Before we begin, take a second and read The State of Split Testing survey and report to see how you and hundreds of your close and personal email marketing friends are benefiting (or not) from split testing.
We’ll give you five minutes to skim through it.
OK, are you ready to get into the nitty gritty of why Phrasee is the biggest revolution in email marketing since the launch button?
Email marketers want to test their subject lines, body copy, and calls-to-action (CTAs) better. The State of Split Testing report made that crystal clear. But, there are substantial limitations – including robust test design, thinking of original language, and importantly learning from the results.
That’s why we built Phrasee. We have done the research and understand the problems email marketers face.
The best way to explain Phrasee in gory detail is by splitting our methodology into three areas – language, experimental design and artificial intelligence.
WARNING: This is a very long and very detailed explanation. If you only have 5 minutes then bookmark this page and come back later when you have more time! It’ll take you about 13-15 minutes to read this full description. What can we say, there’s a lot going on in Phrasee.
1. Language methodology
Language is difficult to process programmatically. Very difficult. Take, for example, Google Translate. While it’s a great service that we love, it’s still not perfect, despite having some of the best minds in the business working on it.
Why is this? Because humans interpret language in a very unique, complicated and intelligent way.
Give yourself a pat on the back, because you’re unique, complicated and intelligent.
Let us illustrate. Consider this passage from our CEO’s favourite book:
“At present I do not feel that I have seen more than the fringe of poverty. Still I can point to one or two things I have definitely learned by being hard up. I shall never again think that all tramps are drunken scoundrels, nor expect a beggar to be grateful when I give him a penny, nor be surprised if men out of work lack energy, nor subscribe to the Salvation Army, nor pawn my clothes, nor refuse a handbill, nor enjoy a meal at a smart restaurant. That is a beginning.”
There isn’t a machine in the world that could effectively interpret this passage, let alone write it. Humans have evolved to become highly intelligent linguistic entities, and machines haven’t. This is why the Turing Test hasn’t been successful to date.
Machines are not able to write literature any better than a thousand monkeys with a thousand typewriters.
But: marketing language is not literature. It is a small collections of words and phrases that elicit a particular response from your customers.
What does this have to do with Phrasee? Let us explain using an example:
“Get £50 off from our amazing sale on unforgettable holidays!”
This is one of trillions of human-sounding sentences Phrasee can generate. It is made up of a limited set of language constructs. Changing one bigram or trigram (two and three word sets, respectively,) for example, substantially changes the overall human interpretation of the email. For example:
“Receive £50 off from our incredible deals on breathtaking vacations!”
How many other ways can you say the same thing? The answer, in case you’re wondering, is trillions…
These are substantially different subject lines. While the gist may be similar, the overall interpretation of the phrase will be different. Each phrase set is interdependent on the others. Changing one bigram will change the entire interpretation of the surrounding language. This is why looking at individual words is a suboptimal strategy to improve your marketing language.
While the entire variant space is large, by using some email marketing heuristics and real-time training data Phrasee is able to deconstruct subject lines into a language set.
What’s a language set? It’s an algorithm written by our computational linguists specifically for your brand. This allows Phrasee to generate human-sounding subject lines in your brand’s voice, and no one else’s.
This, in turn, is augmented with a proprietary sentiment analyser, so we can determine emotional response to given language sets. It’s further helped by a natural language processing toolkit that avoids about 99.7% of grammatically incorrect outputs.
Old-fashioned copywriters don’t always like the language Phrasee generates. They sound different from what you would write.
And if you’re very precious about your own words being used instead of machine-optimised ones, then Phrasee isn’t for you.
You should only use Phrasee if you care about getting more opens, clicks and conversions.
The results don’t lie. Phrasee-generated language beats humans over 95% of the time, and increases response rates. One of our best yet is a 417% revenue lift – not bad, huh?
This is why people who use Phrasee get better open, click and conversion rates than anyone else.
Marketing language, when generated, tested and learned from effectively, outperform humans.
Phrasee is the software that will prove this to you.
2. Experimental methodology
If anyone tells you that split testing in email marketing is no longer necessary, they are categorically and scientifically wrong.
Split tests, also known as controlled experiments, are how we learn about the world around us. This is a fundamental truth.
The world changes. Your marketplace changes. Your customers change. If you aren’t constantly testing to see what works then you are assuming that things will stay the same forever.
