How to build an experimentation, CRO or AB testing framework

Everyone approaches experimentation differently. But there’s one thing companies that are successful at experimentation all have in common: a strategic framework that drives experimentation.

In the last ten years we’ve worked with start-ups through to global brands like Facebook, the Guardian and Domino’s Pizza, and the biggest factor we’ve seen impact success is having this strategic framework to inform every experiment.

In this post, you’ll learn

• Why a framework is crucial if you want your experimentation to succeed
• How to set a meaningful goal for your experimentation program
• How to build a framework around your goal and create your strategy for achieving it

We’ll be sharing the experimentation framework that we use day in, day out with our clients to deliver successful experimentation projects. We’ll also share some blank templates of the framework at the end, so after reading this you’ll be able to have a go at completing your own straight away.

Using this framework will help you mature your own approach to experimentation, make a bigger impact, get more insight and have more success.

Having a framework:

• Establishes a consistent approach to experimentation across an entire organization, enabling more people to run more experiments and deliver value
• Allows you to spend more time on the strategy behind your experiments and less time on the “housekeeping” of trying to manage your experimentation program.
• Enables you to transition from testing tactically to testing strategically.

Let’s explore that last point in detail.

In tactical experimentation every experiment is an island – separate and unconnected to any others. Ideas generally take the form of solutions – “we should change this to be like that” and come from heuristics (aka guessing), best practice or from copying a competitor. There is very little guiding what experiments run where, when and why.

Strategic experimentation on the other hand is focused on achieving a defined goal and has clear strategy for achieving it. The goal is the starting point – a problem with potential solutions explored through the testing of defined hypotheses. All experiments are connected and experimentation is iterative. Every completed experiment generates more insight that prompts further experiments as you build towards achieving the goal.

If strategic experimentation doesn’t already sound better to you then we should also mention the typical benefits you’ll see as a result of maturing your approach in this way.

• You’ll increase your win rate – the % of experiments that are successful
• You’ll increase the impact of each successful experiment – on top of any conversion rate uplifts, experiments will generate more actionable insight
• You’ll never run out of ideas again – every conclusive experiment will spawn multiple new ideas

As we introduce our framework, you might be surprised by its simplicity. But all good frameworks are simple. There’s no secret sauce here. Just a logical, strategic approach to experimentation.

Just before we get into the detail of our framework a quick note on the role of data. Everything we do should be backed by data. User-research and analytics are crucial sources of insight used to build the layers in our framework. But the experiments we run using the framework are often the best source of data and insight we have. An effective framework should therefore minimise the time it takes to start experimenting. We cannot wait for perfect data to appear before we start, or try and get things right first time. The audiences, areas and levers that we’ll define in our framework come from our best assessment of all the data we have at a given time. They are not static or fixed. Every experiment we run helps us improve and refine them and our framework and strategy is updated continuously as more data becomes available.

Part 1 – Establishing the goal of your experimentation project

The first part of the framework is the most important by far. If you only have time to do one thing after reading this post it should be revisiting the goal of your experimentation.

Most teams don’t set a clear goal for experimentation. It’s a simple as that. Any strategy needs to start with a goal, otherwise how can you differentiate success from wasted effort?

A simple test of whether your experimentation has a clear goal is to ask everyone in your team to explain it. Can they all give exactly the same answer? If not, you probably need to work on this.

Don’t be lazy and choose a goal like “increase sales” or “growth”. We’re all familiar with the importance of goals being “SMART” (specific, measurable, achievable, relevant, time-bound) when setting personal goals. Apply this when setting the goal for experimentation.

Add focus to your goal with targets, measures and deadlines, and wherever possible be specific rather than general. Does “growth” mean “increase profit” or “increase revenue”? By how much? By when? A stronger goal for experimentation would be something like “Add an additional £10m in profit within the next 12 months”. There will be no ambiguity as to whether you have achieved that or not in 12 months’ time.

Some other examples of strong goals for experimentation

• “Increase the rate of customers buying add-ons from 10% to 15% in 6 months.”
• “Find a plans and pricing model that can deliver 5% more new customer revenue before Q3”
• “Determine the best price point for [new product] before it launches in June.”

