Learning from your store
A learning layer. Your rules still lead.
Bring your store’s catalog, historical sales and discovery signals into collection ranking. Indexa Intelligence compares leading candidates, shows which positions changed and keeps publication in your hands.
Outerwear · Intelligence
Illustrative demo · Fictional products and inputs
1 → 1Leather biker jacket
Campaign pin · Fixed position
3 → 2Blush wool overcoat
Moved by Indexa Intelligence
2 → 3Sky-blue trench coat
Moved by Indexa Intelligence
Why this order?
Blush wool overcoat · Won 1 of 1 observed comparison
Preferred over Sky-blue trench coat in this example. The jacket stays pinned. Win share orders evaluated products; ties retain the baseline order and hard rules still apply.
- Units sold · 7 days
- 23
- Units sold · 30 days
- 87
- Available stock
- 46
- In stock
- Yes
Illustrative mapped inputs, not live evidence. In the console, this panel shows actual inputs and observed comparisons, not causal feature importance.
Saving a draft does not change Shopify. Review, then explicitly publish.
Your data, your checkpoints
Training datasets and versioned model checkpoints are scoped to your tenant. Mapped catalog fields, available Shopify historical sales aggregates and Indexa events supply the inputs; missing history is reported. Intelligence is optional and disabled by default for each store.
A focused window, not the whole catalog
Rules create the baseline. Intelligence considers an initial window of at most 32 positions within strategy depth. Your store’s configurable topN and pin exclusions can make the evaluated set smaller. The rest of the order, collection membership, priority groups and exact pins are preserved.
See the movement and the evidence
Changed products carry “Moved by Indexa Intelligence” with their before and after positions relative to the rules baseline. “Why this order?” shows observed comparison wins, opponents and actual numeric or boolean mapped inputs. These describe decisions and inputs, not causal feature contributions.
Training runs away from shoppers
Asynchronous cloud training runs in a separate worker, never in the shopper request. A completed run is evaluated against the active model or approved base on an untouched holdout before it can become a candidate. Offline metrics are quality checks, not evidence of revenue lift.
Review, promote, roll back
Training completion does not activate a checkpoint. Review the evaluation before controlled promotion. Roll back to a previously promoted checkpoint with its matching feature configuration, without retraining.
Preview now. Publish when ready.
An AI preview can run while a collection strategy remains a disabled draft, provided tenant Intelligence is enabled. Saving changes Indexa configuration only. Publishing to Shopify and enabling scheduled updates are explicit actions. Publication computes a fresh order; the preview is not a frozen snapshot.
Clear fallback, controlled publication
If AI cannot produce a valid collection preview, the console reports Rules fallback and blocks AI publication. For eligible shopper collection requests, an unavailable or timed-out model returns the prior deterministic order.
A defined place in discovery
Shopper reranking applies only to the first page of an eligible collection listing without a search query or explicit sort. Hard collection boosts bypass AI. Collection planning has its own inference deadline, separate from the shopper request budget.
Up next