Get 100 email copywriters in a room and ask them one of the things they hate most about their job; chances are you’ll hear “writing subject lines” in the first few seconds.
I know. I know. If you don’t actually write subject lines, it seems fun. Till you read the results.
Writing subject lines that work is like pleasing Goldilocks. The subject lines need to be interesting enough for a user to want to open your email, and they need to be “safe” enough to be delivered in the first place.*
Plus, your emails not only have to avoid the dreaded SPAM and Trash folders, but they also need to get out of your ESP in the first place. (If I had a dollar for all the suppressed names these days….)
Assuming your email has been inboxed, readers will decide to open your emails primarily based on sender name (read: from address) and/or the subject line, so it’s essential to get them right. It’s also one of the toughest.
That’s where Artificial Intelligence/Machine Learning (AI/ML) comes in.
AI/ML tools use NLG (Natural Language Generation) to write subject lines automagically for every individual on your list.
How do the AI/ML systems do it?
They use algorithms based on past/current individual behavior, available 0-3 party data and overlays, combined with marketing and overall campaign analysis.
Will they perform better than what my copywriter, with 25 years of experience, writes?
If your veteran copywriter and/or other people in your organization train(s) the system, absolutely. If not, it will likely happen but it will take longer than you expect. (Training the algos is a crucial part of your success when you’re using AI/ML.)
Remember, the BIG benefit of AI here is that it does things at scale — so EVERY single person on your list gets an individualized subject line. Unless you’re only mailing to a handful of people, it would be impossible for any copywriter to execute this. Coupling human talents, experience and expertise with AI’s prediction and capabilities is a powerful force.
My AI/ML email vendor said that they could do everything. Why do I need to get my team involved?
It’s good to have your folks involved because they know your brand and audience best. If they’re involved from the jump, they can ensure that the AI/ML has some structure and doesn’t run amok. (If you’re going to leave your dog alone in the yard all day, you don’t need to tie it to a post, but it’s probably best for everyone involved that your yard is fenced. Also, please don’t do this.)
What do I optimize for first?
In the beginning, most companies optimize for inbox placement and/or opens. After that is perfected, they move on to test things like how it impacts the NTD (Next 10 Days.) Pinpointing the best NTD path can be incredibly lucrative, and AI/ML is beneficial here.
How many subject lines will AI/ML create?
Approximately eleventy bazillion.
Here’s the thing… If you’re using a package to write the subject lines automatically, you should start small and work your way up. This allows you to keep a tight rein on your brand and how it is being portrayed. As you get more comfortable with the technology and the results you are seeing, you can increase the level of personalization, segmentation, targeting, etc., to maximize your engagement and revenues.
After I’ve gotten my data ready, what do I do?
First, list all your brand rules concerning subject lines, so you have them in one handy place. Be sure to include any rules you have about tone. For example, do you only want to be friendly/upbeat/positive? Perhaps you’d prefer to sound more provocative? Are you ok with using sales/offers/deals in your subject lines? Do you like them to be very deadline-oriented? Are you alright with humor? What do you want to use when you don’t have a name for personalization?
Next, list any biases you may have had. Initially, the algos will learn from past data, so it’s essential to make a list of any biases that may have impacted your results at one time or another.
For example, a couple of years ago, it was all the rage for email consultants to spout that you needed super short subject lines because they were working better on mobile. (Legit.) So, these days you may limit your characters without even recognizing it. Other companies only used questions. Some started all their subject lines with the company name and then a colon. Some were high on emojis. You might not put the product names in your subject lines because you can’t truncate the length, and it “breaks your current provider.” You also might not use the word FREE or exclamation points because you were told to avoid it ten years ago. These are all biases that should be noted.
Whatever biases you have, it’s a handy exercise to closely examine your history. Identify the patterns. Make a list of your observations. If you’ve been emailing for a long time, you may not even remember all the logic/reasoning; just jot down what you can. This will be helpful if/when you want to try/test things you’ve previously shelved. (And yes, your vendor will likely tell you that they’ll consider all this. Please do it anyway. It’s worth the effort.)