A clear goal ensures everyone knows what they’re working towards, and what other teams are working towards. This means you can coordinate work across multiple teams and spot any conflicts early on.

Part 2 – Defining the KPIs that you’ll use to measure success

When you’ve defined the goal, the next step is to decide how you’re going to measure it. We like to use a KPI tree here – working backwards from the goal to identify all the metrics that affect it.

For example, if our goal is “Add an additional £10m in profit within the next 12 months” we construct the KPI tree of the metrics that combine to calculate profit. In this simple example let’s say profit is determined by our profit per order times how many orders we get, minus the cost of processing any returns.

These 3 metrics then break down into smaller metrics and so on. You can then decide which of the metrics in the tree you can most influence through experimentation. These then become your KPIs for experimentation. In our example we’ve chosen average order value, order conversion rate and returns rate as these can be directly impacted in experiments. Cost per return on the other hand might be more outside our control.

When you’re choosing KPIs, remember what the K stands for. These are key performance indicators – the ones that matter most. We’d recommend choosing at most 2 or 3. Remember, the more you choose, the more fragmented your experimentation will be. You can track more granular metrics in each experiment, but the overall impact of your experiments will need to be measured in these KPIs.

Putting that all together, you have the first parts of your new framework. This is our starting point – and it is worth the time to get this right as everything else hinges on this.

Part 3 – Understanding how your audience impacts your KPIs and goal

Now we can start to develop our strategy for impacting the KPIs and achieving the goal. The first step is to explore how the make-up of our audience should influence our approach.

In any experiment, we are looking to influence behavior. This is extremely difficult to do. It’s even more difficult if we don’t know who we’re trying to influence – our audience.

We need to understand the motivations and concerns of our users – and specifically how these impact the goal and KPIs we’re trying to move. If we understand this, then we can then focus our strategy on solving the right problems for the right users.

So how do we go about understanding our audience? For each of our KPIs the first question we should ask is “Which groups of users have the biggest influence on this KPI?” With this question in mind we can start to map out our audience.

Start by defining the most relevant dimensions – the attributes that identify certain groups of users. Device and Location are both dimensions, but these may not be the most insightful ways to split your audience for your specific goal and KPIs. If our goal is to “reduce returns by 10% in 6 months”, we might find that there isn’t much difference in returns rate for desktop users compared to mobile users. Instead we might find returns rate varies most dramatically when we split users by the Product Type that they buy.

For each dimension we can then define the smaller segments – the way users should be grouped under that dimension. For example, Desktop, Mobile and Tablet would be segments within the Device dimension.

You can have a good first attempt at this exercise in 5–10 minutes. At the start, accuracy isn’t your main concern. You want to generate an initial map that you can then start validating using data – refining your map as necessary. You might also find it useful to create 3 or 4 different audience maps, each splitting your audience in different ways, that are all potentially valid and insightful for your goal.

Once you have your potential audiences the next step would then be to use data to validate the size and value of these audiences. The aim here isn’t to limit our experiments to a specific audience – we’re not looking to do personalization quite yet. But understanding our audiences means when we come to designing experiments we’ll know how to cater to the objections and concerns of as many users as possible.

Part 4 – Identifying the areas with the greatest opportunity to make an impact

Armed with an better understanding of our audience, we still need to choose when and where to act to be most effective. Areas is about understanding the user journey – and focusing our attention on where we can make the biggest impact.

For each audience, the best time and place to try and influence users will vary. And even within a single audience, the best way to influence user behavior is going to depend on which stage of their purchase journey the users are at.

As with audiences, we need to map out the important areas. We start by mapping the onsite journeys and funnels. But we don’t limit ourselves to just onsite experience – we need to consider the whole user journey, especially if our goal is something influenced by behaviors that happen offsite. We then need to identify which steps directly impact each of our KPIs. This helps to limit our focus, but also highlights non-obvious areas where there could be value.

As with audiences, you can sketch out the initial map fairly quickly, then use analytics data to start adding more useful insights. Label conversion and drop-off rates to see where abandonment is high. Don’t just do this once for all traffic, do this repeatedly, once for each of the important audiences identified in the previous step. This will highlight where things are similar but crucially where things are different.

