AI can't sell what it can't read

Your storefront was built for a person with eyes. Photography, layout, white space, a font you chose carefully. That still matters - people still buy.

But a growing share of the decisions about your products are now made by something that never sees the page.

Google's shopping systems read a feed. Meta reads a catalogue. Marketplaces read a listing schema. And increasingly, a shopper asks an assistant something like "find me a linen duvet cover in dark green, king size, under €150, in stock" - and the assistant answers from structured data, not from your homepage.

All four of these readers want the same thing: fields. Not design.

Three readers you never see

Ad algorithms. A shopping campaign decides which of your products to show against which query. It decides using attributes. A product with colour, size, material, GTIN and a real description can be matched against hundreds of specific searches. A product with a supplier code for a title can be matched against almost nothing.

Marketplace and site search. Filters are attributes. "Show me green, 30cm, in stock" only works if green, 30cm and in stock exist as data. If they're mentioned in a paragraph of prose, they don't exist.

Assistants. This is the newest one and the one people underestimate. An assistant asked for a recommendation has to compare products across shops. It compares what it can parse. A product with clean, complete, consistent attributes is comparable. A product whose specifications live in a PDF, or inside a photograph of a spec sheet, or in a description written as three sentences of atmosphere, is not comparable - so it doesn't get compared.

You don't get an error message for this. You just aren't in the answer.

What "readable" actually means

It's less exotic than it sounds. Four things:

Complete. The attributes that describe the product exist as fields: brand, identifier, colour, size, material, dimensions, price, availability. Not implied by the photo. Not buried in prose.

Consistent. The same idea uses the same word everywhere. M, Medium and M/L across three suppliers are three different values to a machine, and none of them will group correctly.

Single-sourced. Every channel reads from one catalogue. The moment your store, your Google feed and your marketplace listing hold three slightly different versions of the same product, at least two of them are wrong, and you can't tell which.

Accessible. A system you authorise can read your catalogue and act on it - not by scraping your pages, but through a proper connection.

That last one is the shift most stores haven't planned for yet.

The unglamorous work that makes it true

Nothing here is clever. It's the same list every time:

• Fill the identifiers. GTIN or MPN where the manufacturer assigned one.
• Turn spec-in-prose into spec-in-fields. If the description says "100% French linen, 200x220cm", those are two attributes hiding in a sentence.
• Normalise your vocabulary. One list of colours. One list of sizes. One category taxonomy that you own, mapped from whatever suppliers send.
• Fix the images. A machine can't sell from a 400-pixel thumbnail, and Google starts enforcing 500x500 minimums on 31 January 2027.
• Write descriptions that contain facts. Marketing adjectives are invisible to every reader in this article.
• Translate properly if you sell across borders - and know which translations are stale, because a description that changed in English and didn't change in German is a wrong product in that market.

Do this once, properly, and you get paid three times: better ad reach, better site search, and a catalogue an assistant can actually use.

Why most stores can't do it, even knowing all this

Because the catalogue isn't in one place.

A typical shop runs on six systems that don't talk: the store platform, a spreadsheet of supplier data, a translation file, an ads account with its own overrides, a marketplace back office, and somebody's inbox. Product data exists in all six, complete in none.

You can't make a catalogue readable when there's no single catalogue to make readable. That's the actual blocker - not effort, not intent.

So the sequence has to be: pull everything into one place, fill the gaps there, and let every channel read from it. Fixing the Google feed on its own means fixing the same product again for Meta next month.

Where this goes next

The interesting part isn't that AI writes product descriptions. Plenty of tools do that, and a description generated from nothing is just faster nonsense.

The interesting part is AI operating on your catalogue: find every product missing a barcode, rewrite these 800 supplier descriptions in our tone, tell me which products can't go live on Meta and why, translate the new range into German. Work you'd otherwise assign to a person for a month.

That requires a connection, not a chat window. Feedwrk is built as an MCP server - 33 tools, read and write, scoped to a single workspace, with write access opt-in and high-impact actions running as a dry run first. Built today; hosted access next.

If your catalogue is one clean, connected thing, an assistant can work in it. If it's six systems disagreeing, no assistant can help, however good the model.

What to do this quarter

• Pick one channel and get the fields right there. Usually Google. The attribute list is public and strict, and it's a good proxy for everything else.
• Write down your canonical vocabulary - colours, sizes, categories. One page. Everything maps to it.
• Find your worst 200 products by missing fields, and fix those before touching the rest. The tail is long; the damage isn't evenly spread.
• Stop fixing per channel. Every fix should land in one catalogue that all channels read.

None of this is about AI, and all of it is what makes AI usable on your shop later.

Send us your store address and we'll tell you how readable your catalogue is right now - how many products are complete, what's missing, and which gaps cost the most. Free, no account, no install. Shopify stores today.