Onton launches Ontology 1 search model
Onton has launched Ontology 1, a neurosymbolic search model that outperformed Google and Amazon on an e-commerce benchmark, signaling a shift toward reasoning-based product discovery.
San Francisco-based Onton has launched Ontology 1, a neurosymbolic model designed to handle complex, conversational, and multimodal product searches. In the newly released Subtext-Decor-90 benchmark, which evaluated 90 text queries using three LLM judges—Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5—Ontology 1 achieved a mean precision@10 of 0.630. This outperformed Google Shopping at 0.543 and Amazon at 0.469, despite Onton indexing only about 1% of its competitors' catalogs. Onton won 52 queries outright, compared to 19 for Google and 16 for Amazon. When excluding empty slots from three queries where Ontology 1 returned fewer than 10 results, its score rose to 0.665, while Google scored 0.549 and Amazon scored 0.459. Although the judges' agreement was modest, with a Krippendorff's alpha of 0.465, all three ranked the search engines in the same order.
Unlike traditional keyword or vector search systems that rely on seller-provided labels, Ontology 1 uses a neurosymbolic approach to construct an inspectable world model. For instance, instead of relying on a "pet-friendly" tag, the model reasons through physical attributes like fiber, weave, and construction to determine durability. This knowledge graph runs on Ograph, a custom graph database. Onton reports that a single Ograph core delivers roughly 100 times the throughput per core of SuiteSparse:GraphBLAS running on 14 cores. Additionally, Ograph's GPU build runs 43 times faster than its CPU counterpart, with optimized runs reaching up to a 1000-fold speedup.
For e-commerce practitioners, this development offers a way to handle long, requirements-heavy queries that typically fail on standard search stacks. However, Ontology 1 is currently limited to the home decor and furniture vertical. It also struggles with specific functional queries where Amazon's massive catalog and metadata dominate, such as "lamp that won't wake my partner if I read at 3am," where Onton scored 0.4 compared to Amazon's 0.9, and "something to put on a weirdly deep windowsill," where Onton scored 0.07 against Amazon's 0.67.
Adopting Ontology 1 is currently a bespoke process. The model is live for end users at Onton.com, but there are no public APIs, pricing tiers, or open weights. Onton is granting partner access on a case-by-case basis, meaning integration currently requires a direct partnership rather than a standard software installation.
This is our own summary of reporting by MarkTechPost
