A recommended-product card has little room and several jobs. It needs to show what the item is, help someone decide whether to inspect it and avoid pulling them away from a suitable purchase.
The practical decision is which information earns that space. Treat price, reviews and an explanation of the suggestion as different tools. Evaluate the complete shopping visit, not just the number of people who tap the card.
🔎 What research says
Test a useful cue against a fuller card, then follow the effect from product viewing through purchase. Research distinguishes card-information experiments from observed rating differences and survey evidence about trust. It supports careful local comparisons, not a universal rule to hide price, display fewer facts or maximize clicks.
🗺️ In this guide
1. Test one useful cue against a fuller card
📈 Recommendation
For a crowded recommendation row, test a simpler card against the current version while keeping the suggested products unchanged. Keep full price and product information easy to reach. A card should help someone choose the next step; it need not become a miniature product page.
For a homeware seller, compare a card led by price with the current price-and-rating card. Keep products, positions, promotions and eligibility consistent. This is a proposed store test, not a universal winning layout.
Judge the whole visit: paid orders, order value, product exploration and time spent looking. Inspect recommendation clicks and ordinary search separately. More clicking can also mean more work.
For a rollout decision, check margin, returns and customer understanding alongside orders. A busier journey is not automatically a better one.
🎓 Findings
Fang, Kim and Chintagunta (2026) randomly assigned 47,996 customers at a U.S. retailer to cards showing price, reviews, both or neither. Across 71,807 sessions, single-cue designs had higher modeled total order value than the dual-cue design. This covered all ordered products, not just recommendations.
For price alone versus price plus reviews, Table 2’s adjusted model implies 12.5% lower odds of ending without a purchase. We calculated this from the reported coefficient: exp(−0.134) = 0.875. Odds compare nonpurchase sessions with purchase sessions; this is not a 12.5-percentage-point conversion gain. An approximate interval from the reported standard error corresponds to about 2%–22% lower odds.
Illustration, not an observed study rate: starting from 50 purchases in 1,000 sessions, those odds would imply about 57 purchases, not 175. The hypothetical starting rate determines the absolute difference.
🧠 Why it works
A concise cue may help a shopper decide whether to inspect an alternative. Our operational interpretation is to remove competing messages before removing access to essential information. Test whether shoppers make better choices, rather than engineering extra clicks.
✋ Limitations
The algorithm underperformed, and apps were excluded. The no-cue group searched more but bought less. This does not justify hiding prices, adding friction or expecting the same effect with a better algorithm. The odds calculation is model-based, not a raw group-rate difference.
2. Follow rating cues beyond the product click
📈 Recommendation
When assessing a rating cue, measure whether it attracts attention and whether viewers subsequently buy. Keep the rating and its review count truthful. A rating that helps someone open a product page has not yet demonstrated a paid-order benefit.
For a skincare seller, compare a rating-led card with a price-led card using the same eligible products. Preserve access to review details and negative reviews. This is our proposed cue test; the study below tested recommendation availability.
Report product-view probability, purchase probability after a view and purchase probability across everyone assigned to each version. State each denominator. Do not add percentage changes across those stages.
Inspect different rating ranges as a diagnostic, then test the proposed design change directly. Do not decide that lower-rated products deserve promotion because their additional recommendation benefit might be larger.
🎓 Findings
Lee and Hosanagar (2021) randomly showed or hid a recommendation panel for 184,375 users. Ratings themselves were not randomized. Higher ratings were associated with a larger recommendation effect on viewing, but a smaller additional effect on buying after viewing.
The 2018 author manuscript, Table 5, puts the latter difference at 0.0705 percentage points less additional purchase probability per one-star higher rating, in its adjusted model of viewed user–product records. One such record is one person viewing a particular product. That is about 0.7 fewer additional purchases per 1,000 comparable records, as arithmetic on the model coefficient. It does not mean higher ratings reduced purchases overall.
