Making Magento 2 Agent-Ready: The Correctness Layer No One Is Measuring

Short version: AI shopping agents now read your Magento store and answer questions about it for shoppers who never visit your site. Google's new Lighthouse "Agentic Browsing" score tells you the door is open — whether you publish an llms.txt and register WebMCP tools. It tells you nothing about whether the agent, once inside, says the right thing. That second question — is the agent correct about you — is the layer no one is measuring. This is that layer, and MARS is the score for it.

#The shift, and the half of it Magento merchants are missing

In Spring 2026 Shopify turned on Agentic Storefronts by default: every eligible store's catalog is syndicated to ChatGPT, Copilot, Google AI Mode and Perplexity automatically, no apps, no feeds. AI-driven traffic to Shopify stores grew roughly 8x year over year, orders from AI search over 10x. The plumbing became free.

Magento and Adobe Commerce got none of that for free. There is no default toggle. To participate you implement it yourself — a product feed, structured data, an MCP or ACP endpoint, an llms.txt, WebMCP tools. A healthy ecosystem of modules now does exactly this, and Google's Lighthouse 13.3 (May 2026) even added an Agentic Browsing category that grades whether those signals are present.

So the Magento world is racing to get present. That race is real, and it is the easy half. The hard half — the one that decides whether the agent wins you a sale or quietly costs you one — is whether what the agent reads is true.

#Three failure modes, and only one of them shows up in a checklist

When an AI agent encounters your store, one of three things is happening:

1. It can't read you. A rendering or crawlability problem. Your content is painted in by JavaScript the crawler never runs, or your endpoints aren't reachable. More content won't help; this is engineering.
2. It reads you but describes you wrong. A correctness problem. The served HTML, the schema, or the policy text tells the agent something false, and the agent repeats it with total confidence.
3. It reads you correctly but doesn't pick you. An authority problem. Slow, real, and the only one that isn't a quick fix — reviews, entity authority, third-party corroboration over weeks.

A presence checklist — llms.txt present, WebMCP registered, feed submitted — catches none of failure mode 2. That is the expensive one, because it is invisible. It doesn't show up in your analytics. It shows up as sales that never happen.

#The served-HTML truth gap

Here is the mechanism, stated plainly, because it is the single most common way a Magento store lies to an agent without anyone noticing.

Your browser runs JavaScript. A crawler, and the agent reading through it, mostly does not — it reads the served HTML, the raw document your server returns before any script executes. If your product template hardcodes a sentence like "This item is backordered" into that served HTML and then corrects it with JavaScript after load, then humans see "In stock" and every AI agent sees "backordered." The page isn't broken. The page is lying, correctly, to the only reader that matters here.

We watched exactly this on a live 30,855-product furniture store. Baseline test, five buyer questions across four agents, twenty cells: four of four agents told shoppers an in-stock item was backordered. ChatGPT diagnosed the store's own site as broken and recommended competitors. Gemini decided the retailer was the wholesale brand it carries and said it "cannot sell" one of its own brands — wrong entity, wrong catalog, confident tone. None of it was an inventory problem. It was two lines of template logic. Ten days of correctness fixes later, all four agents stopped; three now explicitly confirm "In stock." (Full teardown.)

That store, incidentally, could have passed a presence check the whole time.

#Why "the door is open" is not the same as "the agent tells the truth"

Google's Lighthouse Agentic Browsing score is a genuinely useful addition — it standardizes the presence signals, and you should pass it. But understand what it measures. It measures whether you published the things agents look for. It does not send four agents to ask about a real product and check whether the answers are right.

Lighthouse tells you the door is open. It doesn't tell you whether the agent walks in and says the right thing.

This gap is about to get wider, not narrower, because the modules that auto-publish llms.txt and register WebMCP tools make it easy to pass presence while your catalog data, your schema, and your policy text quietly contradict each other underneath. The confidently-wrong tax gets paid by whoever optimized for the checklist and stopped there.

#MARS: scoring the layer that's missing

MARS — the Magento Agent-Readiness Score — is an open rubric that scores a store 0–100 across six dimensions, weighted so that correctness carries the most.

• Readability (20, gate): can an agent reach and read the served HTML at all. If a store fails this, nothing else matters, so a failing readability score caps the total.
• Correctness (25): does the served data tell the truth — stock, price, entity identity, no internal contradictions. The heaviest weight, because this is the invisible-loss dimension.
• Structured Data (20): valid, deduplicated JSON-LD; brand attached to product; no competing schema graphs.
• Answerability (15): policy and FAQ coverage machine-readable, stated clearly enough that an agent answers from your page instead of inventing a number.
• Transactability (15): a reachable, protocol-agnostic agent endpoint (UCP/ACP/MCP), on a supported Magento version, with a catalog an agent can actually query.
• Observability (5): can you separate AI-channel traffic and see which queries surfaced — and failed to convert.

Two things make MARS different from a connectivity checker. First, it is protocol-agnostic — it doesn't grade you on whether you installed a particular module. Second, part of it can only be scored by asking real agents real buyer questions and checking the answers against the truth — the twenty-cell test. That human layer is the point. It is the only way to catch an agent that is confidently wrong.

#What to do this week

You don't need to rebuild anything. You need to audit, in this order.

Check readability first: view-source a product page and confirm your title, price and stock status are actually in the served HTML, not painted in afterward. Check robots.txt and your CDN aren't blocking GPTBot, PerplexityBot, ClaudeBot or Google-Extended, and that /.well-known/* is reachable. Then check correctness: is the stock sentence in the served HTML conditional, or is a "backordered" string hardcoded for every product? Ask ChatGPT, Gemini, Perplexity and Claude one real buyer question about one real in-stock product, and read what they say back. Then look at your schema for duplicate or brand-less Product nodes, and your policies for numbers an agent could get wrong.

If you pass presence but fail correctness, you are in the most dangerous position available in 2026: visible to every agent, and wrong. That is fixable, usually in days, and it is worth more than any amount of new content.

Want your store scored on MARS — including the twenty-cell agent test? We run it as a free async audit: luma-e.com/audit.