Personalization pitfalls: over-fitting to a customer's last purchase

This is the most common way personalization embarrasses itself, and it is fixable with a few rules rather than a new engine. The rules below apply whether you run a recommendations app, a hand-built Klaviyo flow or a UGC gallery that reorders by what a shopper looked at.

Why a single order is weak evidence

A purchase tells you that someone, on some day, for some reason, paid for a thing. It does not tell you who it was for, whether they liked it, or whether they will ever want anything like it again. A pattern across several orders answers those questions far better: the same size four times, the same colour family, the same price band.

An illustrative example. A customer has bought four pieces in neutral linen over eighteen months, all size M. Her fifth order is a bright children's raincoat, shipped to a different postcode in November. A recency-weighted engine now leads with children's outerwear. A pattern-weighted engine files the raincoat as probably a gift, keeps leading with linen, and waits to see whether a second children's order follows.

The five purchases that lie

The last row is what makes this hard. The same observation (a recent order that doesn't match history) can mean "ignore this" or "update everything". The engine cannot tell the difference from one order, so the right move is to wait for a second signal rather than guess.

The fix: pattern weight over recency weight

Recency still matters, just not as the dominant term. The workable structure has three parts: every signal decays over time, consistency across orders adds weight, and a single out-of-pattern order is capped until something corroborates it.

A worked scoring example. Using the illustrative customer above, give each order one point to its category, halve the value of anything older than six months, and cap a single-order category at one point until confirmed. Linen scores 3 (two recent orders at one point each, two older ones at half a point); children's outerwear scores 1, capped. A pure last-order rule would have scored outerwear highest. Nothing about this needs machine learning, only the discipline not to let one row win.

Let explicit signals override inferred ones

The cheapest fix is to ask. Amazon's "Improve Your Recommendations" page lets customers exclude purchases, "for example, gifts you purchased", from being considered in their recommendations, which is a public admission that even the largest recommendation system on the web cannot reliably tell a gift from a preference by itself.

• At checkout. A "this is a gift" checkbox is useful for gift messaging anyway; route its value into your customer data so the order is excluded from the taste profile.
• In the account. A "don't use this for recommendations" option on order history, even if few customers use it, gives the ones who notice a bad rail a way to fix it.
• In preference centres. Size, fit and style preferences a shopper states should be read before any behavioural score, and shown back to them. See personalization without creepiness for label wording.

Gartner's 2025 research points the same way: it recommends shifting "from passive inference to active customer involvement", after finding that 53% of customers had a negative experience with personalized marketing and were 44% less likely to purchase again afterwards.

Don't recommend what they just bought

A close cousin of over-fitting is recommending the item a customer has just purchased. For durable goods (a sofa, a pair of boots, a coffee machine) the purchase closes the need, and the next useful recommendation is a complement or care product, not a second one. Consumables are the exception, where the right message is a replenishment reminder timed to typical usage. Split your catalogue into "repeat" and "one-and-done" categories and treat them differently in every flow. The UGC email retention playbook covers post-purchase sequences built around that split.

How to audit your own recommendations in an afternoon

1. 1Pull twenty customers whose latest order sits outside their usual categories.
2. 2Look at what your site, emails and ads currently recommend to each of them.
3. 3Count how many recommendations are driven by that single latest order.
4. 4Check whether gift messages or alternative ship-to addresses are available in your customer data at all. If not, that is the first thing to fix.
5. 5Find the decay setting (or its absence) for each behavioural signal your tools use.
6. 6Test any change with a holdout rather than a before-and-after comparison; the A/B testing guide explains the setup.

Where UGC galleries fit

Galleries of customer photos and videos have an advantage here: they can reorder rather than exclude. Idukki's widget uses that approach. With retargeting enabled on an embed, UGC tagged to products a shopper recently viewed or saved is boosted up the gallery, but nothing else is filtered out, and viewed-product signals expire after a window the merchant sets. A shopper who glanced at a one-off gift still sees the rest of the catalogue's content, and the nudge fades on its own. The broader case for segment-level ordering is in personalizing UGC display by shopper segment, and the data foundation in first-party data after cookies.