The reported standard error gives an approximate interval of 0.028–0.113 percentage points less additional benefit. The final publisher abstract confirms the direction; equivalence of the older manuscript’s numeric table with the final article remains unverified.
🧠 Why it works
Opening a product page and deciding to buy are different decisions. A visible rating may attract attention, while detailed fit, ingredients or delivery information determines the next step. This is an interpretation to investigate, not proof that a particular badge caused the stage difference.
✋ Limitations
This was a recommendation-panel experiment, not a test of changing or hiding review cues. Product ratings were observed. The older manuscript also found no clear search-versus-experience product difference in the extra purchase effect after viewing. Do not turn category labels into a universal card rule.
3. Offer a truthful explanation without crowding the card
📈 Recommendation
Test a short optional explanation when shoppers cannot tell why an item was suggested. State the real basis, such as related purchases or compatibility, and let shoppers reach more detail. Do not replace a useful product cue with a vague claim that the algorithm knows best.
For a lighting seller, an optional explanation could say that a bulb fits the lamp being viewed, if catalogue data genuinely establishes compatibility. A purchase-history recommendation needs a different, equally truthful explanation.
Ask whether shoppers understand the reason and trust the suggestion; separately record exploration, purchases and unsuitable-item returns. Those are proposed evaluation measures, not outcomes all measured by this study.
For an executive, require evidence that the explanation is accurate and useful before adopting it. A persuasive explanation that conceals the real basis is not an improvement in decision quality.
🎓 Findings
Nilashi, Jannach and colleagues (2016) studied 150 postgraduate students using Amazon and Lazada. Participants browsed both sites, added items to baskets and answered questionnaires; they did not make required paid purchases.
In Tables 3–4, perceived transparency’s modeled relationship with trust was 0.401 for Amazon and 0.445 for Lazada, with statistically clear positive associations. These are standardized relationship strengths: they compare variation on different survey scales using a common unit. They do not mean 40% or 45% more trust, sales or customers. Recommendation quality and general website quality also mattered.
🧠 Why it works
A shopper can see a relevant-looking product and still wonder why it appeared. A truthful explanation gives them something they can check. Keep that explanation optional so it does not compete with the immediate product decision; this layout is our proposal.
✋ Limitations
This survey model cannot establish that adding an explanation caused trust or paid orders to rise. The same participants evaluated both sites, so these are not independent replications. The sites differed in many ways; their coefficients cannot establish that one explanation design was better.
💡 The takeaway
Give the card a clear job, and judge whether it helps shoppers complete a suitable purchase.
Start with one placement and unchanged suggested products. Compare the cue designs through exploration, orders and customer understanding; provide truthful optional explanations where useful. Roll out a design because the whole decision improves, with economics and unsuitable purchases checked separately.
📚 Sources and study types
Fang, X., Kim, S., & Chintagunta, P. K. (2026; online 2025). Too Many or Too Few? Information Cues in Recommender Systems and Consequences for Search and Purchase Behavior. Journal of Marketing, 90(1), 9–28. Randomized retailer card-cue experiment; adjusted session outcomes. The final article incorporates the October 2025 correction to the Sales standard deviation.
Lee, D., & Hosanagar, K. (2021; online 2020). How Do Product Attributes and Reviews Moderate the Impact of Recommender Systems Through Purchase Stages?. Management Science, 67(1), 524–546. Randomized recommendation availability, with observed ratings as moderators. Numeric moderation estimate checked in the institution-hosted 2018 author manuscript; final-table equivalence unresolved.
Nilashi, M., Jannach, D., Ibrahim, O. bin, Esfahani, M. D., & Ahmadi, H. (2016). Recommendation quality, transparency, and website quality for trust-building in recommendation agents. Electronic Commerce Research and Applications, 19, 70–84. Laboratory browsing and questionnaires on two real websites, analyzed through a survey relationship model; no required paid purchases. Author-hosted published PDF.
