The Quick Take
AI product matching is for the catalog you actually have, not the one you wish you had. If every Shopify variant and every Odoo product shared a barcode, you would not need it. Most stores that grew on both systems do not, and the gap between "BLK-M" and "TSHIRT-BLACK-M" is where connectors either create duplicates or stall.
Synco's matching runs as a background job, links the exact matches by code, asks an AI model to judge a short list of candidates for the rest, and puts the result on the product mapping page for you to review. This guide covers what it does, how the shortlist is built, what the confidence score means, and when to skip it.
Related reading: how AI helps sync Shopify and Odoo, AI-assisted sync page, product matching by SKU, barcode, or name, mapping existing products, Shopify Odoo product sync page.
Why Name Matching Alone Fails
Connectors that match by name do one of two things wrong. Either they require an exact title match, which almost never happens across two systems edited by different teams, or they fuzzy-match and silently link "Blue Mug 300ml" to "Blue Mug 350ml". The first leaves most of the catalog unlinked. The second creates sale orders against the wrong product, which you discover at stock count or, worse, at invoice.
Identifier matching is right when it is available. SKU and barcode are unique in both systems, so an exact match there is safe. The problem is the long tail: products created before anyone agreed on a SKU scheme, variants whose barcode lives only in Odoo, bundles, and anything imported from a CSV twice.
The AI step exists for that tail. It is not a replacement for identifiers. It is what happens after identifiers have been tried.
What The Matching Job Does
The job runs per shop, one at a time, in the background. You start it from the product mapping page and can leave the page; progress is saved in Redis and the results page stays available to review later.
Step by step:
1. Read both catalogs. Every Shopify variant with its product title, options, SKU, barcode, price, vendor, and product type. Every Odoo product with its template name, variant attribute values, default code, barcode, list price, and category. 2. Link exact matches by code. A variant whose barcode or SKU matches exactly one Odoo product is linked without asking the model. Variants that already carry an Odoo link are respected and skipped. 3. Build a shortlist for the rest. For each unmatched variant, candidates are ranked by word overlap across titles and option values, with vendor, product type, and price used as tie-breakers. The top 10 go forward. Products that share a SKU prefix are never cut from the list. 4. Ask one question per variant. The model, TypeSafe's Jev, is given the variant and its shortlist and asked which candidate it is, or none. It returns a probability per option. 5. Enforce one-to-one. Each Odoo product can be suggested to at most one variant per run. If two variants want the same product, the higher-probability match keeps it. 6. Write the review page. Results are paged 50 at a time with the probability, the match case, and the reason shown. Nothing is saved to Shopify or Odoo.
The whole run is read-only against both systems. The only writes happen when you click Map on a row, and those go through the same product mapping path as a manual link.
What The Confidence Score Means
The percent next to each suggestion is the probability the model put on that candidate relative to the others and to "none". It is not an accuracy guarantee. It is a sorting key and a filter.
In practice:
- Above 90 percent is usually the same product with a differently formatted title. Accept in bulk after a glance.
- Between 60 and 90 percent is worth a look. Often it is the right product but a sibling exists (a different size, a bundle versus a single).
- Below 60 percent is not suggested at all. The row stays unmatched and you map it by hand or fix the identifier.
The list is sorted by score with the match case (exact identifier or AI) and the reason next to each row. A sensible first pass is to map the rows at 90 and above, then read the rest.
Where The Same Logic Shows Up Again
The matching model is also used by the first-sync check. When the pre-launch scan finds a Shopify product that could link to more than one Odoo template, or an Odoo product with a plausible but not exact Shopify twin, it labels the row "Likely match" with the same probability and asks for a yes or no. So the review you do before go-live and the review you do on the mapping page are the same kind of decision, presented the same way.
When To Use It And When Not To
Use AI matching when:
- the store has history in both systems and identifiers were never reconciled
- you are migrating from another connector that stored links somewhere you cannot read
- a supplier catalog was loaded into Odoo with different codes than the storefront uses
- the mapping page has hundreds of unmatched rows and the team is doing them by eye
Skip it when SKUs and barcodes are already clean on both sides. The exact-match step will link everything and the model never runs. Also skip it for products you intend to exclude from sync; put the hold tag on them instead so the job ignores them.
Common Mistakes
Mapping everything at 60 percent without reading the reason. The floor exists so the suggestion is plausible, not so it is right. Review the middle band.
Running it before deduplicating Odoo. If Odoo has two active products with the same barcode, the one-to-one rule will pick one and the other looks unmatched. Fix the duplicate first; the first-sync check flags these.
Expecting it to create products. Matching links what exists on both sides. Products on one side only are a create decision, handled separately with its own preview.
Treating the link as temporary. Once saved, the link is a persistent metafield on the Shopify variant holding the Odoo id. Every later sync reads it. If a link is wrong, break it on the mapping page rather than re-running matching.
Frequently Asked Questions
What does AI product matching do when the SKU already matches? It links the variant directly and never sends it to the AI. Exact barcode and SKU matches are handled by code first.
Can two Shopify variants get matched to the same Odoo product? Not in one run. Each Odoo product can be suggested to at most one variant per run. If two variants compete, the stronger match wins and the other is left for review.
Are AI product matches applied automatically? No. Matches appear on the product mapping page with a match percent, the match case, and the reason. You map each row you agree with.