Boosting Basket Size with Proximity-Based Product Recommendations
AI-powered proximity recommendations lift basket sizes. Here's the mechanics of how they work, the ethical limits, and the ROI math.
E-commerce has had "customers who bought X also bought Y" recommendation strips since the early 2000s. They lift average basket size by 10β25% depending on category and placement. Physical retail has, until recently, had nothing comparable β because there was no interactive surface in the shopper's hand.
Indoor navigation PWAs change that. Once the shopper has a search bar and a map, the "recommendation strip below search results" is a natural placement. And the data behind what should appear there is substantially richer than online.
Here's the mechanics.
The signals available in a physical store
Online stores base recommendations on:
- Browsing history
- Purchase history
- Collaborative filtering (what similar shoppers bought)
- Content similarity (same category, same brand)
A physical store's PWA can use some of those (co-search within sessions, content similarity) but also layer in:
- Aisle proximity. Products in the same or adjacent aisles are literally easier for the shopper to add β zero extra walking cost.
- Route-adjacency. Products on the way between the shopper's current position and their intended destination.
- Impulse-stage. Products near the exit or checkout, where shoppers are in last-chance buying mode.
- Time-of-visit patterns. Different recommendation sets for a Monday morning versus a Saturday afternoon shopper.
The proximity signals are more powerful than the online equivalents because they align with actual walking cost. A shopper will add a complementary product that's five steps out of their way far more readily than one that's three aisles over.
The scoring approach
A simple blended scoring model that works well:
- Same aisle β 1.0
- Same shelf or adjacent shelf β 0.9
- Within 15% of canvas distance β 0.7
- Same category β 0.5
- Same brand β 0.4
- Co-searched by other sessions β 0.3β0.8 (scaled by frequency)
- Trending in last 7 days β 0.2 (as filler)
The highest-scoring candidates become the recommendation strip. Blend multiple reasons for the same product and add a small bonus per signal β so a product that's both "same aisle" and "co-searched" ranks above either alone.
Ethical limits
A few places where retailers should pull back:
1. Health-category manipulation. If a shopper searches for a specific medication at a pharmacy, don't push a higher-margin alternative into the recommendation strip. The trust cost is real and the regulatory risk is non-zero.
2. Price-anchoring games. Showing an expensive product next to a cheaper one you want the shopper to accept as "a deal" is manipulative. Don't do it.
3. Dark patterns around "limited stock." If you're not actually running low, don't fake urgency.
4. Children's products. When a shopper is clearly in the toys or children's section, keep recommendations family-friendly.
AI vs heuristic
The signals above are heuristics β they work from already-available data. More recently, LLMs have become good at generating "creative" pairings that no co-search data could catch.
Example: a shopper searches for a Sony soundbar. The heuristic recommends:
- A HDMI cable (same aisle, same category β reasonable)
- Another Sony soundbar (same brand β a substitute, actually bad)
- A Logitech mouse (same aisle but unrelated β noise)
An LLM given the same product list might suggest:
- A wall-mount bracket (not in the same aisle, but highly complementary)
- A sound insulation foam kit (adjacent-category thinking)
- A universal remote (behavioural: people who upgrade AV equipment often need better control)
The LLM's suggestions are more creative but less proximity-aware. The best approach is to use both: the heuristic for proximity, the LLM for creative pairings, and have the admin tick which ones to turn into permanent "manual pairs."
The ROI math
A typical retailer running recommendations sees:
- 8β15% of shoppers tap a recommendation card.
- 40β60% of taps convert into a route addition.
- 25β40% of route additions actually get added to the basket at checkout.
Compound: for every 100 search-driven PWA sessions, you get roughly 1β3 extra items in baskets. At a modest SGD 8 average item price, that's SGD 8β24 of additional revenue per 100 sessions.
For a store doing 300 PWA sessions/day, that's SGD 24β72/day β SGD 8,000β26,000/year β of recovered impulse revenue that would otherwise never happen because shoppers didn't walk past the right shelf.
What to measure
If you turn recommendations on, track:
- Tap-through rate (% of recommendation views that get tapped).
- Add-to-route rate (% of taps that queue the product).
- Attach rate (% of recommendations that end up in checkout β needs POS integration or inferred from basket).
- Decay by position in the strip (first card gets most taps, drops off; use this to validate your ranking).
Negative signals to watch for: if tap-through drops below 5% over a 30-day window, your recommendations are noise. Rethink the signal weights.
The pitfall to avoid
Over-recommending. If every search surfaces 8 recommendation cards, the shopper tunes out. Three to five, max. Quality beats quantity.
Second pitfall: showing the same product the shopper is already searching for. Dedupe aggressively against the current result set.
Try it
StoreNavigator's recommendation engine runs the heuristic model by default. Optional LLM integration β Claude, OpenAI, Gemini, OpenRouter, or local Ollama β adds creative suggestions. Admins pick which ones to promote into manual pairs.
Try the live demo β β search for anything, scroll down to the "You might also want" strip.