For months now, big tech has been promising AI agents that can buy on your customers’ behalf. On the ground, we’re still a long way off. AI assistants compare and recommend, but your customer is still the one placing the order, on your site.
Before the order, though, comes the cart. And on the sites we work with, AI agents are already adding products to it themselves. It’s still rare, but it happens often enough to make one thing clear. Your pages are no longer read only by humans. Machines now compare them, filter them and, increasingly, act on them.
In other words, your customers aren’t handing payment over to agents yet, but they could soon trust them to prepare their order. Where does each step of the agentic journey stand? What does it change for your site? And what should go on your 2027 roadmap? Let’s take stock!
AI assistant or AI agent: what’s the difference?
Before going further, a quick clarification. AI assistant or AI agent: what’s the difference? The two terms are often used interchangeably. Yet they describe two very different uses of your site, and the rest of this article relies on that distinction.
- The AI assistant recommends. ChatGPT, Claude or Perplexity read one of your pages on demand to answer a user’s question. They compare and advise without touching anything on your site. The user then decides whether to visit you or not.
- The AI agent acts. It browses your site in place of the shopper. It filters the catalog, picks a variant, clicks and adds to cart, whether it’s built into an assistant’s app, hosted in the cloud or installed in the shopper’s browser.
The same tool can play both roles. ChatGPT or Claude are assistants when they answer a question, and become agents as soon as they’re given the browser to act.
But as you’ve gathered, the two have nothing in common. One is the in-store advisor. The other is your user’s personal shopper.
Now that the difference is clear, let’s look at their real impact on the customer journey today, and at what’s at stake next.
Where does agentic commerce really stand?
Whether it’s an assistant that recommends or an agent that acts, we talk about agentic commerce as soon as a customer hands part of their buying journey to an AI. The question is which part. KPMG et la Fevad (the French e-commerce federation) identify three levels of autonomy. An AI agent or assistant can:
- assist the customer by searching, summarizing and comparing information;
- prepare an action by building a cart or choosing between several offers;
- execute certain operations under mandate, such as a booking, an order or a restock.
Mapped onto the buying journey, these three levels give four steps, and they’re moving at very different speeds.
To find out what assistants and AI agents actually do, we analyzed one week of logs, at the end of October 2026, across around a hundred French e-commerce sites. Four figures stand out.
- Fewer than 1 in 1,000 pages served by our customers is read by an AI assistant (0.085%). ChatGPT alone accounts for more than four-fifths of them.
- Nearly 6 in 10 of those pages are product pages.
- But 1 in 9 AI assistant requests ends in an error or a block.
- Less than 0.01% of add-to-carts are made by an identified AI agent.
The volume of actions is still anecdotal, but in six days, AI agents added products to carts themselves more than two hundred times. That’s proof an agent can already use an e-commerce interface, and we can expect these uses to grow sharply in 2027 and the years after.
In short, here’s where each step of the journey stands.
| Step | Where are we? | What AI expects from your site |
|---|---|---|
| 1. Discover | Already happening: 31% of online shoppers already use generative AI for their purchases (Fevad / Odoxa), and assistants mostly read your product pages (Fasterize) | Readable pages |
| 2. Select | Already happening: the assistant shows only a short selection of offers (KPMG), and 35% of French consumers would let an AI preselect a product for them (iAdvize × Ifop) | Comparable offers |
| 3. Act | Emerging: less than 0.01% of add-to-carts (Fasterize) | Reliable interactions |
| 4. Pay | Not yet: shoppers lack trust and platforms are pulling back | A framework of trust |
The agentic journey: what each step demands of your site
Should you wait for these numbers to grow before acting? No. Volumes are still modest, but AI-driven selection is already a habit (nearly one online shopper in three relies on it), and order preparation will follow. Your site is already being consulted differently, and it will soon need to welcome agents that act on it. Let’s go through what that changes for your site, step by step.
Discover and select: product pages AI assistants can read
Take a customer looking for a sofa bed to sleep on every night. Before, they’d open several tabs, compare dozens of products and (maybe) end up on your site. Now, they ask an AI assistant, which compares price, stock, delivery times, return policy and reviews, then shows them only three or four offers.
To make that cut, the assistant reads your product pages first. If price, stock or availability aren’t readable there, your offer is less likely to make the selection, and whatever doesn’t make it, your customer never sees.
This risk is far from theoretical. Nearly 1 merchant in 4 has no structured data at all on their product pages, and for nearly 1 in 2, some prices or stock levels only appear after JavaScript runs, which most AI bots don’t execute (Digital Experience Benchmark 2027). For them, those prices simply don’t exist. Your customer still has the final say, but they choose from the assistant’s shortlist.
