Onton Releases Ontology 1: A Neurosymbolic Search Model That is 2.7x More Accurate than the World’s Best E-commerce Search Engines
MarkTechPost Michal Sutter
Onton just launched Ontology 1, an AI model that reasons through vague shopping queries instead of matching keywords. It beat Google Shopping and Amazon on accuracy while indexing barely 1% of their catalogs.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Search boxes on shopping sites have not really changed since the late 1990s. Type in something specific like "pet-friendly sectional" and most engines just hunt for a tag that says pet-friendly, which is often missing or simply made up by the seller. Onton, a San Francisco outfit, is betting that this narrow, attribute-matching approach is the actual bottleneck in e-commerce search, and its new model, Ontology 1, tries to fix it by reasoning rather than filtering.
Instead of trusting a seller's label, Ontology 1 builds what Onton calls an inspectable knowledge graph. For a query about pet-friendly furniture, it looks at fiber, weave, and construction, checks those properties against the actual listing, and flags claims that don't add up. It also factors in how trustworthy the listing or review source seems, since gamed listings and paid reviews are apparently common enough to matter. When the model hits a concept it doesn't already understand, it works it out from more basic, checkable properties, then reuses that reasoning the next time a similar query comes in.
Onton put numbers behind the pitch with a new benchmark called Subtext-Decor-90, which it released publicly along with code and data. Three separate multimodal LLM judges, Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5, scored the top ten results from Onton, Amazon, and Google Shopping across 90 text queries. Ontology 1 landed a mean precision@10 of 0.630, ahead of Google Shopping at 0.543 and Amazon at 0.469, and it won 52 of the 90 queries outright versus 19 for Google and 16 for Amazon. It did all this while indexing roughly 1% of what its competitors carry, and only in the home decor and furniture vertical so far.
The numbers aren't bulletproof. Agreement between the three judges was modest, with a Krippendorff's alpha of 0.465, meaning the exact scores wobble depending on which judge you ask. But all three judges ranked the three engines in the same order, which is the part Onton leans on. The model also has clear weak spots: on functional-spec queries like finding a lamp that won't disturb a partner reading at 3am, Amazon's deeper category metadata still wins by a wide margin, something Onton chalks up to its narrower, single-vertical index.
Underneath all this sits Ograph, a custom graph database Onton built to run the knowledge graph, which the company says outperforms a well-known open-source graph library by a wide per-core margin, with a GPU version running many times faster still. None of this is available as a downloadable model or public API right now. Ontology 1 lives at Onton.com for end users, and outside partners get access case by case if they're building agentic commerce tools. For now, this is a product pitch aimed at retailers whose search already struggles with long, specific queries, not a research release for the open-source crowd.
My take — AI-written commentary, not fact-checked reporting
A search model that reasons about fiber and construction instead of trusting a seller's tag is a genuinely interesting idea, and the willingness to publish a benchmark with its own weaknesses laid bare is more honest than most launch posts. But shipping this as a closed, partner-only product with no API or weights means the actual claims are unverifiable by anyone outside Onton's chosen circle, and a Krippendorff's alpha of 0.465 is the kind of detail companies usually bury, not headline. Judge it as a promising architecture with a marketing department that's ahead of its openness, not as proof that keyword search is dead.
Read more about this at: MarkTechPost