How many personas to create?
Let’s assume you want the personas to represent your customer base.
Let’s assume you have 1000 customers.
Your upper boundary would be 1000 personas, one for each customer. This would be 100% representative of your data.
Your lower boundary would be 1 persona, describing the most typical or average customer. Depending on how we measure representativeness, this could mean you represent from 0% (i.e., there’s no customer that comprises the average of each attribute) to some higher number, like 30%.
So, both extremes are typically excluded as problematic. Typically, people create less than 10 personas (almost all research papers we’ve reviewed create less than 10 personas). This is influenced by two reasons: (a) the idea that decision makers cannot cope with multiple personas due to cognitive limitations (there’s a popular memory heuristic, “seven plus minus two” postulating that people can hold 5-9 items in their mind simultaneously but not much more) and (b) the clustering algorithms often deployed for persona creation typically yield an “optimal” number of clusters or segments at k < 10. In fact, the data coverage continues increasing after this but the optimal cluster size assignment traditionally aims at finding a point where error or information gain stabilizes and the cluster size at that cut-off point becomes “chosen”. So, practical realities and traditions have led to to a “small” number of personas being preferred.
In our research focused on interactive persona systems, we want to challenge this thinking. Our core argument is that when we offer decision makers interactive tools (search, filter, generation, recommendation) for accessing personas, we can considerably increase the number of available personas. The decision makers can then make use of these features to form their own “consideration set” of personas based on a given decision-making situation — sometimes, they could want 1, sometimes 5, sometimes 50 personas (all numbers could have a reasoning, ranging from focusing on one to obtaining a broad overview of an international audience). So, we don’t want to artificially limit the access to personas “hidden” in the data by pre-selecting them with a clinical algorithmic segmentation.
This also means that we can conceptually distinguish “clustering for persona creation” from “clustering for statistical analysis/machine learning” — the latter aims at compressing information into fewer dimensions or clusters, whereas, for persona creation, this goal of compressing information should take place based on persona users’ needs, not the algorithm’s perspective of what information is to be preserved.
The current persona creation literature, according to our argument, applies clustering without consideration for the specific circumstances of persona creation. These circumstances are very different from the purpose for which clustering initially was developed for, especially when we add interactive persona systems into the mix.
The above reasoning also yields an interesting after-thought: is clustering even meaningful for persona creation? Or, can we better capture decision makers’ information needs in real time and then present them with personas matching those needs without needing to perform clustering as a middle step?