How do you personalize recommendations when inventory is thin?

October 15, 2026

Short answer: When inventory is thin, availability is the personalization. A recommendation for a sold out size is worse than a generic pick that can ship today, so ranking should weigh stock depth, size coverage, and restock timing ahead of preference scores until the catalog recovers.

Thin stock breaks the engine's core assumption

Recommendation models score products by predicted preference and present the best matches. The quiet assumption is that the best match can be bought. In a thin catalog that assumption fails constantly: the top pick is out of the shopper's size, the second is a single unit in a warehouse three zones away, and the rail becomes a museum of almost available products.

Shoppers read that as incompetence, not bad luck. A store that knows their taste but not its own shelves looks worse than a store that never personalized at all, because the promise was higher. The failure is not the model's taste. It is the missing inventory term in its ranking.

Putting stock depth inside the score

Multiply preference by availability rather than filtering after ranking. A strong match with full size coverage should outrank a perfect match with one unit left, because the perfect match will be wrong for most of the shoppers who see it. Size level stock matters for apparel: aggregate counts hide the broken runs that frustrate real buyers.

Add a decay for stock age in constrained categories. Products sitting thin for weeks are usually thin for a reason: broken sizes, a color nobody wants, a fit problem driving returns. Preference data collected when the product was fully stocked describes a product that, in practical terms, no longer exists.

What to recommend when the right product is gone

The best substitute is rarely the most similar looking product. It is the product that serves the same job at the same price with stock to sell: the alternate fabric in the same cut, the neighboring category the shopper also browses, the bundle that includes an available equivalent. Substitution logic should be built by merchandisers in advance, not improvised by the model at the worst moment.

Restock capture belongs in the same rail. When the true first choice is gone, an honest out of stock note with a notify option, placed beside a buyable substitute, respects the shopper's intent twice: it offers something now and remembers what they actually wanted. Stores that hide the sold out favorite entirely lose both chances.

Throttling personalization during sellouts

During launches and sale peaks, segments of the catalog can sell out in an hour while the engine keeps routing traffic by yesterday's preferences. Build a thin inventory mode: when category stock depth drops below a threshold you set, ranking shifts weight from preference to availability automatically and frees merchandisers from manual firefighting.

Communicate the mode to the team. Customer service and email marketing should know when the site is in scarcity mode, because sending a campaign into a category the engine is actively routing around creates the exact sold out landing experience everyone is trying to avoid.

The data to fix before the next thin season

Thin inventory exposes data debt: inaccurate size level counts, slow warehouse sync, restock dates nobody trusts. Log every recommendation impression that ended in an unavailable variant for a month and the pattern will name the fixes in priority order, from sync lag to phantom stock.

When the catalog refills, compare the thin period's substitute performance against normal recommendations. The substitutes that sold and stayed sold, with low returns, are permanent additions to the recommendation map. Scarcity, handled well, teaches a store what its shoppers actually accept as an alternative, which is some of the most valuable merchandising knowledge there is.

Keep reading: How do you personalize product pages for shoppers who arrived from social video?, How do you personalize recommendations for gift buyers shopping for someone else?, and How do you personalize recommendations for first-time visitors with no history? cover the nearby cases in this series.

Frequently asked questions

Why not just hide out of stock products from recommendations?

Hiding is part of it, but the harder problem is near stock: products technically available with one size or one unit left. Ranking needs stock depth inside the score, not a binary filter at the end.

What is the best substitute for a sold out recommendation?

A product serving the same job at a similar price with real stock behind it, chosen from substitution relationships merchandisers build in advance, rather than the visually closest product the model finds.

Should restock alerts appear inside recommendation rails?

Yes, next to a buyable substitute. The shopper gets something they can purchase now and a promise about the product they wanted, instead of a dead end.

How does a store know it is in thin inventory mode?

Set a category level stock depth threshold in advance. When depth falls below it, ranking weight shifts from preference to availability automatically until the catalog recovers.