AI agents are grading your product catalog now

A retailer can look at a product page and think it is finished.

The photography is clean. The brand voice feels warm. The page has a size picker, a few lifestyle shots, and a description that sounds like something a customer might enjoy reading.

Then a machine reads it and comes back with a different answer: the title does not say what type of product it is, the description is too short to match against a specific shopping request, the material is buried, the variants are thin, and the page does not give enough hard evidence for an AI shopping agent to recommend it with confidence.

That gap is becoming a practical ecommerce problem.

A recent public audit from AI Catalog Score gives a useful snapshot. The site evaluated BurtsBeesBaby.com and scored the catalog 77 out of 100, a B, labeled "Sometimes recommended." The audit covered 250 products on June 6, 2026. It marked 92 products as AI-ready, 158 as needing work, and 0 as invisible.

That score should not be treated as law. One tool's rubric is not the same thing as Google, ChatGPT, Perplexity, Claude, Gemini, Mistral, DeepSeek, or any future shopping agent deciding what to show a customer.

Still, the failure modes are familiar. They are also fixable.

The audit found:

• 137 descriptions with only 44 words, which the audit flagged as too short for AI matching
• 121 products where the product type "Top Set" was missing from the title
• 92 titles without a factual marker such as material, spec, or ingredient
• 63 descriptions using fluff words like "premium," "amazing," or "best"
• 24 titles with only 4 words
• 16 products with only 2 images
• 10 products with only the default variant

The public audit says it covers 61 of 100 signals. The app version claims to cover additional fields such as metafields, SEO meta, image alt text, variant barcodes, and inventory. Again, that does not make the tool an objective referee. It does make the direction clear.

AI shopping systems are going to reward catalogs that explain themselves clearly to machines.

Product pages now have two audiences

Most ecommerce teams already write for people and search engines. They think about conversion, page speed, merchandising, SEO titles, collection pages, internal links, reviews, and paid traffic.

AI shopping adds another reader: the agent acting on behalf of the shopper.

That reader is impatient in a different way. It may not admire the mood of your hero image. It may not infer that "cozy little essential" means organic cotton baby pajamas with a two-way zipper, ribbed cuffs, machine-washable fabric, and sizes from newborn to 24 months.

It needs the page to state the facts.

If a customer asks an AI assistant, "Find a soft organic cotton coming-home outfit for a newborn under $40 that ships this week," the agent has to compare products across stores. The page with a poetic description and weak structured data may lose to the page that says:

• what the product is
• who it is for
• what it is made of
• what sizes and colors exist
• whether it is in stock
• how quickly it ships
• what proof supports the claim
• what images show the product clearly

That is the evidence packet.

The product page still has to persuade a human. But it also has to give a machine enough evidence to classify, compare, cite, and recommend the item.

This did not come out of nowhere

The idea of machine-readable product pages is not new.

Google has published Product structured data documentation for years. Google says product structured data can help product information appear in richer ways across Search, Google Images, and Google Lens, including price, availability, review ratings, shipping information, and more.

Google also separates product markup into use cases. Product snippets fit pages where people cannot directly purchase the product, such as review pages. Merchant listings fit pages where customers can buy from you and can include more detailed product information, including apparel sizing, shipping details, and return policy information.

Schema.org Product gives teams a shared vocabulary for products and services. As of Google's May 2026 web index, Schema.org lists Product usage in the 1M to 10M domain range. Its properties include things like brand, category, color, dimensions, identifiers, aggregate ratings, and additional property values.

So the work is not "AI SEO" in the cheap sense. It is better catalog infrastructure.

The difference now is that AI shopping and research agents may use these facts in more conversational, cross-site, task-driven ways. The page is no longer waiting only for a search crawler or a shopper browsing a collection grid. It may be read by an agent trying to satisfy a specific request.

That connects with the same shift we covered in Visa, agent commerce, and machine-readable product data. Once agents can research, compare, and eventually transact, product data becomes part of the sales system, not a back-office SEO chore.

Fluff is expensive because it hides the match

Retail copy has a bad habit of sounding pleasant and saying very little.

"Premium everyday essential."

"Beautifully crafted for special moments."

"Perfect for your little one."

Those phrases are not always wrong for human tone. The problem is that they do not help a machine answer a shopper's request. They do not say whether the product is a romper, top set, pajama, swaddle, crib sheet, bib, or jacket. They do not say whether it is cotton, bamboo, wool, polyester, fragrance-free, waterproof, dishwasher-safe, vegan, refillable, compatible with a specific model, or sold in a pack of two.

In the BurtsBeesBaby.com audit, AI Catalog Score flagged 63 descriptions for fluff words and 92 titles for lacking a factual marker. Whether you agree with the exact scoring or not, the critique is fair.

A product page cannot rely on vibe when the buyer's agent is matching constraints.

This is also where brand hallucination creeps in. When AI systems do not have clean facts, they may guess. We wrote about that problem in AI brand hallucination and structured-data correction: if the public web leaves gaps, models may fill them badly.

Retailers should not make machines guess what the product is.

A practical catalog readiness checklist

This is the work I would start with before chasing any new AI shopping tool.

This checklist is deliberately boring. That is the point.

A small store does not need to become an AI lab to improve its odds. It needs better product facts, cleaner structure, and a process for keeping both current.

