The deliverability discussion calls are going well and I’m going continue to host them on a biweekly basis. Next call will be May 6th, 5pm Ireland time, noon Eastern and 9am Pacific time. Still doing invites manually, so drop me an email at laura-ddiscuss@ the obvious domain.
A few weeks ago, the discussion entered around machine learning in general. As a follow on I’d planned to talk about was how different ML filters are trained.
Almost all filters out there are based on machine learning. The commercial B2B filters, the filters at consumer mailbox providers, all filters have some components based on ML. But there’s a big difference in how those filters are trained and the data they have access to. Gmail, VMG and Microsoft all have access to the the mail client as well as the mail server. They can, and do, track user activity with mail as part of how they teach the engines.
Filtering appliances don’t have the same inputs as the mailbox providers do. They don’t have access to the mail client. That ML is not based on user interaction or engagement at all.
We did talk a little bit about that, and what folks’ experiences were, but then the conversation wandered. It was a good wander. We talked a lot about how we described filtering and filters and delays to people. I described a cake analogy a client shared with me. Another person described filters as tomato sauce.
The cake analogy: making changes at Gmail in particular is like baking a cake. You put all the ingredients together, mix them well and put the cake in the oven. Then you have to wait for it to bake. If you keep poking it, or opening the oven then you just make the cake worse. When you’re trying to fix delivery problems at Gmail, you need to make the changes and then just wait for the filters to catch up.
The tomato sauce analogy: (any errors in transcription are mine) A company wants to make tomato sauce. They want to make the best tomato sauce there is. So they make one kind of tomato sauce. But different people want different kinds of tomato sauce. Some people want chunky sauce, some want smooth sauce, some want really garlicky sauce, some want meaty sauce. A successful company makes all kinds of tomato sauce to meet the needs of different kinds of customers.
We also talked about how the size of the sender does matter. Smaller senders and larger senders are treated differently by the filters. What works when you’re small doesn’t always work when you’re big. And, what works when you’re big doesn’t always translate down to smaller senders.
It was a fun call. Afterwards I got a message from a participant saying they really enjoyed it and found it “fascinating how some scenarios can be so nuanced especially between smaller and larger senders and transactional versus promotional. There has been so much shared from everyone and the machine learning was really helpful as I definitely heard new information.”
Start those emails coming for the next call. Can’t wait to talk again.