Ontology 1: Onton’s Neurosymbolic Search Beats Google Shopping and Amazon

Onton, a San Francisco search startup, released Ontology 1 on August 2, 2026: a neurosymbolic product-search model that scored a mean precision@10 of 0.630 on a 90-query benchmark, against 0.543 for Google Shopping and 0.469 for Amazon. The model does this while indexing roughly 1% of either competitor's catalog, and it is currently live for shoppers at Onton.com. There is no public API, open weights, or pricing tier. Access for teams building on the model is granted through direct partnership, case by case.

What happened

Detail	Value
Company	Onton (San Francisco)
Model name	Ontology 1
Benchmark	Subtext-Decor-90 (90 text queries, home decor vertical)
Onton P@10	0.630 [95% CI: 0.571, 0.688]
Google Shopping P@10	0.543 [0.490, 0.596]
Amazon P@10	0.469 [0.417, 0.521]
Queries Onton won outright	52 of 90
Catalog size vs competitors	~1% of Amazon or Google Shopping
Judge inter-rater reliability	Krippendorff’s alpha 0.465
Availability	Live at Onton.com; partner access case by case

Onton published its benchmark, Subtext-Decor-90, with code and data. Three multimodal LLM judges scored the top 10 visible result cards returned by each engine for each query: Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5. Confidence intervals were computed with 10,000 bootstrap resamples. On three queries, Ontology returned fewer than 10 results; empty slots were scored as non-relevant, which penalises Onton slightly. Excluding those slots shifts its score to 0.665.

Multimodal and image queries were excluded from the 90 because Amazon Lens does not support them and Google Lens does not return products exclusively. Onton runs a separate 10-query image comparison against Google.

Why keyword and vector search struggle here

Standard e-commerce search maps intent onto hard attributes: size, price, material, brand. A filter for “pet-friendly” does not exist in most catalogs, and neither does one for furniture that physically fits a specific room. Onton argues the underlying catalog interface has not changed substantially in nearly 30 years.

Ontology 1 sidesteps this by building an inspectable knowledge graph (stored in Ograph, Onton’s custom graph database) rather than absorbing patterns into model weights. For a query like “pet-friendly sectional,” the model does not trust a seller’s label, which may be absent or false. Instead it reasons from more objective properties: fiber type, weave, and construction. It flags product claims that contradict the underlying data, and it weighs sources, because some listings game the algorithm and some reviews are paid.

When the model has no direct account of “pet-friendly,” it treats that as a gap, works the answer out from cleanability and durability signals, then stores that reasoning for reuse on later queries like “pet-friendly chair” or “cleanable blue couch.” The self-learning loop runs continuously.

On infrastructure: Onton reports one Ograph core outperforming SuiteSparse:GraphBLAS running on 14 cores, roughly 100x throughput per core. A GPU build runs 43x faster than the CPU variant, with early runs reaching 1000x as the implementation is tuned.

Where Ontology 1 loses

Failure cases are concentrated on functional-spec queries where Amazon’s deep category metadata dominates. Two examples from the benchmark results tell the story clearly:

Query	Onton P@10	Amazon P@10
“lamp that won’t wake my partner if I read at 3am”	0.4	0.9
“something to put on a weirdly deep windowsill”	0.07	0.67

Onton attributes these losses to catalog breadth and its single-vertical, non-sponsored index. The company says the self-learning loop should narrow the gap over time.

Why it matters

Precision@10, the share of the top 10 results that are actually relevant, is a direct proxy for conversion. A retailer whose search returns irrelevant products on nuanced queries loses the sale. Onton’s claim is that long, requirements-heavy queries are exactly where keyword and embedding-based retrieval fall apart, and the benchmark gives that claim some statistical backing, even if the judge inter-rater reliability is modest.

The neurosymbolic framing also matters for trust. An inspectable knowledge graph means a product team can audit why a result appeared or was excluded. That is meaningfully different from a black-box embedding model that produces a ranking with no explanation. For regulated categories or high-return-rate products, that auditability has real value. This connects to broader trends in agentic AI workflows, where shopping agents need structured, trustworthy product data to act reliably on a user’s behalf.

The vertical scope is a real constraint today: Onton currently indexes only home decor and furniture. The company says the methodology generalises to other product categories and non-product data with minimal reconfiguration, but that claim has not been independently benchmarked yet.

Our take

The benchmark is self-published, which is worth flagging. Onton designed the query set, chose the judges, and controls the index. A Krippendorff’s alpha of 0.465 is on the low end of acceptable inter-rater agreement, meaning absolute P@10 numbers should be read with some caution. That said, three independent LLM judges consistently rank the engines in the same order, which is harder to dismiss than a single-judge result.

The more interesting thing here is the architecture, not the score. A knowledge graph that decomposes vague predicates into checkable properties, flags contradictions in listing data, and reuses reasoning across queries is a legitimate alternative to pure embedding search. For mid-market retailers running e-commerce platforms where shoppers use natural language with negation and contextual constraints, this is worth watching. If Onton opens broader API access or publishes the methodology for additional verticals, the picture becomes much clearer.

For now, “partnership only” means most businesses cannot test it directly. Keep it on the radar, but don’t plan around it until access opens up. If you’re working on search relevance for your own catalog, talk to us about what approaches are actually deployable today.

Source: Marktechpost