Opus AnalyticsResearch note · Montreal

What AI agents read on a retail page

A controlled test of Meta AI, Perplexity, ChatGPT and Claude against an instrumented storefront, with every request logged.

Summary

  • Meta AI stops reading a store when the customer wants to buy from it. The same page, same model, same session settings: asked as research it was fetched and quoted correctly; asked as a purchase it was never fetched at all, and the answer named eight competing products with retailers and prices.
  • What an agent cannot read, it invents. Perplexity described a contact form it had never loaded — wrong number of fields, wrong labels, wrong button — and cited the site as its source.
  • Two missing HTML elements made an entire form invisible. Adding two label elements, and changing nothing else, took one agent from reporting no fields at all to reporting both fields correctly.
  • No agent executed JavaScript. Across every run, by all four agents, not one script ran.
  • Four agents, four networks, and two of the four are not identifiable by user agent. Bot rules written against user-agent strings block the honest crawlers and admit the rest.

What was tested

A single-product storefront on instrumented infrastructure, logging every request with timestamp, path, user agent, source network and geography. The site carries a visible notice that it is a research site; no order can be placed on it.

The product carries a different price in five places, one per reading method. The price an agent reports identifies how it read the page, without inference.

PriceWhere it appearsWhat reporting it establishes
CAD 149.00JSON-LD Product block onlyStructured data was parsed
CAD 159.00Server-rendered visible text onlyThe raw HTML was read
CAD 169.00Written by script after loadThe inline script source was read
CAD 179.00Public JSON API onlyThe API was found and preferred
CAD 189.00Revealed only after a clickThe click handler was read

One correction matters here, because the obvious reading of the third row is wrong. Reporting the scripted price does not establish that a page was rendered: the value also appears in the served HTML inside the script itself, so it can be read without anything being executed. Rendering is established only by a beacon endpoint that fires when a script actually runs. Every statement below about JavaScript rests on that beacon.

All bot-protection and performance features that could interfere — bot fight mode, browser integrity check, AI crawler controls, managed robots.txt, minification, script optimisation — were verified disabled before each run, so what was measured is the agent rather than the CDN.

Finding 1 — Meta AI stops reading a store at the moment of purchase intent

Two prompts, three minutes apart, same URL, same model, same mode, separate conversations. The only variable was how the request was framed.

Framed as research, Meta AI fetched the page and reported the correct price. One request in the log, from Meta's network, timestamped.

Framed as a purchase, Meta AI made no request at all. It stated that it could not check the link directly, then returned a shoppable carousel of eight competing lamps with retailer names and prices, from Best Buy, IKEA, Vari, Structube, EQ3 and Desky.

Its own reasoning trace records the switch:

I'm reviewing the product page to find the price and availability information.

I can't check live price or stock from that external URL, and I'm searching the catalog for Aurora desk lamp options available near Ottawa instead.

It states that it is reading the page, never requests it, and falls back to Meta's own shoppable catalogue. Because the product is absent from that catalogue, the customer is shown eight other retailers.

Two details are worth noting. No location was supplied in the prompt, so the geographic filter in that last step is Meta's own inference about the user — a variable no merchant can see or query. And this behaviour is Meta's alone: the identical purchase-framed prompt put to ChatGPT and to Claude produced a fetch and a correct answer from both, with no substitution.

Finding 2 — what an agent cannot read, it invents

Perplexity was asked to list the fields on the contact page. It reported three fields — Name, Email, Message — before a "Send Message" button, and cited the site as its source.

The form has two fields, labelled "Your email address" and "Your message", and the button reads "Send".

The logs show why. Perplexity's crawler requested exactly two things: robots.txt, and the site root. Given a deep product URL it discarded the path and fetched the homepage, then described a page it had never retrieved.

Meta produced the same class of error in a different register: asked what the checkout collects, having never fetched it, it reported "the standard contact, shipping address, and payment fields". The checkout asks for an email address and nothing else.

The pattern is consistent across both. An agent that cannot reach a page does not report silence. It supplies the most plausible version of a store of that kind, and the customer receives it as fact.

Finding 3 — two HTML elements determined whether a form existed

The contact form was plain server-rendered HTML: a real form, method post, an email input, a textarea, a submit button, no JavaScript, fully crawlable.

Meta AI fetched it and reported that no form fields were visible in the content it retrieved. The form carried no label elements, and the textarea had no placeholder or aria-label. Form controls with no associated text were discarded by the extractor while headings and prose were kept.

One change was then made: two labels, each associated with its control. Nothing else on the site was altered — not the method, the field names, the button text, or any styling.

The same question, in a new conversation, returned the correct answer: the form asks for exactly two fields, "Your email address" and "Your message".

Complete blindness to a precise answer, from two elements and roughly ten minutes of work.

