A confident mess is worse than a gap.
Catalogue AI is easy to demo and hard to trust. Every mapping carries a confidence score against a threshold you set — above it applies automatically, below it goes to a merchandiser rather than being forced into the nearest category.
Across suppliers, formats and languages.
A call with the people who built it, not a sales pitch.
This is for you if
If none of these are true we are probably not the right call yet, and we will say so.
- Supplier feeds arrive in formats no two of which agree
- Taxonomy drift means yesterday's mapping is wrong today
- Product data spans languages your team does not all read
- Search returns the catalogue rather than the product
What we build for retailers and marketplaces
Supplier titles resolved to your taxonomy at volume, not sampled and hand-corrected.
Every mapping carries a score, so the threshold — not the vendor — decides what auto-applies.
Search that understands the product, including attributes buried in unstructured descriptions.
Range, placement and enrichment actions taken within rules you define.
Forecasts at SKU and category level, including the seasonal behaviour that breaks averages.
Relevance tuned per shopper without hand-built rules per segment.
The long tail is the whole problem.
Select an area for how it is handled at catalogue scale rather than on a clean sample.
Accuracy on the common case is table stakes.
Every vendor maps a clean product title. What decides a catalogue is the tail — ambiguous, abbreviated and multi-language titles that get mapped confidently and wrongly at a scale nobody reviews.
- Confidence scored per mapping
- Threshold you set decides what auto-applies
- Ambiguous items routed, not force-fitted
- Abbreviations and supplier shorthand handled
- Multi-language titles read natively
- Long-tail accuracy measured separately
Yesterday's correct mapping is today's wrong one.
Taxonomies change with season, range and merchandising strategy. Mappings are versioned so a taxonomy change is a reviewable event rather than a silent rewrite of catalogue history.
- Mappings versioned against taxonomy revisions
- Re-mapping proposed, not applied silently
- Impact of a taxonomy change previewed
- History preserved for reporting continuity
- Bulk review tooling for merchandisers
- Rollback of a bad taxonomy change
Supplier data does not arrive in one language.
Product data is read in the language it arrives in rather than translated first, because a translation step is where attribute detail — sizes, materials, pack quantities — quietly disappears.
- Native reading across supplier languages
- Attributes preserved rather than translated away
- Mixed-language fields handled
- Units and pack quantities normalised
- Locale-specific naming retained
- No per-language pipeline to maintain
Search that understands the product.
Attributes buried in unstructured descriptions are what shoppers actually search on. Retrieval over the full product record beats keyword matching against a title field.
- Retrieval over descriptions, not just titles
- Attribute extraction feeding facets
- Synonyms learned from real queries
- Zero-result queries surfaced as a signal
- Relevance measurable, not anecdotal
- Personalisation without hand-built rules
Accuracy on the common case is table stakes.
Every vendor maps a clean product title. What decides a catalogue is the tail — ambiguous, abbreviated and multi-language titles that get mapped confidently and wrongly at a scale nobody reviews.
- Confidence scored per mapping
- Threshold you set decides what auto-applies
- Ambiguous items routed, not force-fitted
- Abbreviations and supplier shorthand handled
- Multi-language titles read natively
- Long-tail accuracy measured separately
Where catalogue AI usually goes wrong.
The failure is rarely accuracy on the common case — it is the long tail being mapped confidently and wrongly, at a scale nobody can review.
You decide what auto-applies. Starting conservative and relaxing it as accuracy is measured is the whole method.
Supplier data in several languages handled natively rather than routed through a translation step that loses attributes.
Taxonomies change. Mappings are versioned so a change is reviewable rather than a silent rewrite of history.
Let's talk about the feeds that break things.
The suppliers whose titles nobody can parse and whose categories never line up. Describe that on a call and we can tell you quickly whether this is worth your time — no data, no feed, no NDA needed to have the conversation.
E-commerce AI — common questions
Accuracy on the common case is table stakes; what matters is the long tail. Every mapping carries a confidence score, and you set the threshold above which it applies without review. That makes the accuracy question answerable per item rather than as a single headline number.
They go to a merchandiser with the candidates and the reason for the uncertainty, rather than being forced into the nearest plausible category. Forcing them is what produces a catalogue that looks complete and is quietly wrong.
Yes, natively. Product data is read in the language it arrives in rather than being pushed through a translation step first, which is where attribute detail is usually lost.