Indexing merely 1% of the massive product catalogs maintained by tech giants, search newcomer Onton is attempting to disrupt e-commerce search through structured physical reasoning rather than raw inventory scale. According to reporting by MarkTechPost, the company has released Ontology 1, a neurosymbolic search model designed to interpret complex human intent in high-value commerce queries.
On the newly created Subtext-Decor-90 benchmark, Ontology 1 achieved a mean precision@10 score of 0.630 with a 95% bootstrap interval of [0.571, 0.688], calculated from 10,000 resamples. In head-to-head testing across the same 90-query suite, Google Shopping recorded a mean precision@10 of 0.543 [0.490, 0.596], while Amazon registered 0.469 [0.417, 0.521].
Evaluated by three distinct large language model judges—Claude Opus 4.8, Gemini 3.1 Pro, and GPT-5.5—Ontology 1 secured 52 outright query wins, whereas Google Shopping won 19 and Amazon won 16. Outright wins sum to 87 instead of 90 because Onton returned under 10 items for three queries due to its narrow catalog. Factoring in ties alongside overall wins brings Onton’s tally to 58, followed by Google Shopping at 25 and Amazon at 19.
Deconstructing Intent Through Explicit World Models
Rather than depending on seller-provided metadata or surface-level keyword matching, Ontology 1 employs a neurosymbolic reasoning process that breaks down complex user concepts into verifiably checkable properties. When faced with ambiguous queries, the engine actively identifies knowledge gaps, such as defining “pet-friendly,” and derives logical proxies—mapping polyester upholstery as a functional indicator of durability and stain resistance.
This continuous learning system allows accuracy to compound dynamically during operational use, bypassing traditional periodic offline retraining runs. Unlike pure neural network models that store knowledge within uninspectable weight parameters, Ontology 1 constructs an explicit, inspectable world model capable of signaling exactly when critical product information is missing.
The system is backed by Ograph, a specialized graph database infrastructure. Ograph’s GPU implementation runs 43x faster than its CPU counterpart, achieves roughly 100× higher throughput per core compared to a 14-core SuiteSparse:GraphBLAS baseline, and exhibits early tuned GPU build speedups touching 1000× over CPU runs.
Benchmark Nuance and Comparative Performance
Scoring across the Subtext-Decor-90 test set revealed consistent relative engine rankings across all judge models despite score variance. Claude Opus 4.8 awarded Onton 0.759, Google Shopping 0.696, and Amazon 0.641. GPT-5.5 scored Onton at 0.603, Google Shopping at 0.507, and Amazon at 0.422. Gemini 3.1 Pro rated Onton at 0.528, Google Shopping at 0.426, and Amazon at 0.342. The judges demonstrated an inter-annotator agreement measured by a Krippendorff’s alpha of 0.465 across the benchmark.
Ontology 1 excelled on highly subjective, conversational queries in the benchmark set. On “Gross looking art,” Onton reached 1.00 precision@10, compared to Google Shopping’s 0.50 and Amazon’s 0.17. For “laundry hamper I won’t hate looking at for 10 years,” Onton earned 0.97 against Google’s 0.43 and Amazon’s 0.30. On “Rug that hides cat puke but isn’t beige,” Onton scored 0.93 to Amazon’s 0.47 and Google’s 0.10.
Additionally, search performance for “Lighting that makes my apartment feel like a Tokyo cocktail bar at 11pm” yielded a score of 0.93 for Onton, outperforming Amazon at 0.57 and Google at 0.37, alongside a 0.60 mark on “A sofa my husband won’t call feminine and I won’t call a man cave” against Google’s 0.03 and Amazon’s 0.00. On “A chair that’s comfortable for crying in,” Onton achieved 0.73 (Amazon: 0.60, Google: 0.70), and on “Bedroom furniture similar to Call Me By Your Name,” Onton scored 0.53 (Amazon: 0.43, Google: 0.33). On a smaller 10-query image and multimodal comparison dataset evaluated against Google, Onton similarly maintained its focus on physical property reasoning.
