Persona Generation

BETA. Persona Generation is a beta module. Features and behaviour may still change.

What it is & the business problem it solves

Persona Generation is a beta module of Inspire (CM's AI personalization product). It uses a large language model (LLM, an AI text model) to find a small set of named shopper personas in your customers' shopping behaviour. Examples (illustrative) are "Value-seeking family stocker" or "Premium skincare explorer". It then assigns each eligible customer to one of those personas.

Category segments ("buys shoes") tell you what a customer buys. Personas describe how a customer shops: which mix of categories, how broad or focused, how price-driven, how often. Marketers get a short, readable segmentation. They can use it to set tone of voice, choose creative and plan campaigns, without having to build rule-based segments by hand.

Personas complement category and RFM segmentation. They are not a prediction of what a customer will do next.

How it works

The module has two steps, and each step runs on its own schedule.

1. Persona discovery (runs rarely).

   • The LLM looks at the column names and the first 10 rows of your product catalogue (product feed). From these it works out which columns hold the product name, category and brand, and which language the catalogue uses.

   • Inspire builds a short behaviour profile for a sample of customers (by default up to 10,000 customers who interacted with at least 3 different products).

   • Over repeated rounds (40 by default) the LLM proposes groups of similar shoppers. Inspire then merges them into a final set of 5 to 20 named personas. Each persona has a name, title, description, main categories, distinguishing traits and typical behaviour.

   • Discovery is not scheduled by default. It is started manually when the module is set up, and again only when a fresh set of personas is wanted.

2. Persona assignment (recurring, daily is recommended).

   • A customer is eligible when they made at least 3 purchases in the last 6 months (defaults).

   • The first assignment run, and the first run after a new discovery, places every eligible customer. Later runs only re-place eligible customers who were active since the previous run.

   • The LLM assigns customers in small batches. When it is not confident enough about a customer, that customer is marked "Unassigned" instead of being forced into a persona.

   • Customers below the purchase threshold get no persona at all.

Example use cases

What you can configure

You can configure the settings below in the cockpit (the front-end tooling where you adjust the business rules behind your predictions). Settings marked Expert need expert rights. CM staff and customer admins can hold expert rights; request them via your CM contact person.

Persona Generation runs as a single module instance (one configured copy of the module) per workspace. All settings are setup settings on that instance. Schedules are cron expressions (time expressions that decide when a module runs) in Amsterdam time. "Not scheduled" means the step never runs automatically.

This module has no on/off switches: you control it through its two schedules and the settings above.

Prerequisites & required CM products

Packaging tier

Persona Generation is a beta module in the Pro tier: it is intended for the MMC Pro and Enterprise packages. The Basic and Advanced packages do not include Inspire.

Data requirements

• Events: purchase events (needed for eligibility) and any other standard interaction events (views, add-to-carts). Discovery uses every interaction type present.

• Catalogue: a product name and a category-like column. Inspire detects these automatically, so no column mapping is needed.

• Coded rules:

  • Discovery only considers customers with at least 3 different products.

  • Assignment only considers customers with at least Minimum Purchases purchases within Lookback Period (default 3 in 6 months).

• Technical minimum requirements: a customer is only considered with at least 3 different products in their history (discovery and assignment) and at least 3 purchases in the last 6 months (assignment, the default; see the settings above). The module defines no minimum number of customers overall. Personas only add value once enough repeat buyers exist to show distinct shopping patterns.

• Cold start: new or one-time buyers get no persona until they meet the purchase threshold.

How to check it works

• The module has no cockpit statistics and no event viewer yet.

• To check the results, ask your CM contact to share:

  • the persona definitions (names, titles and descriptions of the discovered personas)

  • an assignment summary: number of customers assigned, the share per persona, the share "Unassigned", and the date of the last assignment

• A high share of "Unassigned" or very uneven persona sizes can mean there is too little behaviour per customer, or that the persona set should be rediscovered.

Limitations & common pitfalls

• Beta module: features and behaviour may still change.

• Rediscovery breaks continuity: a new discovery produces a new persona set and reassigns every customer. Campaigns or content tied to old persona names must be updated. Keep the discovery schedule empty unless you deliberately want new personas.

• Coverage: only repeat buyers get a persona. One-time buyers, new customers and customers the LLM cannot place confidently ("Unassigned") need a fallback message.

• One persona per customer: a customer belongs to exactly one persona (or none). Mixed shoppers are not split across personas.

• Sampling: if Max Users for Assignment is above 0, customers not drawn in the sample are not assigned.

• Run time and cost grow with audience size. Very large customer bases take longer per run.

• Sensitive assortments: although no identifiers are sent and the prompts forbid sensitive inferences, product names per pseudonymous customer are part of what the model sees.