AI in retail personalisation that customers don't find creepy
Recommendations, search and merchandising — the models that lift revenue, and the line between helpful and unsettling.
Personalisation has a reputation problem because a lot of it is done badly — retargeting a product someone already bought, or demonstrating knowledge the customer never knowingly shared.
Order of return in retail AI
- 1
Fix site search
Where high-intent customers tell you exactly what they want. Usually poor, usually the biggest win.
- 2
Session-based recommendations
What they're doing now beats what they did in March — and sidesteps the privacy discomfort.
- 3
Merchandising models
Demand forecasting and markdown timing move more money than the customer-facing ones.
- 4
Measure incrementally
Hold out a control group, or you're measuring purchases that would have happened anyway.
Fix search before building recommendations
Site search is where high-intent customers tell you exactly what they want, and on most retail sites it is poor. Semantic search that handles synonyms, misspellings and descriptive queries usually returns more than a recommendation engine, for less work.
Recommend for the session, not the profile
What someone is doing right now beats what they did three months ago. Session-based models also sidestep much of the privacy discomfort, because they are visibly responding to the current visit rather than to a dossier.
The creepiness line
Basis
- Helpful
- What they did on your site
- Unsettling
- What you inferred about them elsewhere
Timeframe
- Helpful
- This session
- Unsettling
- A dossier going back months
Explanation
- Helpful
- "Because you viewed X"
- Unsettling
- No explanation offered
Test
- Helpful
- They would expect you to know it
- Unsettling
- They would be surprised you know it
| Helpful | Unsettling | |
|---|---|---|
| Basis | What they did on your site | What you inferred about them elsewhere |
| Timeframe | This session | A dossier going back months |
| Explanation | "Because you viewed X" | No explanation offered |
| Test | They would expect you to know it | They would be surprised you know it |
The creepiness line
The rule of thumb that holds up: personalise on what the customer did on your site, be careful with what you inferred, and never surface something they would be surprised you knew. Explaining recommendations — 'because you looked at X' — converts unease into usefulness.
Merchandising is where the margin is
Customer-facing models get the attention; the operational ones move more money — demand forecasting, markdown timing, and detecting emerging trends early enough to buy for them.
Measure incrementally
A recommender that suggests what people would have bought anyway shows excellent click-through and adds nothing. Hold out a control group and measure incremental revenue, or you are measuring your own shadow.