Then, what do I do?
Some consultants recommend giving the system 12-25 hand-picked subject lines with the best open rates. Others suggest you give the system ALL the data and then have it learn. Most marketers find that somewhere in the middle is probably best – meaning provide the system as much curated and tree data as you possibly can. (For example, if you send out an Abandoned Browse email with the abandoned product name in the subject line, the trunk data would be Still Interested in… and the personalized/branch data would be the name of the products.)
Be sure to give your subject line data BY DEVICE. Many folks forget this step, and it can make a difference, especially when you’re selling to B2B and older audiences. (When your AI/ML-enhanced email system is working correctly, you’ll send out individualized subject lines and, different subject lines work better on different devices.)
My vendor suggested I suppress all our failures and bad/biased data, should I?
This sounds great in theory, but it’s often short-sighted in practice. A better choice is to demarcate the things you don’t want to repeat. (If you isolate the data but allow it to be analyzed and learned from, but not acted upon without your consent, you can learn much more from it.)
Just so you know, how your vendor/program handles “bad” data is often a sign of how sophisticated their intelligence is.
Should I give my vendor my transactional and trigger email information?
Definitely, just make sure the data is marked with the campaign; otherwise, you’ll end up with only Thank You, Oops! and We Miss You subject lines.
Incidentally, if your business is highly seasonal, make sure to list that as a factor when you’re submitting your data. Things that work around the holidays in B2C and retail or near government and education fiscal periods in B2B can inadvertently skew your data if you’re not paying attention to it.
Do I include my split/multi-variate testing information?
Yes, make sure it’s delineated.
Typically, the more “good” data you analyze, the better your initial model(s) will be and the higher open rates and better inboxing it will yield. If you do point testing, it’s often best to eliminate or separate that. If you do a lot of emoji testing, ensure that the emojis are appropriately translated.
My vendor is recommending just the upfront data. Is that ok?
Giving the system data through purchase is a better choice. Several of the sketchier vendors are recommending upfronts only because they think they can hook you with their stellar open rates. Optimizing just for adjusted open rates is exciting, but if you don’t account for engagement and conversion (or at least adoption), it can be very misleading.
What is subject line scoring?
Many software providers have a built-in subject line scoring model that looks at whether the specific word(s) have a positive or negative impact on your subject line. Scoring tools vary in quality, so review what you’re using. They often don’t take brand voice into account, and many are spotty at grammar and acceptable emoji use. By the way, if you’re using a scoring tool, you should review the results on an ongoing basis. More frequently if the scoring impacts ongoing/future results. (Not all scoring tools are predictive.)
What’s the most important thing to remember when using AI/ML in my subject lines?
Your ultimate goal is to increase your revenues and/or decrease your costs. A lot of the software being sold right now has a lot of VERY shiny distractions, but if you don’t make more money and/or save more time, it’s probably not worth it.
#1 best tip?
Remember, when you’re first starting, you only/mostly use past data to predict future outcomes. Email changes a lot, so before integrating AI in your subject lines, ensure you’re committed to training and maintaining the algos. If you plan on this being a set-it-and-forget-it process, do something else. For a lot of reasons (signals, sheer volume, etc.), email can go astray VERY quickly.
If you’re ready, though, it’s a great entry point to learning more about how you can use artificial intelligence in your business.
Plus, the information you get from all your subject line work will be FANTASTIC for your SMS efforts!
Have a question about using AI/ML in your subject lines? Have a tip you’d like to share? Tweet @amyafrica or write email@example.com.
*These days, a HUGE part of successful email deliverability is about your Sender/IP reputation. If your reputation is top-notch, you’ll have a lot of leeway in what you write in your subject lines. If your deliverability is marginal or your provider is substandard, there’s a big chance you’ll have many more “rules” to adhere to.