop at the homepage fills in the rest.
Write for a reader that checks. Take your best product, list the constraints a demanding buyer would state out loud, and see whether the page answers each one explicitly, including the ones where your product is the wrong choice. The limits you admit are what make the rest of your claims quotable.
Audit your promises. Delivery times, stock, return terms. Anything the site implies, an agent will repeat as a commitment, and since June a broken commitment cancels in one click.
Treat your images as product data, because they are being read and not only looked at.
Then do the one thing on this list that costs money. Put something on your shop that answers questions for shoppers and for agents, and that writes down what they asked and where your catalog had no answer.
It doesn't have to be ours and it doesn't have to be the best one available. Whatever you install now gets replaced within a few years. What it recorded doesn't, because the questions your buyers ask are facts about your shop rather than about the software. A shop that starts in 2028 starts with an empty record against a competitor holding two years of one.
We built CAIDEL because we think the shops that do well over the next decade will be the ones represented properly in conversations they never get to see, and that job is larger than most merchants should have to take on alone. If the snapshot showed you something you didn't expect, book a demo and we'll go through your catalog with you and show you what changes when an agent is dealing with a shop that knows how to answer it.
What is agentic commerce?
Agentic commerce is shopping where an AI agent such as ChatGPT, Gemini or Perplexity's Comet researches, compares and shortlists products on a person's behalf, and increasingly completes the purchase. The person sets the intent and the agent does the reading, comparing and negotiating with webshops.
Are AI agents already buying things autonomously?
Not at scale, but that matters less than it sounds. Agents research, compare and shortlist today, and a shopper who receives three recommended options buys one of the three. The decision has already moved into the conversation even where the checkout hasn't.
Does agentic commerce matter for small webshops or only large retailers?
It matters most for small shops. AI models default to big brands when products look identical, but a small shop with a documented, machine-legible advantage for a specific need can break that default. The barrier has moved from ad budget to the work of being precisely described.
Is there a return on this today, or only once agents start buying?
Today. AI tools already shape a large share of purchase decisions, and traffic arriving from them converts better than average: Adobe measured AI-referred visitors converting 42% better with 37% more revenue per visit in March 2026. The same work that makes a catalog legible to an agent also answers the questions human shoppers get stuck on, so the return does not depend on autonomous checkout arriving.
Should I wait until the standards settle before investing?
Waiting costs more than it saves. The tooling will keep changing, but the behavioural record of how agents and buyers interrogate your specific catalog only accumulates once you start, and it can't be bought or backfilled later. Starting early buys a lead in that record while the immediate conversion gains cover the cost of holding it.
How do I know if AI agents visit my webshop?
The fastest check takes a minute. Ask ChatGPT, Gemini, Copilot and Perplexity the question your best customer would ask, and see whether your shop and the right product come up. For the technical view, check your server logs for user agents like GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot and ClaudeBot. Standard analytics won't show you, since a large share of AI-referred visits appear as direct traffic and failed recommendations never appear at all.
Who is liable if an AI shopping agent buys the wrong product from my webshop?
This is unsettled in the EU. The Commission formally withdrew the AI Liability Directive in October 2025, so no specific regime covers it and ordinary consumer law applies, which leans toward the consumer. In practice the shop absorbs the refund and the return costs whether or not the shop caused the error.
What is CAIDEL?
CAIDEL is the reasoning layer that lets WooCommerce merchants sell to AI shopping agents like ChatGPT, not just be found by them. Being found means your products turn up in the agent's list. Being chosen means the agent has a reason to pick them over the other forty, and that is a different job. CAIDEL represents your store to your customers and the AI agents they shop with, and brings the customer data back to you.
AI-referred traffic and conversion. Adobe Analytics, Q1 2026 AI Traffic Report (published 16 April 2026). https://business.adobe.com/blog/ai-traffic-surge-retail-sites-not-machine-readable
Dutch shoppers using generative AI. Q&A Retail, May 2026. https://www.ecommercenews.nl/helft-nederlanders-laat-ai-mee-shoppen/
How shopping agents choose. Allouah, Besbes, Figueroa, Kanoria and Kumar, "What Is Your AI Agent Buying? Evaluation, Biases, Model Dependence, and Emerging Implications for Agentic E-Commerce" (MyCustomAI, Columbia Business School, Yale), arXiv:2508.02630. https://arxiv.org/abs/2508.02630
Brand bias in LLM product recommendation. Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems, arXiv:2606.17443. https://arxiv.org/abs/2606.17443