When filtering out the three queries where Onton’s small index yielded fewer than 10 products, Onton’s precision@10 rose to 0.665, while Google’s adjusted to 0.549 and Amazon’s shifted to 0.459.
However, a clear regime boundary emerges where seller and category metadata dominate over abstract reasoning. On functional specification searches, Onton faltered against Amazon’s established taxonomy. Amazon achieved a 0.9 precision@10 on “lamp that won’t wake my partner if I read at 3am,” whereas Onton managed only 0.4. On “something to put on a weirdly deep windowsill,” Onton plummeted to 0.07 while Amazon reached 0.67.
| Engine | Key Difference | Best For |
|---|---|---|
| Onton Ontology 1 | Neurosymbolic physical property decomposition via Ograph database | Subtextual, highly subjective, and contextual decor searches |
| Google Shopping | Multimodal vector embedding and broad web-scale indexing | General visual discovery across broad web catalog coverage |
| Amazon | Rigid structured category attributes and direct merchant metadata | Exact dimension filtering and functional product specifications |
📊 Key Numbers
- Onton Subtext-Decor-90 Mean P@10: 0.630 [0.571, 0.688]
- Google Shopping Subtext-Decor-90 Mean P@10: 0.543 [0.490, 0.596]
- Amazon Subtext-Decor-90 Mean P@10: 0.469 [0.417, 0.521]
- Onton Outright Query Wins: 52 of 90 queries
- Google Shopping Outright Query Wins: 19 of 90 queries
- Amazon Outright Query Wins: 16 of 90 queries
- Onton Total Wins (Including Ties): 58 wins
- Google Shopping Total Wins (Including Ties): 25 wins
- Amazon Total Wins (Including Ties): 19 wins
- Onton Adjusted Mean P@10 (Excluding 3 Low-Result Queries): 0.665
- Google Shopping Adjusted Mean P@10: 0.549
- Amazon Adjusted Mean P@10: 0.459
- Inter-Annotator Agreement (Krippendorff’s Alpha): 0.465 across test set
- Claude Opus 4.8 Scores: Onton 0.759, Google Shopping 0.696, Amazon 0.641
- GPT-5.5 Scores: Onton 0.603, Google Shopping 0.507, Amazon 0.422
- Gemini 3.1 Pro Scores: Onton 0.528, Google Shopping 0.426, Amazon 0.342
- “Gross looking art” Query P@10: Onton 1.00, Google Shopping 0.50, Amazon 0.17
- “Laundry hamper” Query P@10: Onton 0.97, Google Shopping 0.43, Amazon 0.30
- “Rug hiding cat puke” Query P@10: Onton 0.93, Amazon 0.47, Google Shopping 0.10
- “Tokyo cocktail bar lighting” Query P@10: Onton 0.93, Amazon 0.57, Google Shopping 0.37
- “Sofa non-feminine non-man-cave” Query P@10: Onton 0.60, Amazon 0.00, Google Shopping 0.03
- “Chair comfortable for crying” Query P@10: Onton 0.73, Google Shopping 0.70, Amazon 0.60
- “Call Me By Your Name furniture” Query P@10: Onton 0.53, Amazon 0.43, Google Shopping 0.33
- Onton Lamp Query P@10 (“won’t wake partner”): 0.4 (vs Amazon 0.9)
- Onton Windowsill Query P@10 (“weirdly deep windowsill”): 0.07 (vs Amazon 0.67)
- Ograph Database GPU vs CPU Speedup: 43x faster GPU implementation
- Ograph Per-Core Throughput vs SuiteSparse:GraphBLAS: roughly 100× on 14 cores
- Ograph Tuned GPU Speedup Peak: early runs touching 1000× over CPU
- Catalog Scale Comparison: Onton indexes roughly 1% of competitor catalog size
🔍 Context
In benchmark evaluations reported by MarkTechPost, researchers evaluated search engine performance across non-traditional queries. E-commerce platforms historically relied on vector embeddings and keyword tags, which regularly fail when users query subtle aesthetic or psychological requirements rather than technical brand names. Ontology 1 attempts to address this gap by replacing uninspectable neural weights with explicit physical property graphs. Within the current search paradigm, this approach challenges the prevailing belief that index size and raw catalog volume are the sole determinants of retrieval quality. Onton competes against established search platforms like Amazon and Google Shopping by prioritizing deeper contextual comprehension over sheer catalog scale. As noted in the reporting, “Adoption today looks like a partnership, not a pip install.”
