Managing recommendations

What a recommendation set is

A recommendation set serves product recommendations such as "frequently bought together", related products and what's trending, worked out from your own store's order history. You turn on the strategies you want, optionally pin or exclude particular items, and then your workflows ask the set for recommendations.

The Recommendations page is where you set all that up and see how it is performing. The set is used from a workflow with the Recommendations connector, which also records the views, clicks, cart adds and purchases the strategies learn from.

Creating a set

1. Open Recommendations and select New recommendation set.
2. Enter a Name, for example "Cross-sell".
3. The Slug follows the name as you type and can be edited. Workflow steps refer to the set by this slug.
4. Optionally add a Description. It is only for your own reference and is never shown to shoppers.
5. Select Create.

You land on the new set. It starts with no strategies on, so next you choose which to turn on. The list shows each set's status, its strategies and when it was last computed, and you can search it by name or slug.

The tabs of a set

A set has five tabs: Overview, Strategies, Overrides, Preview and Settings. The header shows the set's status (Active, Computing, Error or Deactivated), its slug and when it was last computed. Recompute now recalculates the strategies straight away; it is disabled until at least one strategy is on, and the page updates while it runs. If a recompute fails, the error is shown in the header.

Strategies

Each strategy has its own switch, and you can turn on as many as you like. A workflow asks for one strategy by name each time it calls the set, so strategies are alternatives, not a blend. Select Save strategies after changing them.

Strategy	What it recommends
Frequently bought together	Products bought in the same order. Transparent, but only catches things bought together in one visit.
Collaborative filtering	Products bought by the same kind of customer, even on separate visits. Catches complementary items that strict counting misses, but needs more history.
Trending	What's selling right now, weighted toward recent orders. Store-wide, or scoped to a category.
Related items	Other items whose content is similar, from a search index attached to the set. Answered live, so it works from day one with no order history.
Personalized	Not available yet. It needs customer identity, and is marked "coming soon".

Each strategy has settings you can tune:

• Max results: how many items to keep per product.
• Lookback (days): how far back into order history to look.
• Min co-occurrence (frequently bought together): ignore pairs bought together fewer times than this.
• Min shared customers (collaborative filtering): ignore pairs with fewer shared customers than this.
• Half-life (days) (trending): how fast an order's weight fades. Smaller means more weight on recent orders.
• Scope facet (trending, optional): an extra attribute on the purchase, such as a category, to group trending scores by. Leave it empty for store-wide trending.

Each computed strategy shows its state: how many pairs it has computed, Computed, but found nothing (it needs more order history), or Not computed yet.

Related items is switched off until you attach a search index under Settings. Once on, it is answered live from that index and has nothing to precompute.

Overrides

Use overrides to force an item to always appear (Pin) or never appear (Exclude) in a set's results.

1. Under Add override, choose the Action: Pin or Exclude.
2. Enter the Item, using the same key your products have, for example SNOWBOARD-POWDER.
3. Optionally set Scope to item to limit the override to recommendations shown for one product. Leave it empty to apply everywhere.
4. For a pin, optionally set a Position: 1-based, where ties are broken by list order. Leave it empty for the first available position.
5. Select Add override.

Overrides are listed with their action, item, scope and position. Edit one or remove it with its icon. Removing happens immediately, without a confirmation. A filter box appears once you have more than five. Overrides take effect automatically whenever a workflow gets recommendations from the set.

Preview

Preview shows what a shopper would be shown, as a single product page or a whole cart, using the set as it is now.

1. Turn on at least one strategy first.
2. Choose a Strategy from the ones that are on.
3. Enter Item key(s): one SKU, or several separated by commas to preview a cart.
4. Select Preview.

Results are ranked, each with a score bar and value, and items matched by a pin are marked Pinned. The result also shows a request ID, which you can copy and pass to Record Recommendation Feedback to test how clicks and purchases are attributed. If nothing comes back, the product may not have enough order history yet, or the set hasn't been computed.

For a cart, results are pooled across the items and an item's score is added up across every cart item it goes with. Items already in the cart are left out.

Overview

Overview shows how the set is doing. Choose Last 7 days, Last 30 days or Last 90 days for the performance numbers.

• Event volume: Views, Clicks, Add to cart and Purchases recorded store-wide. If it shows no events at all, nothing is being recorded yet, and strategies have nothing to learn from. Add the Recommendations connector's Record actions to your storefront and checkout workflows.
• Recommendation performance:
  • Shown: how many times recommendations were requested.
  • Clicked: the share that led to a click.
  • Led to a purchase: the share that led to a purchase.
  • By strategy: the same numbers per strategy, with a Preview button for each.

Clicked and Led to a purchase only fill in when your workflows send the outcome back with Record Recommendation Feedback, using the request ID that Get Recommendations returned. Without it, recommendations are served but their results can't be measured.

Settings

• Active: turn the set off to stop serving recommendations. Strategies, overrides and computed data are kept, and Get Recommendations returns an error until it is turned back on.
• General: change the Name and Description, or the Slug. Changing the slug breaks any workflow step still using the old one, so update those steps first.
• Search Index: attach an index to turn on Related items, which works without order history. Choose None for precomputed strategies only. When an index is attached, Normalize item keys helps if your item keys don't exactly match the index's own document keys, for example because of different casing. Detaching the index turns off Related items and leaves the other strategies alone.
• Delete this recommendation set: removes the set, its strategies, overrides and computed data permanently, and workflow steps that reference it will start failing.

Using a set from a workflow

1. Add a Get Recommendations step, choose the set by its slug and a strategy that is on, and give it the product as the anchor, or the cart's items.
2. Return the items and the request ID to your storefront. The items are item keys with scores, so look up titles and images separately.
3. When a shopper clicks or buys one, call Record Recommendation Feedback with the request ID.
4. Record views, cart adds and purchases with the other Record actions so the strategies keep learning. Record Purchase is the data that frequently bought together, collaborative filtering and trending are built from.

The Recommendations connector lists every action and field.