How to envisage a data strategy that aligns tightly with your growth objectives? In this series, we will breakdown the growth ecosystem into four stages of the user journey: acquisition => activation => engagement => retention. For each stage, we will recommend the key data science projects in order to accelerate growth and maximise return.
Gone are the days when marketing campaigns take months to research and execute, then months to gather ratings and sales figures, then months to analyse and reiterate. Now companies receive instant feedback on digital channels and are able to optimise on the fly. Both Facebook and Google use powerful algorithms to automate this process in real time, making sure the best ad copies are shown to the most relevant audience hence increasing return on your marketing spend.
Now you may ask: if the duopoly had it all figured out, do I need to do anything else? The answer is absolutely yes. If everyone uses the same powerful model, no one has a competitive advantage. The key is to help Facebook or any buying platform algorithms bring you the most desired customers. The secret is in your own data.
The most desired customers are those who return the highest lifetime values (LTV). What defines value for your business? Number of app downloads, number of product or subscription purchases, or total sales amount? Choosing the right metric is important. If you run a freemium model and choose download as your optimiser, you might end up with more users who will never convert to payers. If you run a dating app and choose purchase as your goal, you will most likely end up with way more men than women. The key is to look into your historical data and model the impact of a range of KPIs. If you are a completely new startup, you can make this decision heuristically and improve as you gather more data.
The second challenge is timing. You cannot wait for a user’s ‘lifetime’ to send data back to Facebook. The sooner Facebook receives these signals, the more effective your campaign will be. You can do this by predicting the lifetime value of a customer at the earliest interval based on, again, your historical data. To give a very simple example, if you know your subscribers stay on average 3.5 months and your subscription costs $10 per month, then as soon as someone subscribes, you can predict this person to be worth ~$35.
The prediction model can get a lot more sophisticated. For example, you might already know that iOS users have a higher propensity to convert. Therefore you can infer LTV at the point of download by Estimated LTV * Estimated Conversion Rate. In fact, you can use all the information you have on existing users, including personal settings and behavioural patterns, to build the model and predict for each new user as soon as you have enough information. Please note, it is ideal to limit the number of features going into your model in order to optimise performance. This is a topic for another day.
Keep in mind that there could be more than one type of desired customer for your business. We will revisit this in the audience selection section later.
The other reason you want to centralise budget optimisation within your own reporting is the issue of duplicated conversion tracking. All marketing channels are greedy. Everyone wants to claim credits for their own. If a prospecting user saw and clicked on a Facebook video ad, then a few days later Googled your product, clicked on a Google Ad and purchased, both Facebook and Google will claim this user as a conversion for itself. You want to be able to attribute this user fairly and optimise your budget across all channels.
What is a fair attribution model? By default, companies apply a last-click policy. This tends to favour channels closer to the conversion point. In the above example, Google will get all the credit. In reality, it is the Facebook video that made the first and probably most impact. We will discuss various attribution modelling in more detail in another post. Below is a quick summary.
What about non-digital channels? Are you able to measure it? The answer is yes, you can estimate the impact. We will discuss more about Causal Impact modelling in the future but below is the concept in a nutshell.
To estimate the impact of an event at a point in time for a single time series, you can start by predicting what would have happened without this event, aka the offline campaign. It is a relatively simple forecasting model based on historical data. Once you have the forecast, the difference between what would have happened and what actually happened is the estimated effect of this campaign.
If you are the lucky few with lots of media attention or amazing word-of-mouth from a very sticky core customer base, you may not rely as heavily on performance marketing to bring in new customers. In this case, analysing traffic sources is still important. It can help:
There are many ways to define your target audience. On one side of the spectrum, you could leave everything to external platform algorithms: select your market, set bidding rules, add your creatives, click run and let Facebook or Google optimise on their side. Another strategy is to create audience groups based on your existing user behaviour and tailor marketing materials to each group.
For example, you are a chat app and one of your analysts discovered that your most engaged users are of Gen Z and their favorite feature is sharing Spotify playlist. From this insight you might consider creating a campaign targeting the age group between 18 and 25. You could also create Lookalike Audiences on Facebook, or similar audiences on Google, based on your most engaged users. Facebook and Google will then find users that are most similar using their own algorithm and serve them your ad. With either approach, you should design your ad around the share playlist feature.
Above is a sample use case of a single cluster. It is always good practice to test various buying strategies and compare performance results. We will circle back to this topic in more detail when discussing the next stages of the user journey. Together with user segmentation and LTV prediction, it closes the growth cycle.
Continue reading our growth series with the next article on user acquisition.
By subscribing, you agree to receive a monthly newsletter from 173Tech.