Product Recommendations for Items With No Sales History
Handle recommendation cold start for new products, shoppers and sites. Compare catalogue evidence, images, contextual signals and carefully measured fallbacks.
Product recommendations for new products with no sales history
New products should be evaluated on the evidence available about them, even before they have purchase history. Separate item cold start from a new visitor or a new website, then choose a recommendation approach that fits the placement and the quality of the catalogue.
Identify which cold start you have
A new item can have rich attributes and images but no interactions. A new shopper can provide a query, a product view or a basket without a purchase history. A new website may have little behavioural evidence across the entire catalogue. These conditions need different responses; calling all of them “no data” conceals useful information.
Start with the task. On a product page, the current item can anchor a comparison or complement. In a basket, selected items provide context. On a blank homepage, the evidence is weaker, so a restrained editorial or contextual starting point may be more appropriate than a claim of deep personalisation.
Use product evidence carefully
Attributes and imagery can describe relationships before sales accumulate. PA's recommendations page describes using natural-language and visual signals for new items. That is a capability to evaluate against your catalogue, not a guarantee that every inferred relationship is correct. Recommendation approach.
Check that the selected relationship fits the job. Visual resemblance can help a shopper compare styles. It does not establish that an accessory fits a device, that a product meets a dietary requirement or that two items are useful together. Use authoritative constraints where the relationship requires them.
Design a fallback that stays useful
The proposed sequence is to use the strongest available evidence for the placement, filter out ineligible items and return only recommendations that meet the agreed quality threshold. If the evidence is insufficient, show a clearly labelled broader collection or omit the recommendation rather than disguising a bestseller list as personal knowledge.
As behaviour accumulates, compare it with the original catalogue-based recommendation. New signals should earn their influence through evaluation; popularity alone can reinforce whichever items received the most early exposure.
Measure catalogue discovery as well as clicks
Compare relevance on new and established items, the share of eligible new items receiving useful exposure, downstream engagement and business outcomes. Record which items had sufficient data for each method. A recommendation experiment that excludes all difficult new products does not test cold-start performance.
PA's anonymised fashion range-mix case reports 21.1% more SKUs in the group contributing the top 80% of turnover. The public page does not disclose sample size or test period; treat it as a reported case result rather than an expected outcome for your store. Published fashion case. This measures the breadth of the revenue-contributing range; it does not isolate the performance of newly launched products.
Test before widening exposure
Build a small panel of recently added products, incomplete records and established items. Have category specialists assess the recommendations for the intended placement. Use a controlled live test only after the examples meet the quality and eligibility requirements.
Evidence
- PA reports 21.1% more SKUs in the group contributing the top 80% of turnover in one anonymised fashion case. (Public page omits duration and sample; headline says doubling despite +21.1% body metric, so do not repeat headline. This measures the breadth of the revenue-contributing range; it does not isolate the performance of newly launched products.)
Downloads
- cold-start-evaluation.csv (illustrative template)
Sources
- AI Product Recommendation Engine, checked 2026-10-25
- Multi-brand fashion retailer: range mix, checked 2026-10-25
How Particular Audience does it
Frequently asked questions
Can recommendations work without purchase history?
They can use product attributes, imagery and available context, but the result depends on the quality of that evidence and the recommendation task. Evaluate new items directly instead of assuming an approach that works on established products transfers unchanged.
Is visual similarity enough?
It can support style comparison, but it does not prove compatibility, suitability or complementarity. Match the relationship to the shopper’s task and use authoritative constraints when a wrong recommendation would misrepresent how products fit or function together.
What should a new store show first?
Use the context it actually has, such as the current product, query or basket. Where evidence is weak, a curated or clearly labelled popular collection can be appropriate. Avoid presenting generic recommendations as though they reflect a known personal preference.