So with a good understanding of our audiences and areas we can add these to our framework. Completing these two parts of the framework is easier the more data you have. Start with your best guess at the key audiences and areas, then go out and do your user-research to inform your decisions here. Validate your audiences and areas with quant and qual data.

Part 5 – Identifying the potential levers that influence user behavior

Levers are the factors we believe can influence user behavior: the broad themes that we’ll explore in experimentation. At its simplest, they’re the reasons why people convert, and also the reasons why people don’t convert. For example, trust, pricing, urgency and understanding are all common levers.

To identify levers, first we look for any problems that are stopping users from converting on our KPI – we call these barriers to conversion. Some typical barriers are lack of trust, price, missing information and usability problems.

We then look for any factors that positively influence a user’s chances of converting – what we call conversion motivations. Some typical motivations are social proof (reviews), guarantees, USPs of the product/service and savings and discounts.

Together the barriers and motivations give us a set of potential levers that we can “pull” in and experiment to try and influence behavior. Typically we’ll try to solve a barrier or make a motivation more prominent and compelling.

Your exact levers will be unique to your business. However there are some levers that come up very frequently across different industries that can make for good starting points.

Ecommerce – Price, social proof (reviews), size and fit, returns, delivery cost, delivery methods, product findability, payment methods, checkout usability

Saas – Free trial, understanding product features, plan types, pricing, cancelling at the end of trial, monthly vs annual pricing, user onboarding

Gaming – welcome bonuses, ongoing bonuses, payment methods, popular games, odds

Where do levers come from? Data. We conduct user-research and gather quantitative and qualitative data to look for evidence of levers. You can read more about how we do that here.

When first building our framework it’s important to remember that we’re looking for evidence of levers, not conclusive proof. We want to assemble a set of candidate levers that we believe are worth exploring. Our experiments will then validate the levers and give us the “proof” that a specific lever can effectively be used to influence user behavior.

You might start initially with a large set of potential levers – 8 or 10 even. We need a way to validate levers quickly and reduce this set down to the 3–4 most effective. Luckily we have the perfect tool for that in experiments.

Part 6 – Defining the experiments to test your hypotheses

The final step in our framework is where we define our experiments. This isn’t an exercise we do just once – we don’t define every experiment we could possibly run from the framework at the start – but using our framework we can start to build the hypotheses that our experiments will explore.

At this point, it’s important to make a distinction between a hypothesis for an experiment and the execution of an experiment. A hypothesis is a statement we are looking to prove true or false. A single hypothesis can then be tested through the execution of an experiment – normally a set of defined changes to certain areas for an audience.

We define our hypothesis first before thinking about the best execution of an experiment to test it, as there are many different executions that could test a single hypothesis. At the end of the experiment the first thing we do is use the results to evaluate whether our hypothesis has been proven or disproven. Depending on this, we then evaluate the execution separately to decide whether we can iterate on it – to get even stronger results – or whether we need to re-test the hypothesis using a different execution.

The framework makes it easy to identify the hypothesis statements that we will look to prove or disprove in our experiments. We can build a hypothesis statement from the framework using this simple template

“We believe lever [for audience] [on area] will impact KPI.”

The audience and area here are in square brackets to denote that it’s optional whether we want to specify a single audience and area in our hypothesis. Doing so will give us a much more specific hypothesis to explore, but in a lot of cases we may also be interested in testing the effectiveness of the lever across different audiences and different areas – so may want to not specify the audience an area until we define the execution of the experiment.

Your first draft of the completed framework will have a large number of audiences, areas and levers, and even multiple KPIs. You’re not going to be able to tackle everything at once. A good strategy should have focus. Therefore you need to do two things before you can define a strategy from the framework.

Prioritise KPIs, audiences and areas

We’re going to be publishing a detailed post of how this framework enables an alternative approach to prioritization than typical experiment prioritization.

The core idea is that you need to first prioritize the KPI you most need to impact from your framework in order to achieve your goal. Then evaluate your audiences identify those groups that are the highest priority groups to influence if we want to move that KPI. Then for that audience prioritize those areas of the user-journey that offer the greatest opportunity to influence their behavior.

