A DTC skincare brand we work with had a problem that looked like a recommendation quality problem but was actually a device-context problem. Their personalization was producing relevance scores that performed well on desktop — category affinity matching, return-visit signals, price tier alignment all working as expected. But mobile conversion rates for personalized sessions were barely moving, while desktop showed a clear lift.
When we dug into the session data, the pattern was clear: on mobile, shoppers were bouncing at the grid level without engaging with tiles that ranked high on the relevance model. The first two above-fold tiles weren't the problem — those were performing fine. The issue was tiles 3 through 6, which required scrolling on mobile but appeared immediately in the desktop viewport. The relevance model was ranking those tiles well for preference matching, but it wasn't accounting for the fact that on a two-column mobile grid, tiles 3 and 4 are below the fold and require deliberate scroll intent to reach. Tiles that were functionally invisible for mobile first-time visitors were being scored the same as tiles that were immediately visible.
This is the core mobile product grid problem: a ranking model that doesn't account for device-specific viewport constraints is a model that's optimizing for a context that doesn't match what mobile shoppers actually see.
The Above-Fold Math Is Completely Different
On a standard desktop product grid at 1280px width with a three- or four-column layout, a shopper sees 8 to 12 product tiles before the fold, depending on tile height and header size. That's enough tiles that ranking signals can distribute relevance across a reasonably large initial exposure set.
On mobile at 390px with a two-column grid, the same shopper sees 2 to 4 tiles above the fold before any scroll. In portrait mode with a standard header height, the number is often just 2. The stakes of tile positions 1 and 2 are dramatically higher than on desktop, because for a meaningful fraction of mobile sessions — especially first-time or low-engagement visits — those are the only tiles that get seen.
This changes what "good ranking" means on mobile. On desktop, a relevance model that places the best-matched product at position 1 and the second-best at position 5 is doing reasonable work. On mobile, placing a high-relevance product at position 5 in a two-column grid is effectively placing it in a graveyard for the significant share of sessions that bounce within the first scroll depth. The model needs to weight toward immediate visibility, not just relevance score.
Thumb Zone and Tap Target Positioning
There's a secondary constraint that desktop personalization doesn't encounter: thumb-zone physics. The ergonomic thumb zone on a phone held in one hand covers roughly the bottom 60% of the screen — the top corners are hard to reach comfortably without adjusting grip. On a two-column grid, this means that tile position 2 (top right) is actually the least ergonomically accessible of the first four tiles for right-handed single-thumb use, while tile positions 3 and 4 (second row, both columns) fall squarely in the comfortable zone.
We're not saying thumb-zone ergonomics should drive tile ranking — the interaction is subtler than that. But it does mean that for a shopper who is browsing one-handed, the friction of reaching a tile at the top-right corner is measurably higher than reaching a tile in the middle of the screen. When your personalization model places a high-relevance product in the top-right position, there is a small but real ergonomic tax on that placement that doesn't exist on desktop.
The more actionable design principle is ensuring that whatever product tiles land in the ergonomically awkward positions (top-right corner, specifically) are large enough and visually distinct enough to attract a tap despite the reach cost. Tile thumbnail quality matters more at high-cost positions. A high-relevance product with a mediocre image thumbnail will underperform in the top-right position relative to a slightly lower-relevance product with a compelling thumbnail.
Mobile Bounce Decisions Are Faster
Desktop browsing and mobile browsing have meaningfully different decision timelines. On desktop, a shopper who opens a product grid has typically committed to a moderate-effort browse session — they're at a keyboard, they can open multiple tabs, search is accessible without switching contexts. The expected dwell time before a bounce or engagement decision is longer.
Mobile browse sessions at the product grid level are faster to terminate. A shopper in a mobile context — on a commute, waiting in a queue, browsing opportunistically — makes a bounce decision within the first 3 to 8 seconds of viewing a product grid. If the above-fold tiles don't immediately signal relevance, the session ends. There isn't the same patience for scroll-and-discover that desktop browsing often allows.
For personalization ranking logic, this means the first-impression signal weight of positions 1 and 2 needs to be elevated for mobile contexts. Not in a way that overrides relevance entirely — a model that puts the highest-thumbnail-quality products in positions 1 and 2 regardless of behavioral relevance is not doing personalization. But within a set of products that score above a relevance threshold, the sort order for mobile should weight toward what looks immediately compelling in a fast-scan, not just what has the highest composite relevance score.
Session-Local Signal Accumulates Differently on Mobile
One operational challenge specific to mobile personalization is that session-local signal accumulates more slowly when browsing patterns are shorter and shallower. On desktop, a shopper who clicks through 6 to 8 products in a session gives the system enough behavioral data to build a confident session-local category affinity profile — you know they're interested in outerwear, not kitchen accessories, and you can act on that with high confidence.
On mobile, session depth is often 3 to 4 product views before purchase or bounce. That's a thinner signal. The personalization model has less behavioral data to work with when re-ranking products mid-session, which means it needs to be more conservative about how aggressively it shifts the grid based on sparse within-session data.
The practical response to this constraint is to rely more heavily on cross-session signals (prior visit behavioral history for identified shoppers) and acquisition context (landing page, campaign, referring source) for mobile sessions, rather than trying to infer a complete preference profile from 2 to 3 product views within the current session. For identified shoppers, cross-session behavioral history is typically available and provides a much stronger prior than anything a short mobile session can build in real time.
How Revlance Handles Device-Aware Ranking
In Revlance, device type is passed as a context signal in the recommendation request. The ranking model uses device type to adjust two parameters: the above-fold weight (how heavily it boosts products that fall in visible grid positions for the given device's viewport configuration) and the signal confidence threshold for real-time grid re-ranking (on mobile, the model requires more behavioral signal before it shifts rankings mid-session, because mobile sessions provide less data).
The above-fold weight adjustment is the more impactful of the two. On mobile contexts, tiles in positions 1 and 2 get a placement multiplier applied to their recommendation scores, which means a product needs to score noticeably higher on composite relevance to displace a candidate product from position 1 or 2 than it would on desktop. The goal is to ensure the most immediately compelling products are what mobile shoppers see first, without sacrificing the relevance quality that makes the personalization system worth having in the first place.
We're not saying mobile personalization requires a completely separate model architecture — it doesn't. The same underlying behavioral signals and ranking logic apply across devices. What changes is the position-weighting configuration and the signal accumulation thresholds, calibrated to the reality of what a 6-inch screen shows and how long mobile shoppers give you before they decide whether to scroll or leave.