Should recommendations differ between mobile and desktop shoppers?
Most recommendation engines are device-blind: the same algorithm serves the same ranked list to a phone and a laptop. But the two shoppers behave nothing alike. The mobile shopper browses in short bursts, decides fast, and buys with a thumb. The desktop shopper opens tabs, compares deliberately, and researches. Serving both the identical recommendation experience ignores the context that shapes every other part of the decision. Device-aware recommendations are not a separate engine; they are the same engine with different presentation and different ranking weights.
How the sessions actually differ
Mobile sessions are shorter, more frequent, and more interrupted. The shopper might browse for ninety seconds on the train, come back at lunch, and buy at night. Each session is a fragment, and the recommendation rail has to work in fragments: immediately legible, low-commitment, easy to resume. Desktop sessions are longer and more deliberate, with tabbed comparison and deeper product page engagement.
The purchase patterns differ too. Mobile skews toward replenishment, gifting, and impulse categories; desktop skews toward high-consideration purchases with bigger carts. A recommendation strategy that ignores this will systematically misfire: showing deep comparison content to a thumb-scroller, or impulse add-ons to a researcher mid-comparison.
What should change on mobile
On mobile, the rail should optimize for glanceability and low friction. Fewer items visible at once (the carousel shows two or three, not six), larger touch targets, and recommendations that require no comparison to evaluate: 'pairs with what you are viewing', 'back in stock in your size', 'reorder your usual'. The ranking should weight recency and simplicity over depth.
Continuity matters more on mobile because sessions fragment. 'Pick up where you left off' is the highest-value mobile recommendation pattern: the products from the interrupted session, restored instantly. The mobile rail is a bookmark as much as a recommender.
What should change on desktop
On desktop, the rail can afford depth. Comparison-friendly layouts (side-by-side specs, wider carousels), recommendations that invite research ('compare with similar', 'complete the setup'), and bundles that assume a bigger cart. The ranking can weight margin and cross-category discovery more heavily because the shopper has the attention to evaluate them.
Desktop is also where social proof carries the most weight in recommendations. Review counts, ratings distributions, and 'bought together' data get read carefully on a big screen and glanced past on a phone. Weight the proof signals higher in desktop ranking.
One engine, two presentations
You do not need two recommendation engines. You need one candidate generation layer and device-aware ranking and presentation layers. The candidates (what this shopper might want) are largely device-independent. The ordering and the display (what they see first, how many, in what layout) are where device context applies.
Implement it as weights, not forks. A device parameter in the ranking function that adjusts the importance of recency, margin, proof, and novelty is maintainable; a separate mobile engine is a second system to tune, monitor, and debug. Start with presentation differences (layout, count, modules), which are cheap, then tune ranking weights once the presentation is right.
Measuring device-aware recommendations
Measure per device, always. A global lift can hide mobile gains offset by desktop losses, or vice versa. Track rail click-through and rail-attributed revenue separately for mobile and desktop, and watch cross-device journeys: the shopper who browses on mobile and buys on desktop should see continuity, not a reset.
The key test is whether device-aware ranking beats the device-blind baseline on each device independently. If mobile improves and desktop is flat, ship it. If either device regresses, your weights are wrong, not the concept. And keep a device-blind holdout running; device behavior drifts as your traffic mix changes, and the holdout tells you when to retune.
Reviewed
Published Oct 3, 2026.