For all of the vast resources the big four free inbox providers (Yahoo!, AOL, Gmail and MSN/Live/Hotmail) have expended over the decades to tweak, test and improve their filtering and delivery processes, they’ve been remarkably quiet about it. But their reticence is for good reason; if they were to publish the juicy details, they’d be undermining their own efforts, as spammers would surely make use of the information to evade filtering changes almost as soon as they are implemented.
So when a major inbox provider actually publishes specific details about how they make delivery decisions – as Gmail recently did in a paper explaining how Priority Inbox works – it creates a splash in the deliverability pond.
In January, Google posted on one of its research blogs a four page document containing the mathematical algorithms Gmail employs to decide how important any particular message is to its recipient by gauging the probability that the recipient will act quickly on the message.
It’s nothing short of a mathematical model of recipient engagement.
The real utility of the paper, in my opinion, is that it gives senders a clearer picture of what engaging e-mail looks like to the ISPs. I propose that Priority Inbox is a reasonable proxy for any recipient domain where it comes to measuring engagement. Senders who test and optimize their mail for preferred placement in the Gmail Priority Inbox should see gains in deliverability most anywhere else.
So what does engaging e-mail look like to Gmail? I’m not mathematically savvy enough to evaluate the algorithms in the paper, but the author provides some useful narrative:
Labelling: First, mail that the recipient marks as important is definitively important. The real trick, as the author notes, is how to infer importance without explicit recipient labelling. To do that, the algorithms measure hundreds of “features” of the message that fall roughly into one of three categories: Social, Content, and Thread features.
Social features: This has nothing to do with content-sharing activity on social networks, as some might assume. Instead, it’s a measure of the degree of interaction between the sender and the recipient, like the percentage of a sender’s mail that is opened by the recipient.
Content features: Gmail looks for keywords in the content, header and subject lines of the message that correlate with higher opens (or lack of opens).
Thread features: Gmail looks at aspects of the thread (or “conversation”) to which the message belongs. Has the recipient taken action on other messages in the same thread? If so, what percentage of the overall number of messages in the the thread were acted upon? Did the recipient actually start the conversation? If so, then Gmail assumes the message is important.
Beyond a treatment of the features of a given message, the author addresses a pair of interesting aspects:
Important or merely interesting: Opening a mail is a strong signal of importance to Gmail, but the author notes that many users also open a lot of mail that is “interesting” rather than “important”. So Gmail’s math tries to calculate where to draw the line between the two.
Time to action: The real goal of all the math is to predict the probability that a recipient will act on the mail within a given number of seconds following delivery to the inbox. The less time it takes, the more likely the mail is to be scored as “important”.
Part of the problem, of course, is that where Gmail places the message in the inbox directly affects how long it takes the recipient to notice and act on the message. Gmail tackles this, too, but at this point in my reading of the paper, my slide rule burst into flames.
This mathematical treatment of engagement metrics offers a rare glimpse beneath the kimono of a major inbox provider, and it confirms in a general sense much of what the deliverability community has been able to infer largely through observation: engagement metrics are playing an increasingly large role in automated delivery decisions at recipient domains. Senders who are serious about improving delivery (and, ultimately, ROI) will test and optimize their mail for engagement.