Legibility Gets You Read. Proof Gets You Picked.

Short version: An AI shopping agent looks at your product page and keeps very little of it. Text, a live price, a stock status. The hero image, the trust badges and the 4.8-star widget mostly fall through. So there are two separate tests. First, can the agent read the facts on your page correctly? Second, when it can, does anything outside your site back those facts up, so it picks you over the next store? We call the first one legibility and the second one corroboration. Most stores fail the first without knowing, then spend money on the second.

#What an agent actually keeps from a product page

Ask ChatGPT or Perplexity "is this sofa in stock and what does it cost" and watch what it quotes. It quotes text. A product name, a few spec lines from the description, a price, an availability word. Where it got those matters: the HTML your server returns on the first request, and the JSON-LD in that HTML.

What it doesn't quote is the stuff a merchandising team spends weeks on. A lifestyle photo has no price in it. A "Trusted by 10,000 customers" badge is an image. A star rating rendered by a review app after page load often isn't in the first HTML at all. None of that is useless for humans. For an agent it's close to invisible.

So when people ask "is my store AI-ready", the first honest answer is: open the page source and read it like a machine would. If the price is only in a banner image, or the stock status only shows up after JavaScript runs, the agent is working from a different page than your shopper is.

#Tier 1: legibility

Legibility means the agent reads your facts the same way a shopper would. The fix almost never lives in the theme or the design. It lives in a few boring places:

• A plain-text description with real specs: dimensions, material, model number. Three lines of mood copy gives an agent nothing to quote.
• Price and stock as text in the HTML.
• JSON-LD that says what the visible page says, once. If a theme and an SEO app both emit a Product graph and they disagree, the agent has to guess which one is true.

Home Elegance USA is the clean example. It's a 30,855-product Shopify store selling 30 furniture brands. Four AI agents (ChatGPT, Gemini, Perplexity, Claude) told shoppers that in-stock furniture was backordered. The cause was a template sentence hard-coded into every product page, while the real stock status got painted in by JavaScript later. Shoppers saw "In stock". Agents read the HTML and saw "backordered". We fixed the template and the data around it in 10 days, re-ran the same 20-cell test, and the count of agents getting it wrong went from 4 to 0. Three now say "In stock" outright. The full before/after is on the Home Elegance USA case page.

No redesign. The store looks the same today as it did before the fix.

#Tier 2: corroboration

Here's the part that surprises people. After the fix, Home Elegance was legible. Every agent described its products accurately. And when we asked the broader category questions, the "best affordable luxury furniture" kind, its visibility stayed roughly flat: 4 of 12 before, 3 to 4 of 12 after. We say that on the case page on purpose.

Reading you correctly and recommending you are two different steps. Once an agent can read ten stores that all sell a similar sofa, it needs a reason to name yours. It doesn't take your word for it. It looks for sources outside your site that agree with you: a category roundup or distributor list that names you, the same specs and model numbers showing up on the manufacturer page and the marketplace listing, numbers it can check somewhere other than your own homepage. Your site claiming to be "the trusted choice" counts for roughly nothing. Three other places saying the same spec does.

Oroca shows what it looks like when a store wins that second step. It's a B2B dental-materials distributor in Vietnam that launched a brand-new site, built to be machine-readable from day one, starting at zero presence in AI answers. We track it weekly on 30 real buyer prompts across ChatGPT and Gemini. By week 37 of 2026, Oroca held 45.0% share of voice in its category, first of the six brands tracked, with an AI mention rate of 82% across those prompts. Every mention is logged against a prompt, an engine and a week, so the number can be opened and checked row by row. Method and table are on the Oroca case page.

One thing that number doesn't tell you: which outside sources each engine leaned on when it named Oroca. The dashboard records who got named and cited. It doesn't record the agent's reasoning, and we'd rather say that than pretend it does.

#Why most stores lose at tier 1 and pay for tier 2

Tier 1 failures are silent. Nothing breaks in analytics when an agent reads "backordered" on an in-stock product. Google keeps ranking the page, because Google renders JavaScript and waits. The store owner never sees the agent's version of the page, so they assume it matches theirs.

Tier 2 is where the budget goes, because it's the part with a visible symptom: "ChatGPT recommends our competitor, not us." That sounds like a reputation problem, so the spend goes to PR, listicle placements, review campaigns and content. Some of that is worth doing. But if the agent can't read your price or stock correctly, every outside mention points it to a page that tells it something wrong. You paid to send traffic to a page the agent misreads.

The order matters. Legibility is cheap and quick to verify, usually days of template and data work. Corroboration is slow and runs over months. Doing the slow one first is how a store spends a quarter on mentions and still gets quoted as out of stock.

#How to check both tiers yourself

You don't need a call or a tool subscription for a first read.

Tier 1, about 15 minutes:

1. Open three product pages you care about. Add view-source: in front of the URL. Search the raw HTML for the price and the stock word. If either is missing, or the wrong one is there, that's your first fix.
2. In the same source, find the application/ld+json blocks and count how many say "@type": "Product". More than one on a page usually means two apps are fighting.
3. Ask ChatGPT and Perplexity directly: "Is [exact product name] from [your store] in stock, and what's the price?" Write their answers next to what your admin says.

Tier 2, about 30 minutes:

1. Write down five questions a buyer asks before they know your name, like "best supplier of X in Y" or "where to buy X for a small clinic". Ask each one in two AI assistants.
2. Note who gets named.
3. Search for those brands outside their own sites. Which lists, distributors, manufacturer pages and marketplaces mention them, and do the specs match from one place to the next?
4. Run the same search for your store. The gap between their list and yours is your corroboration backlog.

If tier 1 fails, fix it before you spend anything on tier 2.

Want a second pair of eyes on both tiers? Send one store URL and we'll run a free async AI-correctness read: we ask live AI agents what they see on your product pages, check it against what your store actually says, and send back a named punch-list for each tier: luma-e.com/ai-readiness. No call required.