A recommendation row is a small assortment with a large job: give someone a useful next step. More products can improve the chance of a match. They can also make it harder to decide whether to inspect anything.
The decision is how many products belong together in this placement, how clearly they differ and what shoppers think the list represents. A successful shortlist earns its space through useful choices, not the number of cards it can fit.
🔎 What research says
Test recommendation length where it will be used, and measure entry into the set as well as purchases. Retain the best match, make real differences understandable, and check whether the list’s source changes how people explore it. A winning length in one retargeting message is not a universal layout rule.
🗺️ In this guide
1. Test list length in its own placement
📈 Recommendation
Choose recommendation length by testing that particular message or page. Keep the first product, ranking logic and commercial conditions consistent so the test can reveal what adding alternatives changes.
For a shopper who asked about a backpack but has not ordered, compare the existing follow-up with a shorter or longer alternative. This is our proposed store test; the study did not test this backpack message.
Measure purchases from the recommendation as well as overall orders and margin. A stronger message result can still displace purchases that would have happened elsewhere.
🎓 Findings
Long and colleagues (2025; online 2024) randomized recommendation length in Tmall and Taobao customer-service messages. Messages followed a product inquiry when the shopper had not ordered within five minutes. The analysis used the first viewed message for each of 1,615,325 consumers. Each message showed one to four products, including the product already discussed.
Same-day purchases through the message occurred for 3.08% with one product, 5.13% with two, 4.53% with three and 4.29% with four. Moving from one to two added 2.05 percentage points—about 21 extra purchasing consumers per 1,000 message viewers. Moving from two to three lost 0.60 percentage points—six fewer per 1,000. These are arithmetic restatements of this experiment, not forecasts.
The reported comparisons between adjacent list sizes were statistically clear. Two performed best among the four tested lengths; the paper did not establish a universal two-product rule.
🧠 Why it works
An additional product can offer a useful alternative. Another can add enough comparison work to discourage action. The point where that balance changes depends on the placement and shopper task.
✋ Limitations
This was post-inquiry retargeting through a merchant messenger, not an ordinary product-page carousel or a new visitor browsing a category. Only message-linked purchases on the day of receipt were counted. The products fit without scrolling even in the four-product condition; the result does not compare scrolling interfaces.
2. Measure who never opens a recommended product
📈 Recommendation
Measure whether people enter the recommended set before judging the products they click. A report covering only clickers misses shoppers who turn away at the list itself.
Report any-product clicks divided by recommendation viewers, then message-linked purchases divided by the same viewers. Also inspect purchases among clickers, using that different denominator explicitly.
If fewer viewers open a product, test the initial recommendation set before redesigning only the destination product pages. Keep a comparison group to distinguish the effects.
🎓 Findings
In the same Alibaba experiment, the chance of clicking any recommended product fell from 20.73% with two products to 19.18% with three. That is 1.55 percentage points, or about 16 fewer people starting a product search per 1,000 message viewers.
The researchers decomposed the accompanying purchase-rate drop and attributed about 64% of that specific two-to-three decline to the lower rate of starting a search. For the three-to-four comparison, that share was about 30%. These figures describe where the measured decline occurred; they do not mean that 64% of shoppers were overloaded.
🧠 Why it works
A list can fail before anyone studies its details. Improving the clicked product page cannot help someone who never opens it. Separating entry from later choice tells an operator where to investigate first.
✋ Limitations
The decomposition uses the experiment’s search and purchase rates; it does not independently establish a psychological cause. Regret, perceived relevance and other explanations remain possible. Do not add this percentage to a conversion lift or apply it to another store’s abandonment.
3. Keep the best match and useful differences
📈 Recommendation
Test a set with differences shoppers can use, while retaining the product you expect to fit best. Do not fill a shortlist with nearly interchangeable options—or add poor products just to make one stand out.
