So you have customer data, and you’d like to start segmenting it to get serious with your acquisition and retention efforts. That means going beyond what out of the box analytics tools spit out, and doing data crunching of your own. Where do you start? There’s typical data-mining methodology (ie CRISP-DM) which you can follow but that’s not what this post is about. The focus here is on some of the practicalities of implementing a customer segmentation project that I wish somebody taught me early on.
Why Segment
Let’s start off with the obvious – why segment? Segmentation is done to improve results with your marketing, retention, customer success, and other programs, and it does this through a variety of ways, some that are often overlooked:
- Increasing relevance. If marketing is about delivering the right message at the right time to the right person, then segmentation is what will get you there. Segmentation allows you to personalize a user’s experience, which works because of emotional, trust, and social proof reasons. Read more here. Remember: average message fits no one.
- Uncovering new opportunities. Segmentation might lead to the discovery of groups of customers with needs that you didn’t know about before. Perhaps you’ll find that a large group of commuters are using your app while on the train to work, so you can acquire more users by launching some transit ads. Or maybe you’ll find that your most active users are busy parents who value time more than money, so you decide to create product features that cater to convenience over savings.
- Learning. Experiments are more likely to be successful when you segment first. Read more on how segmentation is better for setting up proper experiments than A/A/B testing, and see how experiment segmentation is done at Pinterest.
- Strategy & Communication. Having a common set of customer segments for your entire team to work against can really help to streamline decision making. Perhaps you want to create a white glove treatment for high value customers, create a mobile app for commuters, or personalize emails for moms. If everyone in the company is looking at common user segments, then it becomes easier to make the most out of these opportunities – marketing will know what personas they’re speaking to, analysts can look at the funnel numbers of each segment differently, and the growth team can figure out which channels to use to acquire more of the required personas.
What can segmentation do for you? A lot. Some examples: SwayChic increased email open rates by 40% and tripled revenue for their email campaigns using segmentation; Neustar claims their segmentation solution lifts web conversions by 27%.
Great, you say, but why do my own segmentation?
Because out-of-the-box analytics tools only get you so far! Analytics tools like Google Analytics, Mixpanel, or Amplitude have great segmentation capabilities, but as your company scales there’ll be a point at which the out-of-the-box implementations just won’t cut it anymore. You’ll probably have a bunch of data points coming from different sources that you want integrated in the segmentation, or you’ll want to be able to tailor the segmentation algorithm to your specific needs, like changing the definition of some retention cohorts to deal with a class of fraudulent users.
The Pitfalls
Once you’ve determined that you want to get started doing your own segmentation, here are some common pitfalls to avoid.
1. No goal – this may sound obvious but know what you’re segmenting for! Segmenting your email subscriber base to create more relevant content is different from segmenting your target market into geographic buckets for sales, and segmenting anonymous users on your site by behavior is different again, etc. Depending on your use case, the data available and the variables you want will be different. A general segmentation conducted without a focus might be great for general discussions, but won’t get you very far.
2. Not planning on maintenance – again this may be obvious but it happens more often then you think. You can’t create a one time segmentation and hope to use it over and over till the end of time. Your customers will change, along with their behaviors and interests. Make sure you run your segmentation often enough to keep the info fresh – the last thing you want to do is send a retention offer to somebody who hasn’t even signed up, or a call-to-action to join your newsletter to people who have already joined. Have a look at the data and determine the frequency that you want to refresh your segmentation, be it daily, monthly, or quarterly.
3. Too many or too little segments – extensive segmentations like the one shown here from Experian looks super impressive on office walls. Before you jump on creating your own though, make sure you can make use of something this extensive.
Do you have the man-power to be sending out this many variations of direct mail, emails, promos, or targeted content and ads placements? If you want to experiment you’ll be sending multiple treatments to each group, which quickly blows up the amount of variations. A wider number of segments are great when you have automated systems handling your operations, but if not, 7-8 segments is a good maximum limit to stick to. The success of the segment relies on your team buying in, which means that ideally you have 5-6 segments that everybody can remember and recall whenever you’re deciding on copy, messaging, and all the other good stuff segmented personas are good for. And lastly, make sure every segment has a good amount of customers in it. If a segment only has 1-2% of your customers, they had better be high value or high potential so that the extra effort you’re putting in to target that small group is worthwhile. And no, individual outliers do not need their own segment!
Too Many Segments
Not Enough Segments
4. Horrible names – names really matter! I would argue it’s one of the most crucial parts of the segmentation exercise. If you’re segmenting based on customer revenue, you might brainstorm and come up with a few possibilities for names. Your lowest spending segment could be cheap, or they could just be new to your product, or you can see them as low hanging fruit. Do you see how the name will drastic shape how your team will treat this segment? If you do things right, your customers should start to see and feel these names in your correspondence – your most high value loyal customers might get white glove treatment, for example. But you don’t want new customers feeling unimportant and labeled as “low value” or “newbies”. So choose wisely.
5. Poor visualization – whatever you do, DO NOT present your Python, R, or Weka results as your final work product! Take the time to really paint the picture of what the people in your segments are like. Find or create some photos of people who personify the segment, and add some color about what they’re really like beyond just the numbers of median income, average education, etc. If you’re looking for inspiration, check out Environics’ Prizm5 segmentation of the Canadian population.
6. Boxed in by what data you have – the power of segmentation comes when you’re able to paint a complete picture, and more often than not, that requires pulling multiple data sources together. The sources though, are usually at different levels of granularity – maybe you have income data for certain zip codes, transaction data at the individual customer level, and mobile app usage for anonymous customers only. At this point, I see many people giving up because there isn’t a precise key that matches all the sources. Don’t let that be you! Often there are useful generalizations that you can extract, even if the base data and mappings aren’t 100% precise. Maybe only 90% of your customers have a zip code logged in your dataset, and using that you see that those in higher incomes areas are buying lots of kids toys at your store. Don’t let the fact that you’re missing 10% of the zip codes preclude you from exploring showing toy sales to these customers. And last but not least: work on getting more data! More data can make your model orders of magnitude better than algorithm optimization. So survey your users, capture more data with your apps, buy data, or use mechanical turks to collect data if you have to. The time spent there can save you months in modeling work where you end up spending months trying to find good proxies for missing variables.
And there you have it. Hope you enjoyed this list. I was hoping to do a larger guide to segmentation that digs into modeling details but haven’t found the time. If you think it’ll be of interest to you, let me know!
Sylvia Ng 