How Machine Learning is Affecting Your Open Rates

Organizations that we work with have been reporting a widespread decrease in open rates on messages sent to Gmail addresses. The finger has been pointed, with much derision and some uncertainty, at Gmail’s recent implementation of TensorFlow, a machine learning tool, to help upgrade its spam-fighting efforts. This isn’t the first time that Google has made changes to its spam prevention methods and legitimate marketers have taken a hit to their bottom line, but this may be the most lethal bodyblow to email marketing yet… and that’s a great thing.

For the record, this is not another “email marketing is dead, unless you pay us to reveal our secret sauce” post. That’s the thing with machine learning – there isn’t a secret sauce anymore, or a hidden way to game the system. What’s going to keep you out of a user’s inbox is no longer a decision a programmer made out in Palo Alto, or a messed up header – it’s user behavior, and the algorithm that feeds it. But the good news is, if email marketers change their behavior, this change should help us all.

Quick machine learning primer, for the uninitiated: machine learning is a form of predictive intelligence, like an AI you would see in the movies (albeit much reduced). Computer scientists create algorithms, based on statistics and a whole lot of math, to classify and sort information into groups that the computer scientist defines, like “Good Email” and “Spam Email”. This is done by taking a large set of data (such as all the emails sent to addresses), manually sorting it into the “Good Email” and “Spam Email” categories, and then showing it to your machine learning algorithm. This is where the magic happens – the algorithm will comb through the data in order to identify the traits of the “Good Email” group, as well as the “Spam Email” group, without being explicitly told what those traits are; the computer will decide that for itself.

With that in mind, imagine Gmail didn’t just categorize email into the “Good” and “Spam” categories – they took the categorization further, adding “Promotions”, “Updates”, and “Forums”. Now after training their machine learning models on these labels and groupings of emails, their systems can automatically push emails that users may not want to read to these different inboxes. The end result is a better user experience, as you mostly only see emails that you truly want in your inbox.

The salient point here is that Google isn’t necessarily just training their models to distinguish legitimate emails from spam . It’s deeper than that – because they’re fundamentally, they’re being trained to distinguish quality and relevance of content. The emails people are most likely to open are the ones that have information they want to receive, or the ones that have interesting, high quality content that they want to read – everything else is going to be pushed into Google’s tab structure. As a result, anyone sending email is going to have to start contending with an algorithm that will judiciously sort any incoming mail based on previous user behaviors, and if those user behaviors indicate no one wants to receive your content, then there’s no way around it – you’re going to get filtered.

While there’s no way to get around the algorithm, there’s a few ways to make sure it’s not pushing your open rates lower unnecessarily:

  1. Use Confirmed Opt-In. If people definitely want to receive your messages, they’re more likely to read them, and therefore the algorithms will treat your mail better.
  2. Send to smaller lists. Machine learning is dependent on data – if you feed the system less data, it’s less likely to pick up trends in your marketing and start trying to clamp down on it. This won’t help you if other marketers are using the same content and practices though, which is why we also recommend…
  3. Send email only when you have good content to share. The better the content in your email, the more likely people are to want to read and share it, and the better you’ll look to email service providers. That’s the biggest, best recommendation that anyone in the industry can make to you right now.

Have questions about machine learning or how it affects your enterprise? Drop me a line at