For camera recommendations, explain real trade-offs such as portability, lens compatibility and low-light performance. Test whether people can identify a suitable option without changing the available range.
Measure the intended customer action and returns, not just time on the list. More comparison time may reflect difficulty rather than interest.
🎓 Findings
Tsekouras and colleagues (2020; online 2019) analyzed 9,330 visits to a Dutch mortgage-recommendation website and ran a separate simulated camera-choice experiment with 399 participants. The field outcome was requesting a quote, not signing a contract; the camera task did not establish paid retail sales.
Using the fitted mortgage model, the researchers simulated removing one intermediate option while retaining the most attractive product. Predicted quote-request probability increased. This was a computer estimate, not an observed rollout gain. A model of quote requests cannot tell an ecommerce store how many paid orders a similar change would create.
The camera experiment also found a negative response to closely matched alternatives among highly engaged participants. Its exact sales impact cannot be read from the reported statistical coefficient.
🧠 Why it works
A set can score well on predicted attractiveness yet make its alternatives hard to distinguish. Useful differences may help shoppers identify which item deserves closer inspection. Whether that improves your customers’ decisions needs its own test.
✋ Limitations
The field analysis was observational and its changes were simulated, rather than randomly implemented. Most evidence concerns mortgages and practical camera choices. The paper cautions that much worse alternatives can be ignored or undermine trust. The pictured camera attributes are our application, not the exact tested screen.
4. Test list length with its stated source
📈 Recommendation
A short expert selection and a short popularity list can communicate different things. Test the combination of length and a truthful explanation of where the recommendations come from.
For a coffee range, distinguish a named buyer’s selection from items ranked by customer ratings. Use the label that matches the actual process; do not present an automated list as expert judgment.
Measure exploration beyond the first item, completed orders and suitability. A shift toward more expensive products alone is not evidence of better customer outcomes.
🎓 Findings
Mejía and Guesmi (2026) report four studies involving 1,100 participants, comparing shorter and longer lists attributed to experts or rating aggregators. The publisher’s summary reports that short expert lists prompted more product learning and less selection of the first item; purchase intention and satisfaction depended on how people understood the recommendations.
The accessible summary does not report the sizes of those effects or the tested list lengths. The participant count describes the research program; it cannot justify a sales-lift forecast or a particular number of products.
🧠 Why it works
Shoppers may read a short expert list as a deliberate selection rather than a compressed popularity ranking. That interpretation can change how they explore it. This is a reason to test source and length together, not to manufacture an expert badge.
✋ Limitations
Full-text estimates were not accessible in this review. The summary mixes behavioral measures with purchase intention, and it does not establish a paid-order effect for a coffee store. Keep this recommendation as a conditional test until the detailed methods and results are recovered.
💡 The takeaway
Give shoppers a useful next step, then test whether they take it.
Start with one recommendation placement. Compare a manageable set of lengths, preserve the strongest match and inspect where people stop progressing. Roll out only when the change improves suitable purchases and commercial outcomes in your own setting.
📚 Sources and study types
Long, X., Sun, J., Dai, H., Zhang, D. J., Zhang, J., Chen, Y., Hu, H., & Zhao, B. (2025; online 2024). The Choice Overload Effect in Online Recommender Systems. Manufacturing & Service Operations Management, 27(1), 249–268. Randomized post-inquiry marketplace messaging experiment; same-day purchases through message links.
Tsekouras, D., Dellaert, B. G. C., Donkers, B., & Häubl, G. (2020; online 2019). Product set granularity and consumer response to recommendations. Journal of the Academy of Marketing Science, 48, 186–202. Observational mortgage-quote data, model simulations and a controlled simulated camera-choice experiment.
Mejía, V. D., & Guesmi, S. (2026). The Singular Effect of Expert Short Recommendation Lists. Psychology & Marketing, 43, 2279–2304. Four studies of list length and source; this guide uses the publisher summary without an unverified numeric effect.
