2026 is the year AI shopping agents exit the demo phase. OpenAI Operator, Anthropic Claude Computer Use, ChatGPT Shopping, Perplexity Buy with Pro, Gemini Shopping — each, in its own way, lets a user execute a purchase on their behalf: find a product, compare, add to basket, even validate checkout. For an e-commerce merchant, this is a paradigm shift: a growing share of traffic becomes agents, not humans, and a poorly prepared catalogue becomes invisible to this new audience.
This article answers three concrete questions: what already constitutes a measurable share of traffic in 2026? How does an AI agent actually “read” a product page? And what technical actions to take on a PrestaShop catalogue to stay discoverable?
What is actually deployed in 2026
The ecosystem at the time of writing consists, roughly, of:
- ChatGPT Shopping (since 2024-2025): a visitor poses a purchase-intent question (“best foldable adult electric scooter under £700”), ChatGPT answers with a clickable short list. The click leads to the source store. Inclusion is conditioned on the quality of structured data and IndexNow/sitemap.
- Perplexity Pro Shopping: integrated into the Perplexity answer engine, surfaces clickable products in the margin of “purchase” answers. Highly visual, sourced directly from Schema.org and Open Graph markup.
- Google AI Overviews Shopping: integrated directly into Google SERPs. Selection based on the Shopping graph (which depends on Merchant Center and Product/Offer markup).
- Operator (OpenAI) and Claude Computer Use (Anthropic): agents that control a browser. Still rare in 2026 but growing. They actually navigate the site, click buttons, fill in forms. A synthetic visitor following the DOM and the visual rendering.
- Agents embedded in other apps: Microsoft Copilot Shopping, Amazon Rufus (on Amazon), nascent WhatsApp Business agents.
Estimated volume on UK fashion / décor / tech B2C mid-market stores in 2026: 2 to 8% of referrer traffic already comes from AI agents, growing fast. Not yet majority, but no longer marginal.
How an AI agent “reads” a product page
Two reading modes coexist and will continue to coexist in 2026:
Mode 1: data-first (the majority)
The agent does not load the rendered page. It queries a public API or scrapes structured Schema.org markup. Priority sources:
- JSON-LD Schema.org embedded in the page (Product, Offer, AggregateRating, hasMerchantReturnPolicy, shippingDetails).
- Open Graph meta tags (og:title, og:description, og:image, og:type=product).
- Google Merchant Center feed (XML or Content API).
- llms.txt file if present (which may point to a structured version of the catalogue).
- Product XML sitemap with modification dates and priorities.
No JavaScript executed, no rendering displayed. If the information is not in the static HTML or in the structured feeds, it does not exist for the agent.
Mode 2: browser-first (Operator, Computer Use)
The agent loads the page as a human would — executes JS, sees the rendered DOM, takes screenshots, identifies buttons, clicks. Its reading is closer to a real visitor, with a few specifics:
- It needs interactive elements that are named and labelled (buttons with aria-label, forms with associated labels). The accessibility work done for WCAG 2.2 directly serves here.
- It is sensitive to blocking popups and modals without a clear close button. A cookie banner impossible to close with the keyboard will block the agent as much as a user with a disability.
- It favours linear paths and single-screen checkouts. A six-step funnel is riskier than a one-page checkout.
The optimisation checklist for a PrestaShop catalogue
1. Complete and up-to-date JSON-LD Schema.org
The foundation. See the dedicated article on Schema.org Product 2026: hasMerchantReturnPolicy, shippingDetails, ProductGroup for variants, real AggregateRating, GTIN/MPN/brand filled. Without this, an AI agent cannot match the product page to a purchase intent.
2. llms.txt file at the root
The llms.txt standard (championed by Anthropic since late 2024) proposes a markdown file at the domain root that describes the site and points to the structured resources most useful to language models. For an e-commerce store, the typical llms.txt contains:
- Merchant identity and categories covered.
