Why AI Can't Read Your Catalog (And Recommends a Competitor Instead)

Your product pages are fine. They convert. Customers buy from them every day, they've been through more than one redesign, and someone spent real money on the photography.

None of that helps you here, and understanding why is the whole game.

Adobe's tooling puts the average retail product page at 66 out of 100 on machine readability, with the blunt conclusion that for retailers with thousands of SKUs, a significant share of the catalog is currently invisible to AI. That number is easy to misread as "your pages are 34% bad." It isn't a quality score. It's a measure of how much of what your page communicates survives the trip to a machine.

The trip your page makes

When a shopper asks an assistant to recommend something, roughly this happens:

The assistant needs to answer a specific question: waterproof, under $150, size 10, ships this week. To do that it needs facts it can commit to. It reads whatever it can reach about your product: the page, structured data if you have it, whatever it retained from crawling you earlier, third-party sources.

Then it has to decide whether to name you.

This is the part worth sitting with. A recommendation engine that hedges is useless, so these systems are heavily biased toward stating things confidently, which means they're correspondingly reluctant to assert a product claim they can't ground. If it can't establish that your boot is waterproof, it doesn't say "possibly waterproof." It moves on to a product where it can establish it.

Your page didn't fail a quality bar. It failed to make a claim in a form that could be repeated.

Five ways this goes wrong

These are the patterns that come up again and again.

1. The attribute is shown, not stated

The single most common one. Waterproofing is conveyed by a photo of the boot in a stream. Capacity is conveyed by a picture of the bag with a laptop in it. Fit is conveyed by three reviews saying it runs small.

A human reads all of that instantly. A machine reads a paragraph about weekend adventures and a filename. When the query filters on the attribute, you're not a candidate.

2. Variants that can't be resolved

You sell one product in four colours and eight sizes. A human sees a swatch grid and understands. A machine sees a page whose price changes depending on selections it can't make, whose availability is behind a dropdown it can't reliably operate, and whose title matches all thirty-two combinations equally.

Asked for "the black one in a 10," it can't confirm that specific thing exists at that specific price. Ambiguity is indistinguishable from absence.

3. Price and availability that require execution

If price only appears after JavaScript runs, or stock status is fetched by a call the reader never makes, then from the outside your product has no price and no availability. Both are query filters. Both being unknown removes you from consideration for any question containing a budget, which is most of them.

4. Specs that live somewhere else

Dimensions in a downloadable PDF. Materials in a tab that loads on click. Compatibility in a support article. Every one of those is a place a human will happily go and a machine probably won't.

5. Descriptions written to persuade rather than describe

This one stings, because it's your best copywriting. "Engineered for those who refuse to compromise" is good marketing and carries zero extractable facts. When the assistant compares your page against a competitor's boring spec table, the boring table wins, not because it's better writing, but because it answers the question.

Why this is a revenue problem and not a technical one

Any of the above, on any one product, costs you nothing measurable. It's the aggregate that hurts, and it hurts in a way that's specifically hard to notice: you lose the queries you would have won.

Nobody bounces. Nothing 404s. The shopper asked for exactly what you sell, got a confident recommendation for someone else, and bought it. Your logs are silent. Your conversion rate doesn't move: the visit never happened.

And it lands hardest on the traffic that matters most, since shoppers arriving from AI convert around 30% better than shoppers from traditional search. You're not losing a marginal segment. You're losing your best-converting one, silently, before it arrives.

Do not buy the score

A whole category of free tools has appeared that will grade your site's AI-readability. Some are decent. Cloudflare has one, Adobe has one, and there are half a dozen weekend projects with similar names.

They all stop at the same place: a number.

A number tells you nothing actionable. It doesn't say which of your 4,000 products are affected, which of those anyone is actually asking about, or which one to fix first. A score of 66 and a score of 71 imply the same Monday morning: nothing.

What you want instead is a list, ordered by money:

• Which specific products fail, and on which attribute
• Whether anyone is asking about those products, since a broken page nobody queries is not a priority
• Which competitor is being recommended in your place
• What it's worth: that query volume, against your AOV and conversion rate
• Whether the fix worked: the same question, asked again, with a different answer

That last one is the part almost nothing offers, and it's the only proof that any of this mattered.

How to check your own catalog today

Pick your best-selling category. Then:

1. Write twenty questions a real shopper would ask before buying from it. Full sentences, with budgets and constraints, like "waterproof hiking boots under $150 for wide feet," not "hiking boots."
2. Ask each one of the assistants your customers actually use.
3. For each answer, record: were you mentioned? Was anything about you wrong? Who was recommended instead?
4. For every miss, open your own product page and find the fact the assistant would have needed. If you can't find it stated plainly, you've found your cause.

Twenty questions takes about an hour and the pattern usually shows up in the first five.

Fixing it is mechanical: state the attribute where a machine will find it, make variants and price unambiguous, move the specs out of the PDF. Fixing it across four thousand products, and knowing which four hundred to start with, is the part that needs tooling. That's what we do: find the failures, price them, repair the pages, and re-run the same questions to prove the answer changed.

But run the twenty questions first. You'll know within the hour whether any of this applies to you.

---

Machine readability and revenue-per-visit figures: Adobe, reported via Marketing Week, March 2026. Conversion comparison: Adobe Analytics, holiday 2025.