This is why EVERY scientific study involves controlled tests. Be it clinical trials of pharmaceuticals through to psychology experiments at your local university – anyone who wants to learn about their customers and marketing environment needs to develop hypotheses and then run tests to learn about them.
There is a science to marketing language. That’s why Phrasee treats every split test like a clinical trial.
There is much scientific rigour involved in developing a robust experiment. It involves a few steps, which we’ve massively abridged for your reading pleasure:
- Calculating an appropriate number of control and test groups
- Defining required size per test group to ensure statistical significance
- Developing hypotheses that define specific planned learning outcomes
According to The State of Split Testing study, most email marketers only split test once in a while… and when they do, they don’t follow robust experimental design practices. This is a bad thing.
But good news! Phrasee automates and optimises the experimental design process.
Phrasee does the following things, once again massively abridged for your reading pleasure:
- Phrasee calculates the optimal number of splits for you to run, given your list size and previous results.
- It tells you know how big each split of your data should be to both learn but also maximise your results from sending the winner to everyone else. We often throw in random curveballs to control for experimental bias.
- It then predicts what hypotheses should be tested to maximise learning while driving as many results as possible in the meantime (more on this in the next section.)
We believe that marketers should be scientists. That’s why we built Phrasee to follow the scientific method.
Phrasee’s testing methodology ensures both short- and long-term optimization of your subject lines, body copy and CTAs… and your response rates.
The results don’t lie. Like that time a holiday company told us they calculated their ROI from Phrasee was >10,000%. Boom!
3. Artificial intelligence methodology
Artificial Intelligence (AI) is a buzzword these days. Many claim to do it but actually don’t. At Phrasee, well, we actually do it. Our chief scientist Dr Neil Yager wrote the book on it (literally…)
AI as a discipline is often misinterpreted. Lots of people seem to think that you feed a computer some data, it will magically create an algorithm based upon matching some sort of patterns and boom, AI. This isn’t the case.
AI works best when it is combined with human heuristics and intuition. This is where Phrasee’s combination of algorithmic know-how and email marketing expertise really shines.
Phrasee holds lots of data – millions and millions of records of subject line tests that have been created and sent through Phrasee’s platform. Sure, we could train a machine to go through this data and find patterns. Others have tried and wasted a lot of time doing so. Why is it a waste of time? It would be telling us what we already know!
Think about it like this – say you have a huge set of weather data for London, England. This data covers every minute of every day over the last 100 years. You could train the smartest machines in the world and they would eventually infer that, guess what, in February it rains a lot in London. Wow. That’s good to know.
Oh, btw, we already knew that.
The Phrasee team already knows a thing or two about marketing language. In fact, our CEO is probably the world’s leading expert in the advanced statistical analysis of it (a fact which doesn’t go down well at parties, unfortunately.)
We don’t see the point of wasting your time or ours letting a machine confirm what we intuitively know about email marketing.
Instead, we use our email expertise to pre-define algorithms to expedite the learning process. We use machine learning to find intricate patterns invisible to human eyes to powerfully move things forward.
Phrasee’s optimisation engine follows a few key steps:
- First, it looks at the probabilistic relationship between thousands of variables (including a time decayed series of your past results, syntactic and semantic variance, normalised quality scores, interdependence of individual language elements, and external factors such as language, currency and seasonal holidays.) From here, it makes some statistical predictions about what ought to work.
- Next, it runs millions of simulations to predict the outcomes of sending various subject lines to your audience. This lets it know if its assumptions are likely to be correct or not. The engine adjusts itself on the fly to improve the predictions.
- It then funnels the results into a genetic mutation algorithm that creates millions of permutations of each of the selected subject lines. This is recursively entered back into step 2 until the engine is satisfied that the resultant subject lines present the highest possible learning outcome.
The output you see in Phrasee – that list of language for you to approve – is the result of a huge number of language variants being considered, simulated, and refined.
When you feed your results back into Phrasee, the statistics get fed directly into its machine learning engine. These are the strongest signals around, and are key for maximal optimisation.
The more you tell Phrasee about your results, the more Phrasee learns, and the better your results get.
4. What does this mean for you
AI-generated marketing language beats humans >95% of the time.
Phrasee is the only software on the planet that combines real-world marketing expertise with cutting-edge artificial intelligence.
This is how Phrasee will turn you into a marketing language superhero.