Google Shopping scores 0.543 mean precision@10 on complex product queries. Amazon scores 0.469. A startup just beat both by 2.7x while indexing a fraction of either catalog.

Keyword search has been failing shoppers for years and retailers just absorbed the loss

Standard e-commerce search breaks on queries that carry real intent: negations, room constraints, mood-driven descriptions, multi-requirement filters. Retailers have no filter for “pet-friendly” or “fits a narrow hallway,” and every failed search is a cart abandonment with no error log.

A neurosymbolic model now parses the queries that keyword and vector retrieval drop

Onton Releases Ontology 1 takes conversational, multimodal product queries as input and returns ranked results scored for precision, trust, and contextual fit. Users type or speak complex requirements at Onton.com, and the model resolves negations, listing noise, and vague aesthetic intent into specific product matches. On a 90-query benchmark judged by three independent LLMs, it reached a mean precision@10 of 0.630 against 0.543 for Google Shopping and 0.469 for Amazon.

The teams with the most to gain are already losing revenue to bad search

• E-commerce search product managers whose relevance metrics plateau because long-tail queries never resolve cleanly against attribute-based filters.
• Agentic commerce engineers who need a grounding layer that handles negation and multimodal input without custom prompt engineering for every query type.
• Marketplace operators in high-SKU, high-noise verticals where listing quality variance tanks precision and no reranker currently fixes it at scale.

The model currently indexes home decor and furniture only, which is where listing noise and complex spatial requirements make failures most costly and most visible.

Agentic shopping is arriving faster than most search stacks can handle it

As shopping agents move from prototype to production, the retrieval layer underneath them becomes the critical failure point, and neither BM25 nor dense vector search was built for multi-constraint conversational queries. Ontology 1 is the first released model to benchmark directly against Google Shopping and Amazon on this exact failure mode, which means the comparison is now on the table for every retailer evaluating their stack.

What you can actually do with it today

• Run complex conversational product searches with negation and spatial constraints at Onton.com.
• Evaluate precision against your current site search on requirements-heavy queries.
• Contact Onton for partner access if you are building a shopping agent that needs a grounding layer.
• Use Ontology 1 for moodboard-driven discovery where image and text inputs need to resolve together.

Pricing is partnership-based and granted case by case, with no public API or pricing tier listed.

One real constraint before you get excited

Access is not self-serve, the model weights are not public, and the only indexed vertical right now is home decor and furniture, so any production use outside that category is still theoretical.

The alternatives are not standing still

Coveo and Bloomreach both offer AI-powered site search with enterprise support and live integrations across multiple verticals. Neither has published a direct benchmark comparison against Google Shopping at the query-complexity level Ontology 1 is targeting.

Neurosymbolic retrieval is about to reframe what enterprise search vendors have to prove

When a model indexed on 1% of a catalog outperforms Google Shopping on precision, the benchmark conversation for every search vendor just shifted. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.