For twenty years, e-commerce optimisation meant optimising for two audiences: human visitors and Google's crawler. A third audience has arrived. AI assistants and answer engines now research products, compare options and recommend sellers on behalf of their users, and agent-driven checkouts are moving from demo to reality. When that agent picks three coffee grinders to show its user, the shops it could not read were never in the running.
What is agentic commerce?
Agentic commerce is the shift from people browsing shops to software agents acting on a person's behalf: researching, shortlisting, comparing and increasingly transacting. The spectrum runs from an answer engine citing two shops in response to “best waterproof cycling jacket under 200 euro” to an assistant that completes the purchase once its user approves. The market is early and moving; nobody can honestly tell you what percentage of purchases will be agent-mediated next year. What is already observable: assistants shape shortlists today, and a shortlist you are not on is a sale you never see.
Why can't AI agents read most webshops?
Because most modern shops render in the browser. Product data arrives as JavaScript, and the visible page is assembled client-side. Human browsers execute that JavaScript; many AI crawlers and agents do not. They fetch the raw HTML, and if the raw HTML is an empty application shell, that is what they learn about your shop: nothing.
This is the single most common reason shops score poorly on agent readability. It is not a content problem; the content exists. It is a delivery problem: the content never reaches the reader that matters.
What do AI agents actually need from your shop?
Agents want what careful humans want, in machine form: facts, structured and stable.
- Server-side rendered content. The information must be in the HTML response itself, not assembled by scripts afterwards.
- Structured data. Schema.org JSON-LD for products, offers, FAQs and the organisation, so an agent does not have to guess prices and availability out of prose.
- Answers, not just attributes. The buying questions that decide purchases, in a structured, citable form. This is exactly what BIQs are for.
- Stable, crawlable URLs. An agent that cannot address a page cannot cite it or send a buyer to it.
What are llms.txt and agents.md?
Two young conventions try to give AI systems a front door. llms.txt is a plain-text file at your domain root that points language models to your most important, machine-friendly content, a sitemap for models rather than crawl bots. agents.md goes a step further and describes how automated agents should interact with your site. Neither is a magic ranking switch, and both are still settling as standards. But they are cheap to publish, they signal that your machine-readable content is intentional rather than accidental, and the shops experimenting with them now are the ones learning fastest.
How do you become agent-readable without replatforming?
The uncomfortable answer would be “rebuild your frontend with server-side rendering”. Fortunately there are lighter paths:
Add Schema.org markup to your existing pages. Works well if your platform already renders server-side; does not help if the whole page is client-rendered.
A server-rendered, structured mirror of your catalog and answers on its own subdomain, built for machine readers while your main shop stays untouched.
Pull your curated content as ready-made structured data via API and serve it wherever you need. Integration is around ten lines of code, and the entity type is controllable per entry.
Klariton supports the second and third path directly: a crawlable agent subdomain fed by your published answers, or the Schema Export API, which returns your content as structured entities you can drop into any page. The integration is around ten lines of code, the Schema.org entity type can be set per entry, and the details live in the Schema Export docs.
How do you measure whether agents actually see you?
Optimising for an invisible audience without measurement is guesswork, so instrument two things. First, bot read tracking: which AI crawlers and agents fetch your content, what they read, and how often. That turns “are agents visiting?” from a hunch into a log. Second, purchase attribution: when a buyer arrives from an assistant's recommendation or an agent completes a checkout, record it. Together they connect the machine audience to revenue, which is the only argument that survives a budget meeting.
The strategic point is simple: agent readability compounds. Structured, source-backed, server-rendered answers serve human visitors, search engines and AI agents from a single effort. The shops that treat the machine audience as a first-class reader now will be the ones agents quote by default later.