Mail w/ confidence, not crossed fingers — an intro to propensity scores

So,

In our last newsletter, we shared some Postie research data that stated 90% of marketers are prioritizing prospecting in 2025.

Remember that? If not, give it a read.

What followed was an in-depth breakdown of LAL audiences and how, when done right, they amplify the performance of your prospecting campaigns.

We even have the numbers to back up the jibber-jabber. 👇

But this month, we’re peeling back the next layer of the onion and how Posite takes LAL audiences one step further into the future — propensity scores.

Let’s dive in.

————-

For those who didn’t read the last newsletter, here’s the TLDR courtesy of Wilson, my GPT.

TLDR

Marketers are shifting focus hard toward prospecting in 2025 (90% of ’em, to be exact), but most are still relying on outdated lookalike models.

This newsletter breaks down:
🔍 Why clean, enriched CRM data is your LAL secret weapon
🤖 How machine learning ranks prospects by likelihood to buy
💥 The difference between table-stakes audiences and turbocharged targeting

It’s got strategy, science, and a little serotonin boost. Dive in.

We glossed over the aspect of applying a ranking structure to the LAL model audience, which isn’t fair because the implications of executing this are profound.

Let’s start at the beginning

You are a marketer for a brand looking to run prospecting campaigns. No surprise there.

Here’s a couple of ways you might go about doing that:

• Buy a list of non-customers (expensive and often littered with junk prospects)
• Lean into digital audiences and the walled garden audience-building tools (zero transparency, fluctuating ad costs)
• Brand play with out-of-home mediums (billboards, blimps!, planes!!)
• orrrrr, self-reflect, look into your CRM, find customers that fit your ICP, and use them to build audiences (this works in digital and DM)

While all have their time and place, in the name of efficiency, we’re major advocates of the last option.

The table-stakes approach ingests your seed audience (whatever that looks like) and spits out people who look similar to your seed audience.

Then, you take that audience and run digital ads against them. If you’re lucky and can get a respectable email match rate, you can start emailing them too (hopefully with limited spam reports).

But this method falls short in two areas:

If you’re a marketer right now, odds are you’re underwater with work. You wish you could do it all, but if you’re realistic, you have to triage.

Additionally, even if you did have a copious amount of time to achieve your tasks, you most likely won’t have the budget.

It’s a one-two punch – and it sucks. I know!

Not to sound promotional (🤮), but this is where Postie SERIOUSLY can add value to marketers’ lives.

We don’t stop at giving you a LAL audience. We tell you who, out of that audience, is most likely to convert NOW and who you shouldn’t waste your precious budget on.

We do this by building a ranked list of every household in the U.S. and run your new audiences up against it.

Here’s how we do it.

Models on Models on Models

At Postie, we start with looking at your customer set (audience) and another set of non-customers (population) and run an ML model to find the differences between non-customers and customers.

Then, we evaluate these trained ML models on the entire population of the continental United States via a variety of data sets (axiom, experian, epsilon).

It looks like this.

Each model delivers different propensity prediction scores.

To the untrained eye, some might see the predicted performance discrepancy as a negative indicator.

At Postie, it’s data fidelity and an increase in data fidelity leads to more accurate predictions and peak campaign performance because of the variety of variables being considered.

Variety is the spice of life, after all.

Calibrating LAL Models

The first thing to know is the more you send, the FASTER your model calibrates or learns.

It’s LITERALLY like riding a bike; the more you do it, the better you get, and faster.

A brand sending 1MM pieces of mail a year is going to calibrate slower than a brand sending 10MM pieces a year.

Simple.

But it’s not just about how MUCH you send, but also HOW you send.

‘Ryan, what in the H-E-double hockey sticks are you talking about?’ you ask.

Well, if that brand sending 10MM pieces a year only used one model to define who they prospected to, two things would happen:

1. They would have limited and homogeneous performance data to optimize their model with (slower calibration, despite high send volume)
2. Their model performance will decay (more on that later),and they’ll have nowhere to turn to for better performance

What we do with all our customers to avoid these pitfalls is deploy a lil’ sampler, so to speak, of models (over time).

Here, you can see we have a variety of different models we’ve used to send to prospects.

[MAYBE DELETE THIS] Each dot represents a cluster of 25k households that shared a similar conversion propensity score.

What you also might notice is each line isn’t stacked on top of each other meaning that the predicted and actual performance varied based on the model.

This diverse performance data is used to build robust and highly performant prospecting models that are optimized for your brand and your brand alone.

With all this crystalizing into what we call a universal model score for each household in the continental United States. More on that in another newsletter.

What about point 2?!

Ahh yes, point 2, performance decay.

Look, all great things come to an end. It’s a fact of life, and marketing campaign performance is not immune to that.

But you can plan for it, and it makes the sting a lot more tolerable.

Referring back to the visual of the different models and their performance, what do you see there?

I see redundancy (and decay).

Inevitably, as you nail down an individual model, you begin to run out of prospects in the top decile, and your campaign performance will drop.

That’s ok, it’s natural.

If you only had that model to mail though, then you would have to continue to mail down that model until your performance dropped so low, that you’d question all your life decisions.

Don’t do that to yourself, you deserve better.

If you experiment with multiple models, once you notice one’s performance beginning to slump, you can easily pivot over to another one to maintain your campaign performance.

And the beauty is all Postie models come equipped with propensity scores baked in, so when it’s time to mail your dream prospecting campaign, there’s a data driven strategy on who to mail first, who to not wast your budget on, and when to pivot over to a new model of prospects ready to convert, NOW.

As always, if you have any questions, shoot them over, I love nerding out on this stuff.

Ok, see ya next month. ✌️