Recommendation block strategies

​

Interaction strategies

​

How interaction results are ranked

• Category backoff. When an anchor has little or no interaction history of its own (for example, a newly added product), Layers backfills the block with the top products for the anchor’s product category, then its parent category. Backoff results always rank below results computed from the anchor’s own history, and the anchor itself is never recommended.
• Same-category demotion. The purchase-driven recipes (Frequently Bought Together, Viewed Then Bought, and Customers Who Bought Also Bought) demote recommendations from the anchor’s own category below cross-category matches. The block favors products that go with the anchor over alternatives to it.

​

Product families and interaction strategies

• Behavioral signal from all family members is combined at the family level.
• Products are never recommended alongside another member of their own family, and a family’s members are not recommended for its own anchor products.
• Each family appears as one representative product in results. For families built from combined listings, Layers picks a representative that matches your combined listings display setting. A parent-only store shows the parent listing, and a child-only store shows a child.

​

Collection interaction strategies

​

Similar products strategies

​

Similar Products

• Results are recomputed daily when the existing data is more than three days old or new products have been added since the last run
• When a product’s content is updated (for example, after editing its title, description, or images), that product’s similar products are recomputed incrementally
• An hourly coverage check finds products that are missing precomputed results, such as products added through a bulk catalog import, and queues them for computation. It also rotates the oldest precomputed results so recommendations stay current as the catalog changes
• Products whose index data isn’t ready yet are retried automatically on the hourly cadence instead of being skipped
• All filters, facets, pagination, and sort orders continue to work as expected on precomputed results
• No configuration is required. Precomputed caching is enabled automatically for stores with active similar products blocks

• Match: recommendations must share one of the anchor’s values for the attribute. Multi-valued attributes (like tags) match any of the anchor’s values.
• Within % band: recommendations must fall within a percentage tolerance of the anchor’s value. Band rules are available only for numeric attributes, such as price. Enter the tolerance as a percent in the editor.

• Constraints filter both precomputed and real-time similarity results. A product that fails a rule is excluded, not just demoted.
• If the anchor has no value for a rule’s attribute, Layers skips that rule rather than emptying the block.
• On collection anchors, rules resolve against the representative product Layers selects for the collection.
• Layers automatically widens the similarity candidate pool for constrained blocks, so filtering doesn’t starve the block of results.
• If constraints still leave too few products, the minimum products safeguard and fallback chain apply as usual.
• Block preview in the dashboard reflects constraints, so you can verify the effect before publishing.

• “Similar Products” on product pages
• “You May Also Like” recommendations
• Visual discovery and exploration
• Collection-based similarity recommendations
• Category- or price-consistent alternatives (“more dresses like this one, in the same price range”)

​

Behavior-seeded similarity

​

Shop the Article

1. Layers extracts product links from the article’s body: <a> tags whose path matches /products/{handle}. Linked products appear first, in the order they’re linked, deduplicated. Handles resolve only against your own catalog.
2. Layers seeds a similarity search from the article’s indexed content and fills the rest of the block with matching products, excluding the linked products.
3. Linked products pass through the same eligibility checks as a manual block, so unpublished or channel-restricted products are skipped.

• “Shop this story” rails under a blog post
• Gift guides and lookbooks where the article links the featured products
• Editorial content that should convert without manual curation

​

Manual strategy

• Hand-picked products: Select specific products by ID in the dashboard. The display order matches the order you define. When hand-picked product IDs are configured, they take priority over collection-based selection.
• Per-anchor product picks: For product-anchored blocks, you can hand-pick different recommendations for specific anchor products. When a customer views a product that has anchor-specific picks configured, those picks are returned instead of the default list. If no anchor-specific picks exist for the current product, the block falls back to the default hand-picked products.
• Collection anchor blocks: The collection is determined dynamically from the API request’s anchor_id parameter (no collection selection needed in dashboard)
• None anchor blocks: Select a collection to pull products from in the dashboard
• Optionally apply a sort order for custom sorting
• Apply merchandising rules: For collection-sourced blocks, toggle on to inherit the collection’s merchandising rule pins and banners. Off by default.

​

Apply merchandising rules

• Pinned products surface at the top of the block, matching their pinned positions on the collection page.
• Banners configured on the collection’s merchandising rule are returned alongside the products.
• Other rule effects (for example, sell-out deletion) apply to the block’s results.

