A prospect opens ChatGPT and types: “find me a booking tool for my restaurant, compare prices, suggest the best one”. The assistant searches, reads a dozen websites, returns an answer with two or three names. The shortlist is settled before the first human click.

That is agentic commerce: the AI takes a mission and carries it out for its user. Your website is the ground where that mission succeeds or fails.

1. Agentic commerce, in plain terms.

The word is new; the shift comes in two stages.

First stage, already here: delegated research. Your prospects ask an AI the questions they used to type into Google, and the assistant compares on their behalf before returning a short list. Being on it or not is the ground of GEO: getting recommended by ChatGPT, Perplexity, Google’s AI answers. We detailed how those answers get made.

Second stage, just starting: delegated action. ChatGPT completes purchases with partner merchants in the United States. Perplexity launched Comet, a browser that follows links, fills in forms and runs a booking end to end. ChatGPT’s agent mode does the same. The scope is still narrow, but the direction is set: after the recommendation, the transaction.

The shift has a measure. According to an October 2025 IDC prediction, 64% of customer interactions will run through agents and answer engines by 2029, rather than through brands’ own websites. The number seems to sideline the website; it does the opposite: to recommend you or act on your pages, an agent needs a website to read.

For a SaaS, a startup or a solopreneur, the consequence fits in one sentence: your buyer delegates the sorting, and soon the first step. Your website therefore has to pass two tests: get picked by the AI, then let it act.

2. Three surfaces, only one belongs to you.

Your clients meet AI in three places, and those three surfaces are not equal.

The point chatbot and messaging vendors leave out: all three surfaces feed from the same place. The assistant that recommends you has read your website. The agent booking for its user goes through it. The chatbot that answers correctly draws its answers from it. Investing in channels before the base means plumbing three taps into an empty tank.

3. Getting picked: what the AI must read on your pages.

To put you on a short list, an assistant needs comparable facts.

• A named offer and a published price. “Custom website from €3,500 excl. VAT, training included” fits into a comparison. “Solutions tailored to your challenges” does not. A number beats an adjective, for a reader and for a machine.
• Short answers to buyer questions. Timelines, scope, what is included: a FAQ visible on the page gives the AI sentences it can quote as they are. It is one of the rules of a mention that counts.
• Readable groundwork. Schema.org structured data, llms.txt file, sitemap, speed: what helps Google understand your website helps agents understand it. On our projects, that groundwork is part of every site’s scope.

A page that saves its information for the sales call protects nothing: it drops out of the comparison.

4. Being actionable: what an agent must be able to do.

The next test is action. An agent that picked your name must be able to complete its mission: book a slot, request a demo, create an account.

What makes it fail is well known, because it is what already drives your human visitors away:

1. The key action buried three clicks from the homepage.
2. A button whose label does not say what it does.
3. A form that demands eight fields to ask a question.
4. The decisive information locked in a PDF or an image.

The path that converts a human in a hurry is the one an agent can follow. It is what Cyril Ben Said, founder of Studio Lumen, describes after his website’s redesign:

Mathieu delivered a real showcase: a custom page for my equipment, a gallery of my past projects, Calendly booking built in. Clients arrive, see my work, pick their slot. Result: more appointments, more signed clients.

Arrive, understand, pick a slot. An agent follows that path without help. A website that does not convert humans will not convert their AI any better.

5. Where to start, in this order.

1. The readable base. Offer, prices, dated proof, visible FAQs, schema.org, llms.txt. It is the same work as SEO and GEO: done once, it serves all three surfaces.
2. The action path. A short, standard route to the booking, the demo or the sign-up. Test it: ask an agentic browser to book with you, note where it gets stuck.
3. Conversational channels, after that. Chatbot, WhatsApp: useful to extend a client relationship, provided the base exists. Never as the first investment.

This work is designed in from the start, not patched on after launch. If a launch, a fundraise or a trade show is coming up, now is the right time to lay it down. Book a call, thirty minutes to scope your project.