Shopper intent, classified in 37ms.
A custom commerce model — not a GPT wrapper. Classifies shopper intent in 37 milliseconds and routes every question to the right skill. The taxonomy grows from your catalog, real shopper conversations, and AEO/GEO analytics — so it expands with your brand. 11 points more accurate than GPT-4o.
A model trained for shoppers, not chatbots.
Frontier LLMs are general. Shoppers aren't. We trained a 66M-parameter classifier on the actual question shapes that show up in commerce — and the taxonomy keeps learning from your catalog, conversations, and AEO/GEO analytics so it never goes stale.
Harvested from your data, not bolted on
Three signals train an intent taxonomy unique to your brand — your catalog structure (categories, attributes, policies), real shopper conversations across chat and FAQ, and AEO/GEO analytics from agent traffic. Add a product line, a new policy, a new buyer persona — the taxonomy expands. It is never capped at a fixed thirteen.
Routing, not generation
Once intent is classified, the right skill runs — catalog search for discovery, structured-attribute lookup for fit, the FAQ index for policy, a recommendation engine for gifts. The LLM doesn't get to hallucinate the answer.
37ms classification
Single forward pass on a fine-tuned DistilBERT. Compatible with streaming flows, edge deployment, and conversational latency budgets that frontier LLMs can't meet at any cost.
Predictable per-request pricing
Bundled into your plan and billed per classified request. No token-based surprise bills when agent traffic spikes — and orders of magnitude cheaper than a frontier LLM call for the same classification.
Benchmark methodology, training data and per-class results: why we built our own intent classifier.
Common questions.
01What is shopper intent classification?
It's detecting what a shopper actually wants — discovery, fit check, inventory, comparison, gifting and more — before answering, so the question can be routed to the right deterministic skill instead of a generic LLM response.
02Why not just use GPT-4o for this?
On our 12,400-query ecommerce benchmark the 66M-parameter classifier scores 81.0% versus 70.4% for GPT-4o, responds in 37ms instead of hundreds of milliseconds, and costs nothing per query. Small and specialized beats large and general at this narrow job.
03Does the taxonomy adapt to my store?
Yes. The intent taxonomy is harvested from your catalog structure, real shopper conversations, and AEO/GEO analytics from agent traffic — it expands as your product lines, policies and buyer personas do.
04Where can I see the benchmark methodology?
We published the full write-up — training data, per-class results including the weak spots, and comparisons against seven frontier LLMs — on our blog: 'Why we built our own intent classifier'.
Get a free product data quality audit.
Find out which product attributes LLMs are missing on your catalog — and what it's costing you in agentic search visibility. No call required, the report lands in your inbox.