Why Your AI Shopping Assistant Recommends the Wrong Products

First: do not blame every recommendation problem on the website

An onsite AI shopping assistant can get its information from several different places.

It may use the public product page, a product feed, a commerce API, a search index, a vector database or a combination of these.

Poor recommendations can also be caused by stale indexing, retrieval configuration, ranking logic, prompts, personalization or bugs inside the assistant itself. ClearSiteScore does not inspect those private systems.

What it can inspect is an important part of the foundation: whether the public product information and product pages contain clear, complete and machine-readable signals that an AI-powered commerce system can potentially use.

Imagine your customer asks this

Start with product identity.

A product title such as “Model X Premium” may make perfect sense inside your company, but it tells an external system very little.

Good product data makes the category, product type, brand and important distinguishing attributes explicit.

Product recommendations often depend on attributes rather than product names.

A customer may ask for a sofa suitable for small apartments, a monitor with USB-C charging, a hypoallergenic mattress or a shoe designed for wide feet.

If those characteristics are not explicitly available in the product information, the assistant may have to infer them — or ignore them entirely.

• Dimensions
• Material
• Color
• Size
• Compatibility
• Water resistance
• Use case
• Capacity
• Technical specifications
• Target customer

If an attribute influences the buying decision, make it explicit.

Few recommendation errors destroy trust faster than suggesting a product that is unavailable or outside the customer's budget.

Price and inventory data should be accurate, consistent and accessible in a form that machines can identify.

For public product pages, Product and Offer structured data can make these facts less ambiguous.

• Current price
• Currency
• Availability
• Variant-specific price where relevant
• Consistency between visible page and structured data

Product JSON-LD gives machines an explicit description of important product facts.

Depending on the product and implementation, this can include the product name, brand, SKU, GTIN, images, price, availability, offers and ratings.

Structured data is not a magic recommendation engine. It is simply a cleaner way to express facts that already exist.

{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Trail Pro X7 Waterproof Hiking Shoe",
  "brand": {
    "@type": "Brand",
    "name": "Example Brand"
  },
  "sku": "TPX7-BLK-44",
  "offers": {
    "@type": "Offer",
    "price": "129.00",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock"
  }
}

A product may exist in five colors, eight sizes and three configurations.

The assistant should not recommend a black size-44 shoe simply because the base product is in stock if that particular variant is sold out.

Variant information should therefore be explicit and consistently connected to price, availability and identifiers where your commerce stack supports it.

Many ecommerce descriptions are written to sound attractive rather than to help someone make a decision.

An AI shopping assistant benefits from the same information a knowledgeable salesperson would need.

Explain the problem the product solves, who it is designed for, important limitations, key specifications and meaningful differences from alternatives.

AI shopping is conversational.

People do not always search for exact product names. They ask questions such as “Which one is better for a small apartment?” or “Will this work with my MacBook?”

Pages that explicitly answer common pre-purchase questions provide more useful material for both customers and machines.

• Who is this product best for?
• Who should not buy it?
• What is it compatible with?
• How does it differ from the next model?
• What is included?
• What are the important limitations?
• How long is the warranty?

SKU, GTIN and MPN values are not glamorous, but they help identify and reconcile products across catalogs, feeds and external systems.

For retailers carrying products sold elsewhere, globally recognized identifiers can be particularly valuable.

Missing identifiers can make product matching less reliable.

• SKU
• GTIN
• MPN
• Brand

Specifications tell the assistant what a product is supposed to do.

Customer reviews can reveal how it performs in real situations.

They may expose recurring themes such as fit, comfort, installation difficulty or durability. Only genuine reviews should be displayed or represented in structured data.

Perfect product information has limited value if the system trying to use it cannot retrieve it.

Important public product content should be technically accessible and ideally visible in the server-delivered page rather than existing only after fragile client-side interactions.

Crawler access, HTTP responses, JavaScript rendering and CDN/WAF behavior therefore remain part of the picture.

The recommendation chain

WHERE CLEARSITESCORE FITS

Check the foundation before replacing the AI layer

ClearSiteScore can scan ecommerce pages for a range of issues that affect how clearly products are represented on the public web.

That includes machine-readable product information, structured data, pricing and availability signals, identifiers, product descriptions, images, reviews, canonicalization and technical accessibility.

The goal is not to tell you whether a specific chatbot algorithm is good or bad.

The goal is to show whether your product layer gives machines enough clean information to work with.

• Product structured data
• Offer and price data
• Availability
• SKU / GTIN / MPN
• Brand
• Product descriptions
• Product images
• Canonical URLs
• Reviews and ratings
• Server-visible content
• Crawler accessibility

Before adding another AI tool, fix the product layer underneath it

Frequently asked questions

Continue reading