Your Customers Are Asking Questions Your Search Can’t Answer
Imagine this: a customer lands on your online store. They type “sofa that won’t get ruined by my cat’s claws” or “meja kecil untuk ruang sempit” into your search bar. And your store returns… nothing. Or worse, it shows products that are completely irrelevant.
That customer doesn’t blame your search engine. They blame you. They close the tab, and they buy from a competitor who may have worse products but a search box that actually understands them.
This isn’t a niche problem. It’s the gap between how humans describe what they want and how online catalogs are organized. A new AI model called Ontology 1, developed by San Francisco-based company Onton, just exposed how deep that gap is — and what’s possible when a search system stops just matching keywords and starts understanding intent.
TL;DR: Onton’s new search model, Ontology 1, outperforms Amazon and Google Shopping on complex product queries while indexing only 1% of their catalog. The deeper lesson for Malaysian SMEs: search is moving from keyword matching to reasoning. You don’t need this exact tech, but you need to prepare your product data for a generation of search tools that will demand much more than “name, price, category.”
What This Means: Search That Reasons, Not Just Matches
Conventional e-commerce search works on assumptions. When someone types “blue sofa,” it looks for products tagged as “blue” and “sofa.” But what if the customer means “sofa suitable for a small apartment with kids and a pet”? There’s no filter for “suitable for chaotic family life.” So the search fails.
Ontology 1 takes a different approach. It’s a neurosymbolic model — a blend of neural networks (which learn patterns from data) and symbolic systems (which use explicit rules and facts). Instead of just absorbing patterns into weights, it builds an inspectable knowledge graph that reasons through the properties of a product. For “pet-friendly,” it doesn’t trust the seller’s label. It looks at fiber, weave, and durability to decide whether a product claims are actually consistent with being pet-friendly according to Onton.
This approach works. In a 90-query benchmark called Subtext-Decor-90, Ontology 1 scored a mean precision@10 of 0.630, compared to 0.543 for Google Shopping and 0.469 for Amazon as measured by three independent LLM judges. It won 52 of 90 queries outright — while indexing roughly 1% of the catalog size of its competitors Onton reports. That’s a striking result, even with the caveat that judge reliability was modest (Krippendorff’s alpha of 0.465) shown in the article.
| Search Engine | Mean P@10 | Queries Won (out of 90) |
|---|---|---|
| Ontology 1 | 0.630 | 52 |
| Google Shopping | 0.543 | 19 |
| Amazon | 0.469 | 16 |
The biggest challenge in search isn’t the algorithm. It’s the gap between what your customer describes and what your catalog actually knows.
How This Applies to Malaysian SMEs
You may be thinking: “This is San Francisco stuff. My store uses Shopify or EasyStore, and I’m on Shopee/Lazada. I can’t build an AI model.” You’re right. But this benchmark is a wake-up call about how your customers already search — and what that means for your business.
First, think about the language mix. Malaysian shoppers don’t type clean product names. They type “kerusi gaming yang selesa untuk 12 jam” or “kain katil yang tak panas” or even a full sentence in Manglish. Standard search tools struggle with these long-tail, intent-heavy queries. The article notes that both Google Shopping and Amazon lose on “requirements-heavy queries” with attributes like “something to put on a weirdly deep windowsill” and the failure mode scales with catalog size and listing noise. Malaysian marketplaces are full of listing noise — re-listed items, duplicates, and products in categories that don’t match their purpose. Even a basic search “sengal” back pain? The system doesn’t connect the dots because the product data is sparse.
Second, you can start fixing this regardless of your platform. The core lesson from Ontology 1 is that it reasons from objective properties, not seller labels. It looks at fiber, weave, and construction to judge “pet-friendly.” For your SME, that means you should enrich your product listings with concrete, measurable specifications. Instead of writing only “nice sofa,” write “oak frame, polyester-blend upholstery, removable cushion covers, stain-resistant finish.” Instead of “air fryer,” write “1500W, 5L capacity, dishwasher-safe basket, digital timer to 60 minutes.” These are the kinds of attributes that a reasoning-based search can use to answer vague queries like “easy to clean” or “compact for small kitchen.” Your catalog becomes the raw material that smarter search engines will increasingly rely on.
Third, consider the trend toward conversational and multimodal search. Onton’s model handles images and moodboards in addition to text as Onton notes. Malaysian customers increasingly send voice notes, screenshots, and Pinterest boards to your WhatsApp business chat. They don’t use your search bar at all. When that happens, your staff or chatbot needs product data that can handle non-text queries. If you can’t yet support “photo of a chair I like” in your search, at least add more images of your products from multiple angles — clean, with dimensions, and lifestyle shots. Those visuals become metadata for future multimodal search.
Finally, for small catalogs, Onton itself acknowledges the benefit is less because the failure mode scales with catalog size from their company fit description. But that doesn’t mean you’re exempt. As marketplaces and platforms adopt better search backends, your competitors will benefit — and so will you, but only if your product data is structured enough to be understood.
Practical Takeaways: What You Can Do This Week
- Audit your product titles and descriptions. Do they include objective facts (material, size, weight, washing instructions) or only marketing words? Rewrite at least your top 10 selling products with explicit properties.
- Test your search like a customer. Ask two friends to type queries the way they’d speak, not the way your product names are written. Screenshot what comes up. You’ll see the gaps immediately.
- Add “solution-based” fields to your product data if your platform allows. For example: “pet-friendly” as a boolean, “easy to clean” as a feature flag, or “space-saving” as a tag.
- Review your category structure. If your products are misfiled, a reasoning-based search won’t save you. Clean up duplicates and move items to the most accurate category.
- Follow developments in agentic search. Shopping assistants are coming. When they browse, they’ll look for well-structured data. Don’t be the store they avoid because your product feeds are messy.
The Bigger Picture: Search Is Becoming an Understanding Engine
What Ontology 1 is showing us isn’t just “another better search engine.” It’s evidence that the architecture of product search is shifting from pattern-matching to world-model building. The model doesn’t just map text to products — it builds an explicit, inspectable chain of reasoning. The article highlights that it even flags claims that contradict product data and weighs sources because “some listings game the algorithm and some reviews are bought” as reported based on Onton’s approach.
For Malaysian SMEs, the long-term implication is clear: your product data is your most valuable asset. Tools like Ontology 1 currently target large marketplaces and agentic commerce platforms per their stated focus. But the logic will trickle down. The more you invest in structured, honest, property-based product data now, the better you’ll position yourself for the next five years of search and shopping agents.
And even if you never adopt this technology, the benchmark gives you a simple standard to demand from the platforms you use: can your search answer a real customer’s full question, not just their exact phrase? If not, start fixing what you can control — your data.
Search’s job is to translate what customers say into what you offer. If your data can’t support that translation, no algorithm can help you.
Your competitors are already losing sales to bad search. The question is: will you take advantage of the gap before your customers give up on search altogether and just call you on WhatsApp instead?
Ready to Streamline Your Operations?
Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →