Kato: Making Health-Product Decisions Easier to Understand

The Problem Kato Addresses

Finding a product is not the same as understanding it.

Online stores can make millions of products searchable. AI assistants can make those catalogs easier to use. But being able to find a product is not the same as understanding it.

In supplements and wellness, a useful recommendation may depend on details such as:

• the shopper's goal, preferences, and non-negotiable constraints
• the ingredients and specific forms used in a product
• the quality and limits of the supporting evidence
• relevant safety considerations
• whether important information is missing or uncertain

A system that overlooks any of these details can produce a confident answer without a sound basis. Kato is designed to slow that jump from product match to product recommendation.

What Kato Does

From an everyday request to the information needed to evaluate products responsibly.

Kato turns an everyday request — such as “I want something for sleep, but I don’t want melatonin” — into the information needed to evaluate products responsibly.

It then works through three separate questions:

1. What is the shopper asking for?

Kato identifies the person's goal, preferences, restrictions, and any context that could materially affect the decision.

2. What is the product?

Kato looks beyond a product title or marketing claim to understand its ingredients, ingredient forms, and overall formulation.

3. What can Kato responsibly conclude?

Kato compares the shopper's needs and the product's contents with reviewed evidence and relevant safety information. It may recommend or compare products, ask a follow-up question, suggest that professional guidance is appropriate, or decline to recommend anything when support is insufficient.

Each outcome — including a decision not to recommend — is intended to be explainable and traceable.

The Principle at the Center of Kato

Understanding a product does not automatically justify recommending it.

Understanding a product does not automatically justify recommending it.

This is the most important distinction in the system.

Kato may be able to identify a product and understand its formulation without having enough reviewed evidence to say that it fits a particular person's needs. Product recognition is therefore treated as permission to consider a product, not permission to recommend it.

That difference appears in Kato's current internal benchmarks. The system reaches product-level understanding for 73.7% of the products in its test catalog, while verified support for a product's fit with a specific shopper need is about 9.1%.

The gap is intentional. Kato is designed to give fewer answers when necessary rather than turn incomplete knowledge into confident advice.

How the Information Is Kept Trustworthy

Three separated layers of information.

Kato separates its information into three layers.

1. Reviewed Knowledge

This layer contains structured information about ingredients, ingredient forms, shopper needs, evidence, product claims, and relevant safety considerations.

New information does not publish automatically. It must pass a human review process that considers the source, the strength of the evidence, the limits of the claim, and appropriate safety context. Kato records what it recognizes, what the evidence supports, what it can reasonably conclude, and what remains uncertain.

2. Market-wide Product Intelligence

This layer applies the reviewed knowledge to products found across external commerce catalogs. Kato currently works with real Shopify catalog data through Shopify Partner APIs, rather than relying on a small demonstration catalog.

As it reviews more products, Kato can see where its understanding is strong and where more product research or evidence review is needed.

3. Private Merchant Intelligence

Kato Commerce applies the broader intelligence inside an individual retailer's or brand's operating environment. It can help identify missing data, conflicting information, catalog-quality problems, claim concerns, operational risks, and market opportunities.

Private merchant information remains separate from Kato's market-wide intelligence. A merchant can use Kato's shared product intelligence, but Kato's shared intelligence does not depend on that merchant's private data.

This separation protects private information and helps prevent one merchant's internal data from shaping conclusions used across the wider market.

Finding the Cause Before Suggesting a Fix

Identify what kind of problem it is before telling an operator what to change.

When Kato detects a problem, it first tries to determine what kind of problem it is.

For example, a weak or uncertain conclusion could result from:

• missing information supplied by the merchant
• a gap in Kato's own product knowledge
• limited or conflicting scientific evidence
• missing external context
• a genuine issue in the merchant's operation

These situations require different responses. Kato is designed to identify the likely source of the problem before telling an operator what to change.

How Kato Measures Progress

Whether its decisions improve — not how many answers it produces.

Kato does not measure success by how many answers it produces or how many products it recommends. It measures whether its decisions improve.

Its internal tests examine whether the system:

• understood the shopper's request correctly
• preserved important restrictions and preferences
• relied on reviewed support
• avoided claims that the evidence did not justify
• handled relevant safety context appropriately
• asked for clarification when needed
• declined to recommend when the available support was not enough

Serious safety failures and unsupported decisions are tracked separately so they cannot be hidden inside a favorable average score. Kato also uses automated checks to protect source traceability, the review process, and the separation of private merchant information from market-wide intelligence.

What Has Been Built So Far

Core intelligence, testing, and review systems operating against real commerce data.

Kato's core intelligence, testing, and review systems are operating against real commerce data. It is not yet publicly available.

Governed Knowledge Base

• 51 standardized ingredients
• 65 standardized ingredient forms
• 63 reviewed statements
• 37 evidence records
• 22 safety records
• 11 active shopper-need categories

The current catalog benchmark covers 548 real Shopify product cards. In that test set, Kato:

• reaches product-level understanding for 73.7% of products
• recognizes 74.1% of the active components observed in multi-ingredient products
• encounters a completely unknown product identity in 1.1% of cases

Internal language tests currently report no hard-constraint violations, no invented shopper constraints, correct safety escalation in all tested cases, and correct decisions to withhold an answer in 96% of tested cases.

These are internal Kato results based on current fixed test sets. They have not been independently validated and will change as the system, product coverage, and reviewed knowledge base grow.

Why This Matters Beyond Supplements

A separate layer of understanding before a transaction should occur.

Today's commerce systems generally answer one of three questions:

• What products are available?
• Which products match these search terms?
• Which product is this person likely to buy?

Kato is built to answer a different question:

Given what this person actually needs, what are these products, how are they meaningfully different, what does the evidence support, and what can we responsibly conclude?

That capability is especially important as AI agents begin to shop and transact on people's behalf. In simple categories, moving from a request to a catalog and then to a purchase may be enough. In health-related categories, an additional layer of product understanding and judgment is needed before a transaction should occur.

Supplements are Kato's starting point because the challenge is easy to see there. The same need exists in any category where a confident but poorly supported recommendation can have meaningful consequences.