This then gives you a narrower initial focus. You can return to the other KPIs at a later date and do the same prioritization exercise for them.

Validate levers

You need to quickly refine your set of levers and identify the ones that have the greatest potential. If you have run experiments before you should look back through each experiment and identify the key lever (or levers) that were tested. You can then give each lever a “win rate” based on how often experiments using that lever have been successful. If you haven’t yet started experimenting, you likely already have an idea of the potential priority order of your levers based on the volume of evidence for each that you found during your user-research.

However, the best way to validate a lever is to run an experiment to test the impact it can have on our KPI. You need a way to do this quickly. You don’t want to invest significant time and effort testing hypotheses around a lever that turns out not have ever been valid. Therefore for each lever you should identify what we call the minimum viable experiment.

You’re probably familiar with the minimum viable product (MVP) concept. In a minimum viable experiment we look to design the simplest experiment we can that will give us a valid signal as to whether a lever works at influencing user behavior.

If the results of the minimum viable experiment show a positive signal, we can then justify investing further resource on more experiments to validate hypotheses around this lever. If the minimum viable experiment doesn’t give a positive signal, we might then de-prioritize that lever, or remove it completely from our framework. We’ll also be sharing a post soon going into detail on designing minimum viable experiments.

Creating a strategy

How you create a strategy from the framework will depend on how much experimentation you have done before and therefore how confident you are in your levers. If you’re confident in your levers then we’d recommend defining a strategy that lasts for around 3 months and focuses on exploring the impact of 2-3 of your levers on your highest priority KPI. If you’re not confident in your levers, perhaps having not tested them before, then we’d recommend an initial 3-6 month strategy that looks to run the minimum viable experiment on as many levers as possible. This will enable you to validate your levers quickly so that you can take a more narrow strategy later.

Crucially at the end of each strategic period we can return to the overall framework, update and refine it from what we’ve learnt from our experiments, and then define our strategy for the next period.

You can have a first go at creating your framework in about 30 minutes. Then you can spend as long or as little time as you like refining it before you start experimenting. Remember your framework is a living thing that will change and adapt over time as you learn more and get more insight.

1. Establish the goal of your experimentation project
2. Define the KPIs that you’ll use to measure success
3. Understand how your audience impacts your KPIs and goal
4. Identify the areas with the greatest opportunity to make an impact
5. Identify the potential levers that influence user behavior
6. Define the experiments to test your hypotheses

The most valuable benefit of the framework is that it connects all your experimentation together into a single strategic approach. Experiments are no longer islands, run separately and with little impact on the bigger picture. Using the framework to define your strategy ensures that every experiment is playing a role, no matter how small, in helping you impact those KPIs and achieve your goal.

Alongside this, using a framework also brings a large number of other practical advantages:

• It’s clear– your one diagram can explain any aspect of your experimentation strategy to anyone that asks or if you need to report on what you’re doing
• It acts as a sense check – any experiment idea that gets put forward can be assessed based on how it fits within the framework. If it doesn’t fit, it’s easy rejection with a clear reason why
• It’s easy to come back to – things have a nasty habit of getting in the way of experimentation, but with the framework even if you leave it for a couple of months, it’s easy to come back to it and pick up where you left off
• It’s easier to show progress and insight – one of the biggest things teams struggle with is documenting the results of all their experiments and what was learnt. With the framework, the idea is that the framework updates and changes over time so you know that your previous experiment results have all been factored in and you’re doing what you’re doing for a reason

As we said at the start of this post, there is no special sauce in this framework. It’s just taking a logical approach, breaking down the key parts of an experimentation strategy. The framework we use is the result of over 10 years of experience running experimentation and CRO projects and it looks how it does because it’s what works for us. There’s nothing stopping you from creating your own framework from scratch, or taking ours and adapting it to suit your business or how your teams work. The important thing is to have one, and to use it to go from tactical to strategic experimentation.

You can find a blank Google Slide of our framework here that you can use to create your own.

Alternatively you can download printable versions of the framework if you prefer to work on paper. These templates also allow for a lot more audiences, areas, levers and experiments than we can fit in a slide.

 

If you would like to learn more, get in touch today!