For over two decades, eCommerce development has been aggressively focused on user experience - optimising visual layouts, DOM structures, and client-side rendering for human interaction. However, the rapid emergence of LLM-driven autonomous web agents introduces a new variable: agent experience (AX).

We hypothesised that current eCommerce architecture, which relies heavily on probabilistic UI interactions and dynamic DOM states, is fundamentally hostile to autonomous AI agents. To test this, we engineered the AX Auditor - a Python-based testing harness leveraging Google’s Gemini models and thebrowser-use framework.

We tasked this autonomous agent with navigating a standard Shopify storefront, locating a parameterised product (e.g. a specific price range, size and colour), and executing an add-to-cart action. Our objective was to measure how easily an AI could parse and interact with the storefront purely through machine-to-machine methods, rather than relying on visual or DOM-based navigation.

Making an agent shop like a customer

Initially, we deployed the AX Auditor against a baseline Shopify environment. We observed the agent’s behaviour as it attempted to evaluate the DOM and parse the rendered viewport to execute its pipeline.

To resolve the latency and unreliability of visual navigation, we developed a new Shopify app, codenamed Agenticify. Rather than relying on the agent to interpret the UI, Agenticify injects a structured, machine-readable application/ld+json (Schema.org) data layer directly into the storefront. It uses server-side AI to parse store policies, catalogue data, and pages, building this semantic payload with minimal configuration.

We then conducted A/B testing, running the AX Auditor before and after installing Agenticify, to measure changes in execution speed, accuracy, and reliance on visual fallback.

When the interface becomes the problem

Baseline testing revealed clear limitations in standard front-end architectures. When evaluating DOM content, the AI proved surprisingly brittle. Forcing the LLM to interact visually with the browser - mapping screenshot coordinates to interactive elements - was computationally expensive and unreliable. Transient UI states, such as cookie consent modals or dynamic search overlays, frequently caused execution failures.

The testing phase also acted as a QA loop. The AX Auditor flagged that our initial Agenticify build was missing ListItem schema on search results pages - a gap we subsequently patched.

Here is the telemetry from our AX Auditor runs:

Without Agenticify installed

Without a structured data payload, the agent was forced into computationally heavy DOM traversal and visual interpretation.

AX Score: 50
Schema Usage: Medium
Friction Points:
- Lack of SearchAction schema on the homepage and search modal necessitated visual search.
- Absence of Product or Offer schema on the search results page required visual identification and selection of the product.
- Lack of BuyAction or direct add-to-cart schema required visual interaction to add the item to the cart.

With Agenticify installed

With the structured data layer active, the agent bypassed the visual rendering engine entirely. It successfully parsed the semantic JSON-LD payloads, extracted the exact machine-to-machine endpoints, and executed the transaction programmatically.

AX Score: 100
Schema Usage: High
Friction Points:
None. The site provides a complete machine-to-machine path via Schema.org, including SearchAction for discovery and variant-specific BuyAction URLs for direct cart integration.

Schema isn’t just for search

Is optimising for AI agents overhyped? From an engineering perspective, the answer is definitively no.

The key takeaway from this RD exercise is a needed shift in how developers view Schema.org structured data. Historically, retailers have treated JSON-LD payloads purely as a marketing and SEO requirement for Google's rich snippets.

That era has ended. Schema.org can also act as a machine interface for AI agents.

Relying on computer vision or complex DOM traversal for agentic workflows is fragile. If a headless architecture or storefront relies exclusively on visual UX and lacks a comprehensive, machine-readable semantic layer (AX), autonomous agents will fail to execute confident state transitions, leading to automated cart abandonment. As LLM shopping assistants mature, providing that data layer via tools such as Agenticify moves from nice-to-have optimisation to a baseline requirement.