As we said above, the AI assistant plays the in-store advisor. The traffic it generates is still small (around 2 visits for every 100 from search engines, according to our benchmark), but the visitor it sends you arrives already advised, and converts more easily. In the US, they convert 42% better than other channels (Adobe, March 2026), and we see it with our customers too, where the AI channel often shows the best conversion rate.
Act: the agent moves from reading to doing
This is the newest step, and the one that interested us most. During our six days of observation, AI agents filled carts themselves. Almost all of them came from an AI assistant’s desktop app, whose built-in browser identifies itself as such. The agent opens the page, finds the button and adds the product. Agents that drive the shopper’s usual browser directly remain invisible in our logs.
The volume is marginal, but the signal is clear. Agents can already use an interface designed for humans, as long as it lends itself to it. Filters, variant selectors, add-to-cart buttons… Everything was designed for a human who clicks. It will also have to work for an agent that reads the page structure or finds its way from a screenshot. One fake button, one page that shifts while loading, and the agent misses its target.
Yet according to our Digital Experience Benchmark 2027, covering more than 1,400 of the most visited sites in France, 65% of sites aren’t ready for these agents yet.
Pay: neither trust nor the offering is ready
On the shopper side, the trust isn’t there yet. 55% of Americans and Britons aren’t comfortable with an agent completing a purchase for them (Riskified), and in France, only 21% of respondents say they’re ready (iAdvize × Ifop). On the platform side, the step back has already happened. In 2026, OpenAI dropped Instant Checkout, which let users pay directly in ChatGPT without going through the merchant’s site, and handed payment back to the merchant (Checkout.com).
Agentic payment can therefore take two paths. In the first, the agent browses your site like a shopper, fills the cart and goes through your checkout. It requires no integration, but depends entirely on the quality of your interface. In the second, the transaction is completed inside the assistant itself. The merchant exposes its catalog and checkout through a direct connection, as provided for by OpenAI and Stripe’s Agentic Commerce Protocol, often via its e-commerce platform or payment provider, and nobody loads your pages. Both paths should coexist, and they rest on the same foundation: offers that machines can read. The first adds one requirement: an interface an agent knows how to use.
Staying in control of agents on your site
Whatever the step, the same question comes up. Which agents access your catalog? Which ones act? What can they do, and with what effects? Today, most sites can’t answer.
The first difficulty is identifying them. Only agents that declare themselves show up in your logs, and those installed in the shopper’s browser go unnoticed.
Next, you need to understand what’s blocking them. As we saw, part of AI assistant requests fail. In our measurements, two-thirds of access denials come from the site’s own rules or its anti-bot tool.
Finally, you have to decide which access to allow. Blocking an agent can be a legitimate choice, as long as it’s an informed one. This raises a new question of architecture, security and governance.
Agentic commerce: four workstreams for your 2027 roadmap
There’s no need to overhaul everything at once, but the topic deserves a place on your roadmap now. Waiting for agentic payments to get started would mean arriving after the battle, because every step leading up to it is prepared now. By 2030, Bain estimates that 15 to 25% of US e-commerce will go through agents, starting with simple, recurring purchases.
Here’s how to spread the effort over 2027.
| Order | Workstream | The question to ask your teams |
|---|---|---|
| First | Offer readability | Are our prices, stock, variants, delivery times and reviews readable on our product pages, without relying on JavaScript? |
| Next | Measuring agents | Which agents visit our site, which pages do they read, how many of their requests fail, and can we identify their actions? |
| Then | Reliable interactions | Can an agent filter the catalog, pick a variant and add a product to cart without errors? |
| Monitor | Standards and transactions | Which protocols should we test, and how do we govern permissions, consent and checkout? |
On standards, let’s be direct. It’s too early to choose a final architecture. WebMCP, proposed by Google and being tested in Chrome, is meant to let a site expose to agents the actions it opens to them. The Agentic Commerce Protocol, launched by OpenAI and Stripe, covers product feeds and commercial interactions. Experiment, without locking yourself in.
Is your site ready to welcome AI agents?
Preparing your site for agents starts with knowing what’s really happening on it. That’s Fasterize’s role. The platform combines several data sources about your site (crawl, logs, third-party data, tag) so you can:
- analyze who views your pages (humans, bots or AI agents), what they read and the errors they hit;
- prioritize actions, with recommendations generated from data and AI that your teams approve;
- activate fixes directly on the live site if you wish (JavaScript-loaded content made readable, structured data, Markdown versions for AI bots, more stable pages), or hand them to your IT team;
- measure the impact on your KPIs.
To know where to start, the simplest way is to measure where your site stands today. That’s what the Digital Experience Index is for: a free diagnostic that scores your site out of 100 and compares it with your industry median across five pillars:
- Search, to be found by search engines;
- AI Search, to be technically ready to be read by AI assistants;
- Experience, to offer your visitors smooth navigation;
- Accessibility, to be usable by everyone, as the law requires;
- AI Agents, so an agent can find a price and act on your site.
