How AI Shopping Assistants Are Changing Your Store’s Shelf Space – and What You Can Do About It

AI shopping assistants are no longer a futuristic idea – they’re the first thing many South African shoppers see when they ask a chatbot or voice‑assistant for product recommendations. That extra "middle‑man" between your catalogue and the customer can boost sales, but it can also steal revenue if you don’t plan for it.

Why AI Mediation Matters for Your Store

When a generative AI (like the ones from OpenAI, Anthropic or Google) is asked to recommend products, it creates a ranking of the items it thinks are most relevant. The shopper only sees the top K products – the AI’s exposure set – and then decides what to buy from that short list. This extra step is called AI mediation.

Research that queried leading models across three product categories and three shopper personas found that:

Ignoring the AI’s ranking when you choose which items to list can cut revenue by 2.5%‑11.4% on average, with worst‑case losses over 57% [1].
A simple deterministic plug‑in that orders your catalogue by the AI‑estimated utility loses no more than 0.53% compared to the optimal stochastic model [1].
The AI tends to shrink the assortment it shows – often fewer than its own display limit K – meaning you may be pushing fewer items than you thought [1].

In plain English: if you keep picking products the old way (e.g., “best‑selling items” or “full inventory”) while the AI is silently filtering the list, you’re likely leaving money on the table.

The Two Ways AI Can Rank Your Products

Deterministic ranking – the AI always puts the same items in the same order for a given customer segment. Think of it as a fixed “top‑10” list.
Stochastic ranking (Plackett–Luce model) – the AI adds a touch of randomness, so the order can vary slightly each time.

Both models are not captured by traditional assortment‑optimization tools that retailers use today, which assume the retailer directly controls the shopper’s consideration set.

What the Numbers Really Mean for You

Scenario	Typical revenue loss vs. optimal	What it looks like in practice
Ignore AI mediation (direct‑MNL)	2.5%‑11.4% average, up to 57% worst case	You stock 1.5‑4 more items than the AI would show, but many never get seen.
Deterministic plug‑in (fit AI utilities)	≤0.53% loss	You reorder your catalogue to match the AI’s ranking; the AI still shows its top K, but you lose almost nothing.
Full stochastic optimisation	Baseline optimum	Hard to compute in real time; requires specialised algorithms.

The study also showed that when the AI’s ranking aligns closely with real customer preferences (high alignment parameter ρ), retailers are more likely to let the AI filter heavily – up to two‑thirds of the time – which can be good if you’ve accounted for it, but disastrous if you haven’t [1].

How to Adapt Your Assortment Strategy Today

1. Find Out How Many Products the AI Will Show (the K value)

Ask the AI provider or test it yourself. Run a few queries like “show me the top shoes for a 30‑year‑old male” and count the items returned. Most platforms cap the list at 5‑10 items.

2. Estimate the AI’s Utility Scores for Your Products

You don’t need a data‑science PhD – a simple approach works: 1. Create a spreadsheet with your SKU list, price, and a brief description. 2. Prompt the AI: “Rank these 20 products for a budget‑conscious shopper looking for summer dresses. Give a score from 0‑100 for each.” 3. Record the scores – these are your deterministic utilities. 4. Repeat for each major customer segment (e.g., price‑sensitive, brand‑loyal, trend‑seeker). The paper used three personas; you can start with two.

3. Apply the Deterministic Plug‑in

Once you have segment‑specific scores: * Sort each segment’s catalogue by the AI score. * Trim the list to the AI’s display limit K (or a little less if you want a safety margin). * Publish this trimmed, ordered list as the catalogue you feed into the AI (or directly to your website if you bypass the AI).

Because the plug‑in’s revenue loss is under 0.5%, you’re essentially matching the optimal stochastic solution without heavy maths.

4. Use an Approximation Scheme if You Have Many Segments

If you serve more than three distinct shopper types, the exact problem becomes NP‑hard (computationally tough). The authors provide a Fully Polynomial‑Time Approximation Scheme (FPTAS) that runs quickly when the number of segments m is fixed. In practice, you can: * Group similar customers into broader segments (e.g., “budget” vs. “premium”). * Run the FPTAS using an open‑source library or a simple Excel macro that iterates over possible assortments and picks the one with the highest estimated revenue.

5. Monitor and Re‑Calibrate

AI models evolve, and so do shopper tastes. Set a monthly check: * Re‑run the ranking prompts for a sample of SKUs. * Compare the new scores to the old ones – a shift of more than 10% suggests you should update your assortment. * Track sales of items that fall out of the AI’s top K; if they still sell well offline, consider a dual‑channel approach (online via AI, offline via traditional stocking).

Quick Wins for South African Small Businesses

Action	Time needed	Cost	Expected impact
Identify AI’s K (display limit)	30 min	Free	Prevents over‑stocking of unseen items
Generate deterministic utilities with a free AI trial	1‑2 hrs	Free‑to‑low (e.g., OpenAI free tier)	Cuts potential revenue loss from 2‑11% to <0.5%
Trim catalogue to top K per segment	1‑2 hrs	Free	Aligns stock with what shoppers actually see
Set up a monthly re‑ranking routine	1 hr/month	Free‑low	Keeps you in step with AI updates

What to Do Next

Test the AI ranking on a small product range (10‑20 items) and note the top K.
Create a simple spreadsheet of your SKUs and ask the AI to score them for each key customer segment.
Reorder and trim your online catalogue based on those scores.
Schedule a monthly review to refresh scores and adjust segments.
Consider a paid AI plan only if you need higher query limits or faster responses; the free tier is enough for a pilot.

By treating the AI assistant as a partner, not a black box, you can keep most of your revenue while still benefitting from the convenience and reach of generative AI. The deterministic plug‑in is the low‑tech, high‑impact tool that fits a small South African retailer’s budget and skill set.

Sources: [1] AI‑Mediation in Retail Markets: Revenue Effects and Assortment Optimization (arXiv).

Sources

Researched and drafted with AI, then reviewed by the Vicinic team.