​

Per-anchor product picks

1. You configure a default set of hand-picked products for the block
2. Optionally, you add per-anchor overrides for specific products
3. When the block is requested with an anchor_id:
   • If the anchor has specific picks configured, those products are returned
   • Otherwise, the default hand-picked products are returned

• “Best Sellers” block (collection of top products, sorted by sales)
• “New Arrivals” block (collection of recent products, sorted by date)
• “Editor’s Picks” (curated collection with custom order)
• “Staff Picks” (hand-picked products in a specific order)
• Seasonal promotions (collection of seasonal products)
• “Complete the Look” with product-specific accessory pairings
• Targeted cross-sells that vary by anchor product

​

Metafield reference strategy

• metafield_namespace (required): the namespace of the reference metafield
• metafield_key (required): the key of the reference metafield
• metafield_scope (optional): product (default) or variant

{
  "metafield_namespace": "custom",
  "metafield_key": "wear_it_with",
  "metafield_scope": "product"
}

• Products appear in the exact order they’re authored in the metafield. Nothing re-ranks them.
• The block applies the same eligibility rules as a manual block. It skips products that are deleted, disabled for blocks, unpublished on the request’s channel, outside the request’s B2B catalog, hidden by your combined listings display setting, unavailable in the request’s market, or out of stock (when hide out-of-stock is on).
• The block’s matching rules, exclusions, safeguards, and pagination all apply as usual.
• If the anchor product has no value for the metafield, or none of the referenced products are eligible, the block returns empty and defers to its fallback chain.

• “Wear it with” outfit pairings authored by your merchandising team in Shopify
• “Refills for this” or “Replacement parts” rails driven by existing product relationships
• Any cross-sell relationship you already maintain in reference metafields

​

Set strategies

​

Photographed Together

• Layers detects the individual garments in each product’s images and matches every companion garment to a product in your catalog, with a confidence score. Layers discards weak matches below the confidence threshold.
• Matching respects what the photograph shows. A companion only matches catalog products in the detected garment’s category. When the detected garment has a color, Layers prefers a candidate whose variant options name that color. The preference only applies within a small similarity band of the top-scoring candidate. Olive shorts still favor the olive colorway when it’s a close match. A clearly better product wins over a same-color one that’s a weaker match. When no candidate names the color, or the closest color match is outside the band, Layers falls back to the most similar candidate.
• Color agreement is read from variant options only (for example, a Color option whose value contains the word olive). Product titles and variant titles aren’t consulted.
• The set for an anchor product combines both directions: companions detected in the anchor’s own photography, and products whose photography features the anchor.
• When a companion match resolves to a specific variant on color, Layers pins that variant as the product’s first_or_matched_variant in the response. The storefront can then show the exact colorway the photograph shows. Without a color match, the product’s default variant is used.
• Results are sorted by match confidence, deduplicated by product, and capped per anchor so a heavily photographed hero piece doesn’t fill every other product’s set. The cap counts only eligible companions. A detected match that is unpublished on the shopper’s channel, disabled for blocks, or excluded by the block’s rules doesn’t use up a slot. The next eligible match takes its place.
• Matches are refreshed periodically, so a catalog that has since gained the right product picks it up on the next pass.
• Replacing a product’s photography doesn’t break the set. Layers retires matches built from the old image and rebuilds them from the new shots on the next refresh. Matches from the product’s other images keep serving in the meantime.

• The block’s matching rules, exclusions, and safeguards apply as usual. An apply_filter rule drops members before pagination and totals.
• With Hide out of stock enabled, an unavailable pinned variant drops the pin, not the product. The garment the photograph shows stays in the set with its normal available variant. Availability follows the same search rules under a narrowed inventory location scope. A pinned variant that Shopify keeps selling when out of stock, or whose inventory is untracked, keeps its pin.
• While detection and matching are still running for your catalog, the block returns the same “data is being computed” state as other precomputed strategies and defers to its fallback chain. In the dashboard, the block shows a training badge in the blocks list and a notice on its edit page. Both appear from the moment you publish the block until the first detection pass completes.
• A first pass with some photos still unprocessed doesn’t hold the block back. The block goes live with the matches it has, and Layers completes the remaining matches on later passes.

• “Shop the look” rails on product pages
• “Complete the set” cross-sells grounded in your own styling
• Outfit-driven merchandising without manual curation

​

Related content strategy

​

Related Reading

• On product and cart anchors, Layers seeds the search from the anchor product (or the products in the cart) and finds the content most similar to it.
• On article anchors, Layers seeds the search from the current article’s content and excludes the article itself from the results.
• Results are ranked by semantic similarity and deduplicated. Each item includes the content’s title, handle, summary, author, tags, image, publish date, blog title, and URL.

{
  "content_type": "article"
}

• “Related reading” rails on product pages that link to buying guides
• “Keep reading” rails under blog posts
• Care guides or sizing pages surfaced from the cart

​

Personalized strategies

​

Trending strategies

​

Contextual strategies

​

Strategy configuration keys

​

Interaction

​

Collection interaction

​

Interaction ordering

{
  "time_window": "30d",
  "order_by": "confidence_score"
}

​

Trending

​

Contextual

​

Personalized

​

Manual

​

Metafield reference

​

Set

​

Related content

​

Similar products constraints

​

Prefer anchor option

{
  "prefer_anchor_option": true
}

​

Validation

​

Strategy types by anchor type

​

Strategy availability by anchor type