Personalisation at Scale: Why Most Programs Plateau at "Good Enough"
By Gabbi Stubbs
Most personalisation programs deliver a strong first year and then stall. Open rates lift, a few journeys improve, and the dashboard looks healthy, yet customers still receive messages that feel built for a type of person rather than for them.
The cause is usually the same: the program has hit the ceiling of rules-based segmentation.
Rules-based segmentation groups customers by shared traits or recent behaviour, then applies if-then logic: lapsed in 60 days gets the win-back offer, bought running shoes gets the apparel cross-sell. It is a sensible starting point. It is easy to explain, quick to build and simple to govern.
It also stops working in predictable ways.
The ceiling: rules-based segmentation
Segments are averages. Everyone in the group is treated as the average member, so individual intent is invisible.
Rules multiply. Every new product, channel or exception adds more logic, until no one can say which rule fired or why.
Nothing learns. A rule does what it was told until a person changes it.
Context is missing. Rules struggle with what a customer is doing right now.
Success is flattering. Results are compared with sending nothing, so the program looks effective while leaving value untouched.
Teams stay here because the next step is not another tool. It is a change in data, decision-making and measurement.
What it takes to reach 1:1 relevance
True 1:1 means choosing the best next action for an individual, at the moment of interaction, and learning from the result. That rests on five prerequisites.
A resolved customer identity. One consistent view of each person across channels, devices and systems.
Timely behavioural data. Event-level signals that arrive in near real time, not in weekly extracts.
Built-in consent and governance. Personalisation that customers trust depends on clear permissions and purpose limits.
A decision layer. Models or decisioning engines that rank options per person, replacing hand-written rules.
Causal measurement. Holdout groups that show what personalisation added beyond what customers would have done anyway.
One more requirement sits underneath: modular content and offers that are tagged well enough for a system to choose between them. Without a rich set of options, even the best model has nothing to choose from.
Two organisations that broke through
Netflix
Netflix treats personalisation as the product rather than a marketing layer. Its research team has argued publicly that the share of viewing that comes from recommended titles is a misleading measure, because recommendations target existing demand. Instead, it separates underlying preference from the incremental effect of recommendations, using variation in the system and randomised experiments. It has also spoken about consolidating many specialised recommendation models into a shared foundation model.
Why it works: decisions are made per member and per moment, the data is event-level, and measurement is causal. Netflix can tell whether personalisation adds value, not just whether it correlates with it.
Starbucks
Starbucks shows the same principles in a physical retail business. Its Rewards program ties an identity to purchases across app, drive-through and store. By fiscal Q3 2022 (April to June 2022), Rewards members accounted for 53% of U.S. company-operated revenue. Its Deep Brew AI platform runs on an enterprise data platform that unifies data from multiple sources. An earlier analytics profile reported that personalised email offers more than doubled response rates compared with previous segmented campaigns.
Why it works: first-party, consented identity is captured at the point of purchase, data is unified on one platform, and personalised offers were tested directly against segmentation. It is not flawless. A 2016 rewards redesign confused members and failed to lift spend per member as expected, which is a reminder that offer design matters as much as the engine.
Where to start
Audit your journeys to see where rules dominate and where individual signals are ignored.
Fix identity and consent first. Models cannot personalise what they cannot recognise.
Pilot one decision, such as next best offer, with a holdout group, and report incremental value to the board rather than raw engagement.
The ceiling is not a technology limit. It is the point where personalisation stops being a campaign activity and becomes a data and decision capability. The organisations that cross it invest in the foundations first.
Sources: Netflix research on measuring recommendations (CESifo/Kellogg working paper); Netflix Tech Blog repost, October 2026; Restaurant Dive on Starbucks Rewards, 2022; Drexel LeBow Analytics 50 profile of Starbucks; Towards Data Science on Deep Brew.
Want to talk strategy?
If this resonates, let's have a direct conversation about your business.