Finding 4 — no agent executed JavaScript

The product page loads a beacon script that calls an endpoint when it runs. An ordinary browser produces that request every time, which establishes the control.

Across every run by all four agents, the only beacon requests in the log came from a conventional browser. None came from an agent.

ChatGPT illustrates why the distinction matters. It reported the scripted price on both prompts, which resembles rendering. Its fetch was a command-line HTTP client, no beacon request was made, and the script file was never requested. It read the value out of the script source and reported it as the displayed price. On this page that produced the right answer. On a store where a script applies a promotion, recalculates tax, or falls back to a default, it would not.

Anything a page paints with JavaScript is not visible to these agents, and anything merely referenced in a script may be reported as though it were on screen.

Finding 5 — four agents, four networks, two unidentifiable

AgentSource networkIdentification
Meta AIMeta Platforms Ireland (AS32934)meta-webindexer/1.1
PerplexityAmazon (AS14618)PerplexityBot/1.0, with documentation link
ChatGPTCloudflare (AS13335); robots check via Microsoft (AS8075)curl/8.5.0; a generic browser string for its cloud browser
ClaudeGoogle Cloud (AS396982)An identified Claude user agent

Two consequences follow. A WAF allowlist covering one cloud covers one agent; these are four unrelated networks and the list will change. And bot policy written against user-agent strings is unsound — Perplexity and Claude identify themselves, Meta uses a string few merchants would recognise, and ChatGPT's content fetches arrive as curl while its browser arrives as a generic desktop client. Such a policy blocks the identifiable agents and admits the rest.

Separately, ChatGPT's browser path required explicit user consent before the site could be opened. Only the lighter fetch ran without it, which bounds what page-side work can achieve for that agent.

What this means for the stores you manage

Ordered by what decides the outcome, not by what is cheapest to fix.

1. Catalogue presence, because it decides purchase-intent answers

When a question carries purchase intent, Meta answers from its own shoppable catalogue rather than from the merchant site. A technically flawless store that is absent from that catalogue is replaced in the answer by retailers that are present: in this test, eight of them, by name and price. Shopify now exposes a Global Catalog searchable across every merchant from a single endpoint, so for a Shopify store the question is no longer whether an agent can read the page. It is whether the product is represented correctly and chosen over the one beside it. What reliably drives that selection is not established here, or anywhere we can find, and it is stated as unresolved rather than guessed at.

2. The platform configuration the platform does not do for you

Shopify ships UCP and UCP-compliant MCP servers, and product discovery is automatic. Two things still require the merchant: hosting a UCP profile at a well-known URL for capability negotiation and signed-request verification, and subscribing to order lifecycle webhooks for fulfilment, returns, refunds and edits. Neither is a default.

3. Readability, which governs every question that is not a purchase

Informational questions still resolve against the live page, and that is where the findings above apply. Label every form control, since an unlabelled control is discarded by the extractor. Serve price, stock and shipping in the initial HTML, because values painted by script are not seen and values merely referenced in script may be read and reported incorrectly. Return real status codes on error pages, since a 200 response with an error rendered inside it is a common route to an agent stating something false with full confidence.

4. Bot policy by network, not by user agent

Four agents, four clouds, and two of them unidentifiable from their user-agent string alone. A policy written against user-agent strings blocks the identifiable agents and admits the rest.

One approach to avoid entirely: embedding instructions to agents in page content. Meta own reasoning trace states that it reproduces page content without following embedded instructions, so it does not work, and it is search spam under a different name with the attendant risk to the domain. It is currently being sold.

Limits of this test

One trial per condition, one product, one day. The two framings differ in wording as well as intent, so "intent" here means the whole framing rather than an isolated term. Prompt order was not counterbalanced. The label test altered two elements rather than one, since an association requires an identifier. One agent's contact-form prompt did not complete, for reasons unrelated to the agent, so that result is absent rather than negative.

These findings are a dated snapshot of systems that change weekly, and Meta Connect falls on 23 and 24 September 2026. The value of any figure here is the date attached to it. They will be re-verified.

Conduct: the instrument runs only on infrastructure under our control, never against a client's site or a third party's. No bot protection or CAPTCHA was circumvented. No test transaction was placed on any live store. No personal data was collected. Where an agent was blocked, the block was recorded rather than worked around.

About this work

Opus Analytics is an independent practice in Montreal working on agent readiness for ecommerce: establishing what AI agents currently read, get wrong, or cannot reach on a merchant's storefront, and remediating it.

The engagement begins with a fixed-price audit at CAD 3,000, delivered in five business days: what four agents currently report about a store against what is true, scored by severity, with a prioritised remediation plan and the in-house fixes marked as such. Remediation and monitoring follow where they are warranted.

White-label and referral arrangements are available for agencies and development shops managing merchant portfolios.