💡 AIUniverse Analysis
Our reading: Onton demonstrates that reasoning over explicit physical properties yields a genuine structural leap in matching subjective human intent. By mapping concept hierarchies inside Ograph—such as linking pet-friendliness to woven synthetic fibers—the model circumvents the keyword manipulation and SEO spam that routinely degrade vector search on legacy platforms.
However, the model’s limitations are substantial. Ontology 1 struggles significantly when user queries pivot to functional dimensions or strict category constraints, trailing Amazon 0.07 to 0.67 on windowsill specifications and 0.4 to 0.9 on nightstand lighting. Furthermore, relying on LLM-as-a-judge scoring creates noticeable volatility, illustrated by Gemini 3.1 Pro scoring Onton at 0.528 while Claude Opus 4.8 scores it at 0.759. Operating without open weights, a public API, or availability beyond a single home decor vertical, Onton remains an unverified enterprise prototype at scale.
For explicit physical reasoning to challenge broad e-commerce architectures in 12 months, Onton must demonstrate that its four-stage learning loop—spanning gap detection, decomposition, grounding, and generalisation—can expand beyond furniture into multi-vertical retail without requiring custom enterprise integration for every catalog.
⚖️ AIUniverse Verdict
👀 Watch this space. Although Ontology 1 achieves a 0.630 precision@10 on conversational subtext queries, its steep drop to 0.07 on functional specification searches and lack of public APIs leave its enterprise scalability unproven.
🎯 What This Means For You
Founders & Startups: Founders building agentic commerce applications can integrate neurosymbolic grounding layers to evaluate long-tail consumer constraints without building massive catalog scrapers.
Developers: Developers will not have access to open model weights, public APIs, or standard installation packages, requiring direct vendor partnerships to access the underlying Ograph database architecture.
Enterprise & Mid-Market: Mid-market and enterprise retailers can improve relevance on long, requirement-heavy consumer queries that typically fail on conventional vector and keyword retrieval stacks.
General Users: Consumers can search using complex, subjective requirements like “pet-friendly” or “cleanable” without relying on seller-provided tags or rigid attribute filters.
⚡ TL;DR
- What happened: Onton released Ontology 1, a neurosymbolic search model that achieved a 0.630 mean precision@10 on the Subtext-Decor-90 benchmark, outperforming Google Shopping and Amazon while indexing 1% of their catalogs.
- Why it matters: Symbolic physical reasoning enables search engines to decipher subjective, long-tail intent without relying on seller metadata, though it lags on strict functional specifications.
- What to do: Evaluate neurosymbolic architectures for high-intent search applications while monitoring whether Onton opens API access beyond its current private enterprise partnerships.
📖 Key Terms
- Subtext-Decor-90
- A 90-query benchmark designed to test e-commerce search engines on complex, subtextual, and aesthetic decor queries.
- Ograph
- A custom graph database driving Ontology 1 that runs continuous four-stage learning loops on GPU hardware.
- Neurosymbolic search
- A search architecture combining neural language models with explicit, rule-based symbolic reasoning models.
- Precision@10 (P@10)
- A retrieval metric measuring the proportion of relevant results contained within the top 10 search outputs.
- Krippendorff’s alpha
- A statistical measure assessing agreement among different annotators or LLM judges across structured evaluations.
Analysis based on reporting by MarkTechPost. Original article here.