Product recommendations

Recommending the right product at the right moment is the core of what Zubby AI does. The agent suggests products inside chat - cross-sells, bundles, alternatives, “more like this” - and every suggestion is scored, logged, and attributed. The Recommendations page is where you see how that engine is behaving and shape it.

Find it at Growth → Recommendations in the sidebar. The page is a configuration and insight surface; the actual boost and exclude rules are managed under Growth → Rules & Voice.

How recommendations are generated

Recommendations are not a static “customers also bought” table. They’re produced live by the agent’s function- calling tools as a conversation unfolds - searching the catalog, comparing items, building bundles and outfits, and finding alternatives - all grounded in your synced catalog and embeddings. The ranking layer then orders candidates using:

• Conversation context - what the shopper said they want, their budget, prior messages.
• Product relevance - semantic similarity from the product embeddings.
• Your rules - boosts and exclusions you’ve set (below).
• Real outcomes - whether past shoppers clicked, added to cart, or purchased the candidate.

Stock and profit in the ranking

One scoring step applies stock and profit signals to every recommendation surface: chat answers, cross-sells, bundles, similar items, and storefront carousels.

• Sold-out products are never recommended, on any plan. Stock is judged per variant: a variant counts as sold out only when your platform tracks its stock, stops selling at zero, and it is at zero or below. Variants that are untracked or allow backorders stay recommendable. If a shopper asks for a sold-out product by name, the agent still finds it and tells them it is unavailable.
• Very low stock is demoted, based on the units left or the stockout forecast, with the same exception for products a shopper asked for by name.
• Profit-aware ranking (Growth and above, opt-in) - turn on Margin-aware recommendations under AI Setup. Among products of similar relevance, higher-margin and overstocked items rank a little higher and frequently returned items a little lower. The effect is capped at 15% either way, so it reorders near-ties and never buries a clearly better match. It needs product costs; see Profit & product costs.

Costs and margins are used only for ordering. They are never shown to shoppers or included in what the agent says.

The KPI row

The four cards at the top summarize the engine at a glance:

• Total Recommendations - all-time count of products the AI has recommended.
• Avg. Confidence - the mean confidence score across recommendations. 70%+ reads as “high confidence” in the UI. This is a model or ranking signal, not a measured probability of a correct recommendation. Inspect source facts and real outcomes before treating a higher score as better shopper experience.
• Boost Rules - how many boost rules are active (and the total defined).
• Exclude Rules - how many exclude rules are active (and total).

Boost and exclude rules

Rules are how you put a thumb on the scale without touching the model. The Active Rules panel summarizes them, and the toggle on each row turns a rule on or off in place. Manage the full set under Rules & Voice (the Manage All Rules button).

Each rule targets a scope - Global, Collection, Product, or Category - so you can boost an entire collection for a seasonal push or exclude a single SKU.

Confidence scoring

Every recommendation carries a confidence score from 0 to 1, derived from conversation context, product relevance, and active rules. In the dashboard it’s shown as a percentage badge: 80%+ is green (strong match), 50–79% is amber, and below 50% is muted. Consistently low confidence usually points at a catalog or knowledge gap rather than a tuning problem.

Attribution & outcomes

Recommendations only get better if Zubby can see what happened after one appeared. The widget records the outcome of each suggestion - clicked, added to cart, or purchased - and those outcomes feed back into ranking.

Revenue is attributed on a last-touch model within a 30-minute window: an ordered product must match a recommendation for the same store and shopper identity within the window before the order. The full order receives credit, which can over-credit influence and does not prove incremental revenue. The three insight cards on the page - Outcome signal, Knowledge dependency, and Ranking inputs - help you sanity-check that impressions, outcomes, and rule pressure are all moving in the same direction.

Recent Recommendations table

The bottom of the page lists the last ten products the AI recommended, with columns for:

• Product - links straight to its catalog detail page.
• Confidence - the score badge described above.
• Context - which tool produced it (e.g. get cross sell), its rank position, and the page type it appeared on.
• Outcome - purchased, added to cart, clicked, or no outcome yet.
• Recommended - a relative timestamp.

When performance looks off, compare your rule pressure against the recent low-outcome rows here to see whether a boost is over-promoting something shoppers don’t actually want.

How to boost a hero product

The most common tuning task, start to finish. Why boost instead of pinning it everywhere: a weight nudges ranking while still letting relevance decide, so the product surfaces when it fits and never hijacks unrelated conversations.

1. Open Growth → Recommendations and check the Recent Recommendations table - confirm the product really isn’t appearing before you add rule pressure.
2. Click Manage All Rules, then Add Rule, and pick the Push a collection harder preset (or set type to Boost with scope Product for a single SKU).
3. Pick the target from the suggestion chips so it matches your synced catalog, and set a weight between 1.5 and 2.0 as an illustrative starting range to test, not a guaranteed effective setting.
4. Save and confirm the rule appears in Active Rules. Repeat the same relevant queries before and after the change, plus unrelated and out-of-stock queries as guardrails. Review Recent Recommendations after sufficient traffic; reduce or remove the boost if irrelevant results increase. A saved boost does not guarantee an impression.

Gotchas

• No recommendations yet? They only appear once the AI has chatted with shoppers and your catalog is synced. Check Catalog and the agent’s training status first.
• Confidence stuck low? Fix description gaps and knowledge coverage before fiddling with weights - ranking trust follows content quality.
• Attribution looks light? Outcomes depend on the widget capturing clicks and cart events, so make sure the widget is installed and tracking on your storefront.

Related

• Catalog & product sync - the source of every recommendable product.
• The AI agent - the tools that generate recommendations in chat.
• Knowledge base - coverage that lifts both copy and ranking trust.
• Reviews & UGC - social proof the agent cites alongside recommendations.