AI-Ready Product Data: Why Agents Can't Read Most Catalogs (and What to Fix)

A buyer used to land on your product page, squint at a half-finished description, and fill in the gaps themselves. They’d notice the unit was probably millimetres, assume the two near-identical SKUs were the same part, and email your team if something didn’t add up. That tolerance is gone. The thing reading your catalog now is increasingly not a person — it’s an agent, and an agent does not squint, assume, or email. It either parses a clean record or it skips you.

This stopped being a forecast in June 2026. Public reporting around Cloudflare Radar traffic data showed automated agents and bots generating about 57.4% of global web requests, with humans down to about 42.6%. Microsoft Advertising also reported that AI-driven sessions nearly tripled in 2025 and that automated traffic is growing roughly eight times faster than human traffic; separate coverage of Microsoft’s agentic-commerce updates noted that about 80% of websites are blocking these agents. The audience for your catalog has quietly changed identity, and most catalogs were built for the reader who left.

The gap shows up in the numbers that matter commercially. Adobe’s Q1 2026 AI traffic report found AI-driven traffic to U.S. retail sites up 393% year over year. Coverage of the same report noted that AI-referred traffic was converting better than other traffic by March 2026, which points to higher-intent visitors who often had the product explained before they arrived. The demand is arriving through a channel most catalogs were not built to be seen in.

Agents fail on ambiguity that humans absorb

The reason is not just model quality. It is the shape of the data underneath. Adyen made this concrete when it announced Adyen Agentic in June 2026, including an Agentic Feed for real-time catalog, pricing, availability, and inventory data across conversational commerce environments. In its agentic-commerce research, Adyen describes the inventory gap plainly: ambiguity is a hard failure for an agent because moving from discovery to transaction requires data precision that many systems were never designed to provide.

A traditional product feed can often get by with a small set of basic attributes: name, price, image, category, and a few merchandising fields. An agent-ready schema needs deeper transactional context: dimensional weight and shipping constraints, real-time regional stock, localized tax and returns rules, compliance detail, and a clear view of which record is the real product. A human customer can wait, refresh, or ask a question when a record is unclear. An agent cannot. Ambiguity that a person quietly resolves becomes a hard stop for software acting on its own.

That single distinction is the whole problem, and it is why this lands hardest on exactly the catalogs Claro is built for: weak-standard, multi-supplier catalogs where the same product arrives three different ways. Picture a hydraulic fitting that supplier A lists as 1/2" BSP, 350 bar, supplier B lists as DN15, 5076 psi, and your own ERP carries under a third part number with the pressure rating missing entirely. A human buyer works out these are the same fitting. An agent sees three products, two unit systems, one blank spec field — and either recommends the wrong one, double-counts your stock, or passes over all three because none is clean enough to act on.

Multiply that across 100,000 SKUs and 40 suppliers and you do not have a content problem. You have an identity problem. That is where entity resolution and a maintained canonical product record become commercial infrastructure, not back-office data hygiene.

AI amplifies what is already in the catalog

The tempting fix is to treat this as a feed exercise — map the fields, fill the gaps, ship a better export. That works for exactly as long as nothing changes, which in a real multi-supplier catalog is about a week. The deeper issue is that automation does not clean a catalog; it acts on it, at scale, without pausing to sanity-check.

A reviewer who mis-keys one record fixes one record. An agent reading a broken canonical record applies that error across every recommendation, every quote, every stock check it touches before anyone notices. Bad data used to slow operations down. Now it executes — automatically, and at volume.

So “make the catalog agent-ready” is not a request for more attributes. It is a request for trust: a single canonical record per real-world product, with the units reconciled, the duplicates resolved, the missing attributes filled from a known source, and a confidence level attached to each value so that downstream systems — search, procurement, pricing, and the agents now reading all three — can act on it safely.

If your catalog already has duplicate clusters, conflicting units, stale supplier rows, or unknown attribute sources, agentic commerce will not hide those problems. It will surface them faster. Start with product deduplication and the supplier intake workflows that introduce most drift in the first place, especially supplier onboarding automation.

What agent-ready actually requires

Generating more product data is easy. Deciding which version of a fact is true enough to trust — and keeping it true as suppliers send new files every week — is the hard part. That is the part that determines whether an agent can read you.

Claro runs this as a continuous loop rather than a one-time export: detect a change, resolve which records describe the same product, validate and enrich the canonical version, write the trusted result back into ERP and PIM with its provenance intact, then keep monitoring for the next change. Identity before enrichment, validation before write-back. The catalog does not just get cleaned once; it gets harder to break every time a new supplier, file, or attribute lands.

Start narrow

You do not fix this with a full-catalog overhaul, and you should not try. Start where the agents and the revenue overlap: your highest-traffic, highest-margin SKUs and the supplier feeds that change most often. Get those to a trusted, canonical state first, prove an agent can now read them, then expand.

The catalogs that win recommendations in the next two years are the ones that started this while it was still cheap to rank, cheap to fix, and ahead of the curve.

See what an agent sees in your catalog — get a free catalog audit

We’ll flag the duplicate clusters, unit conflicts, and missing attributes that make your products unreadable to AI — on your real data, not a demo set.

Sources and article inspiration

This article was shaped by the same macro signals catalog teams are now reacting to:

Related Claro resources

FAQ

What does AI-ready product data actually mean? +

It means a single canonical record per real-world product, with units reconciled, duplicates resolved, missing attributes filled from a known source, and a confidence level on each value — structured so that search, procurement, pricing, and AI agents can act on it without inferring or guessing.

Why do AI agents struggle with catalogs that humans handle fine? +

A human buyer can wait, refresh, or ask a question when a record is ambiguous. An agent acting on its own can’t — so a missing spec, a unit mismatch, or a silent duplicate that a person absorbs becomes a hard failure for software, and gets applied at scale.

Isn't this just a feed or export problem? +

No. Mapping fields into a cleaner feed works until the next supplier file changes something. The durable fix is a continuously maintained, trusted product identity underneath the feed — resolved, validated, and written back as the catalog keeps changing.

Claro keeps multi-supplier product and supplier data clean, matched, enriched, and validated across your existing ERP, PIM, and procurement systems. Your catalog should get smarter every day.