- Product sitemap URL.
- JSON catalogue feed URL (if available).
- Return and shipping policy summary.
- Contact information for agents (machine-friendly email).
At DataFirefly, the dfllmstxt module generates this file automatically for PrestaShop stores, with daily updates and per-shop personalisation in multi-shop setups.
3. Clean Product Open Graph
Beyond Schema.org, AI agents and social networks query Open Graph and Twitter Card tags. og:type=product, og:image at 1200×630 format, og:title clear, og:price:amount and og:availability when relevant. Many PrestaShop themes forget og:price.
4. Stable canonical URLs
An agent returning to a product page must find the same URL. URLs with dynamic parameters (id_lang=1, controller=product, etc.) are less reliable than stable rewritten URLs. SEO and AEO converge here.
5. Reliable server-side availability and pricing
The agent reads the HTML rendered server-side. If the real price is computed in client-side JavaScript (frequent with dynamic promotion modules), the agent will see the default price. Check that the price displayed in the static HTML matches the real price.
6. Robots.txt and explicit crawl policy
Major AI agent crawlers now identify themselves with recognisable user-agents: GPTBot, ClaudeBot, PerplexityBot, ChatGPT-User, Anthropic-Verifier, Google-Extended. In a 2026 robots.txt, you explicitly allow those you want to welcome and block those you refuse. Blindly blocking all AI bots is tempting but cuts the store off from the future audience.
7. Server rendering performance and stability
Browser-first agents give up if the page takes more than a few seconds to load or if the DOM constantly shifts (high CLS). The Core Web Vitals work done for Google directly serves agents — same technical criterion, simply applied by an agent instead of an indexing robot.
8. Machine-friendly checkout
For agents that go all the way to purchase (Operator, Computer Use): one-page or very short checkout, no blocking captcha, express payment available (Apple Pay, Google Pay), clear summary before confirmation. A basket abandoned because of a forced captcha at checkout for an AI agent is a missed sale — and in 2026 that starts to count.
Measuring the agent share of your traffic
AI agents generally leave traces in the user-agent and sometimes in the referrer. The main ones to monitor in 2026 in server logs or GA4:
- GPTBot, ChatGPT-User (OpenAI).
- ClaudeBot, Claude-User, Anthropic-Verifier (Anthropic).
- PerplexityBot, Perplexity-User (Perplexity).
- Google-Extended (Gemini / AI Overviews).
- CCBot (Common Crawl, feeds several models).
- Applebot-Extended (Apple Intelligence).
Adding a custom dimension in GA4 or a filter in a server-side analytics tool to isolate this traffic enables measuring growth and computing the real ROI of AEO investment.
Risks to anticipate
Three traps to keep in mind:
- Agent-driven fraud: a misaligned or hijacked agent may attempt mass purchases, scrape prices, or worse. Protection through intelligent rate-limiting (block suspicious patterns without blocking legitimate agents) and step-up authentication for atypical baskets.
- Brand-customer disintermediation: if the agent loops the funnel end-to-end without the user landing on the merchant site, the brand loses the communication opportunity (newsletter, retargeting, cross-sell). A trend that will amplify and must be integrated into the loyalty strategy.
- AI-generated content in reviews: a pattern already observed. Investing in verified reviews ensures the authenticity of the signal and compliance with Google’s quality requirements on reviews.
Conclusion: preparing your catalogue is not optional
AEO in 2025 was an interesting option. AEO in 2026 is a condition of access to traffic. Stores that have not updated their Schema.org markup, published their llms.txt, and configured their robots.txt for AI agents will see their market share of AI traffic trend towards zero while their competitors capture the first network effects.
Good news: these technical actions are among the cheapest in 2026 digital marketing. A few weeks of well-targeted work is enough to move into the “agent-readable catalogue” category. The real competition now plays out on the quality of the product, the price, and the service — not on the technical friction of accessing information.