The title has to carry more weight

Product titles are doing more jobs than many stores realize.

A human browsing your site may understand the category from the collection page. An agent may encounter the product page out of context. A title like "Daisy Dream" or "Everyday Set" may work inside the brand's merchandising system, but it is thin evidence outside that system.

A stronger title does not have to be ugly. It just needs enough factual load.

Weak:

Example

2-Pack Leggings

Better:

Example

Baby Organic Cotton Leggings, 2-Pack

Better still, if accurate:

Example

Baby Organic Cotton Ribbed Leggings, 2-Pack

The improved title tells the agent product type, audience, material, texture, and pack count. The brand can still add warmth in the description and photography. The title should do its job first.

This is especially important for Shopify stores where product titles often flow into collection cards, search snippets, Open Graph previews, feeds, and structured data. A weak title echoes everywhere.

Descriptions should answer the questions a good salesperson would ask

A good product description does not have to be long for the sake of length. It has to be useful.

For apparel, that may mean fabric, fit, closure, stretch, care, sizing notes, seasonality, and what is included.

For skincare, that may mean ingredients, skin type, scent, texture, size, usage, allergens, certifications, and what the product does not contain.

For hardware or accessories, that may mean dimensions, compatibility, material, included parts, installation notes, warranty, and safety constraints.

If your in-store salesperson would need the answer, your product page probably needs it too.

This is where ecommerce teams can go wrong with "AI optimization." They imagine strange prompt tricks or hidden text. The better move is usually simpler: say the true thing clearly.

Structured data is the public plumbing

Structured data is not a magic ranking switch. It is plumbing.

Google's product structured data documentation makes this plain. Product pages can become eligible for richer search appearances when they expose the right information. Merchant listings can use detailed product information like apparel sizing, shipping details, and return policy information.

Schema.org Product gives teams a shared vocabulary. Use the specific properties where they fit. Use additionalProperty for true extra characteristics when there is no better property. Do not jam every marketing claim into markup and call it infrastructure.

The same discipline applies to newer proposals like /llms.txt. The proposal describes a Markdown file intended to help LLMs use website information at inference time. It argues that LLMs benefit from concise, expert-level information in accessible locations and names ecommerce sites as one possible use case.

That is adjacent context, not a required standard. Retailers should not treat /llms.txt as a replacement for clean product pages, structured data, feeds, and policy pages.

If the product page itself is vague, a separate LLM file will not fix the catalog.

The work belongs to operations, not just marketing

Catalog readiness crosses teams.

Marketing owns language. Ecommerce owns merchandising. Developers own templates, structured data, and feed logic. Operations owns inventory, shipping, returns, and product truth. Customer support often knows the missing details because they answer the same questions every week.

That is why this work gets stuck. Everyone owns a slice. Nobody owns the evidence packet.

Small retailers can make progress by turning catalog quality into a repeatable workflow:

1. Pick a product category with revenue or margin impact.
2. Export the products and variants.
3. Check titles for product type and a factual marker.
4. Check descriptions for concrete attributes and missing buyer questions.
5. Check structured data against the page content.
6. Check images and alt text.
7. Check availability, shipping, and return policy data.
8. Fix the template so future products inherit better structure.
9. Add a lightweight review step before products go live.

That last point matters. Manual cleanup helps once. A better product intake process helps every week.

This is where process automation can save a team from doing heroic spreadsheet work forever. Catalog maintenance should not depend on someone remembering to fix the same missing fields by hand every month.

What small retailers should do next

Start with your top 25 products.

Do not begin with the whole catalog unless you have a clean process and enough time. Pick the products that drive revenue, paid traffic, organic entrances, or support questions.

For each one, ask:

• Could an AI agent identify the product type from the title alone?
• Does the title include at least one useful factual attribute?
• Does the description answer the questions a serious buyer would ask?
• Are materials, dimensions, compatibility, ingredients, care, or included items stated clearly where relevant?
• Are variants complete and named in a way a machine can compare?
• Do images show enough product evidence?
• Does alt text describe the image rather than repeat the product name?
• Does structured data match the visible page?
• Are price, availability, shipping, and returns current?
• Are reviews, ratings, certifications, and claims represented accurately?

If the answer is no, fix the page. Then fix the template or workflow that created the problem.

The best catalog work compounds. Cleaner titles help site search. Better descriptions reduce support questions. Stronger variant data improves feeds. Accurate structured data helps search systems. Better images help conversion. AI readiness is not separate from ecommerce quality. It is a harsher test of it.

The page has to prove the product

The Hacker News thread around the catalog audit is small, so I would not treat it as a mass-market signal. I would treat it as an early builder signal: people are starting to make product catalogs legible to agents and to measure where that legibility breaks.

That is the change retailers should pay attention to.

If shopping agents become part of product discovery, your catalog will be judged less by how much brand language it contains and more by how clearly it proves each product.

BaristaLabs helps small businesses turn that into actual site and catalog work: cleaner product templates, structured data, Shopify implementation, content systems, and maintenance workflows. Our AI-assisted website development work can clean up the storefront layer, and our process automation work can keep the catalog from drifting after the first cleanup.

If you want a practical read on where your catalog is thin, request an AI catalog readiness review.