Episode 54: Real-Time Propensity Modeling: Reclaiming Ad Performance in a Privacy-First Era

How do you optimize your digital ad platforms when strict healthcare privacy laws block your tracking pixels?

In this episode of Digital Clinic, Aaron Burnett sits down with Michael Wiegand, Director of Marketing Sciences at Wheelhouse DMG. Together, they explore how real-time propensity modeling enables HIPAA-compliant, value-based bidding on platforms like Google and Meta. Michael shares how to bypass slow retroactive audience modeling, protect patient privacy, and feed real-time prediction scores back to your ad networks so you can optimize for higher quality leads.

Listen & Subscribe:

Introducing Michael Wiegand 

Aaron Burnett: We all know that digital marketing for healthcare and med tech can be truly challenging. The combination of data privacy laws and platform restrictions mean you can’t target the audiences you need to reach in the way other industries can. And many of these same privacy regulations also mean that advertisers can’t use the kind of ad platform tracking that returns rich conversion signals, the very data required for today’s algorithmically-driven campaign optimization. 

And compounding this challenge is the fact that for most healthcare and med tech organizations, the action they truly want, the thing that actually creates value, a patient attending an appointment, a new customer start, may not happen for months after the type of conversion signal typically optimized for in digital marketing. 

It all sounds insurmountable, but it’s not, and today’s guest will tell you exactly how he solved for these challenges, creating clear line of sight to revenue-generating signals and unlocking phenomenal performance in the process. 

Our guest today is Michael Wiegand, Director of Marketing Sciences at Wheelhouse DMG. 

With more than twenty years’ experience developing and implementing some of the most advanced digital marketing strategies in the industry, he’s an expert’s expert. 

In today’s episode, he’ll share exactly how he solved the unsolvable and developed a system that can be used by anyone to unlock performance for privacy-first clients. 

Let’s get to it. 

A Message from the Sponsor  

Aaron Burnett: This podcast is sponsored by Wheelhouse Digital Marketing Group. Wheelhouse provides exceptional performance marketing for healthcare and medical device manufacturers. Every Wheelhouse client saw record performance in 2025 even after implementing HIPAA-compliant data solutions. Find out more at wheelhousedmg.com.  

Real-Time Propensity Modeling 

Aaron Burnett: So let’s turn our attention to the topic of the day: propensity modeling. Most marketers are familiar with propensity modeling, but what we’re doing is different; it’s real-time propensity modeling. Let’s start with how propensity modeling is typically used by digital marketers. 

What do people usually think about when they think about propensity modeling? 

Michael Wiegand: I think the challenge in describing this to both clients and our internal team is that we often conflate propensity modeling with lead scoring. I think they serve two distinct purposes. Propensity modeling at its core is trying to understand who is going to convert and their likelihood to convert based on signals that you know about them when they become a lead. 

That sounds a lot like lead scoring. But what that ends up doing is limiting you. One is a function of how you want to direct your sales team’s energy, and the other is a function of how you want to send signals back to your advertising platforms to get smarter about finding the next best lead. 

I think that latter piece is what we’re trying to accomplish with propensity modeling. 

Aaron Burnett: I know one of the other things we’ve run into is that propensity modeling for a lot of folks, particularly in privacy-first industries, is a slow-cycle exercise used to model audiences that you might want to target. You go to a third-party audience provider and say, “These are the characteristics of the audience we want to reach.” 

You target that audience, find out how it worked, and then remodel and go after the audience again. It’s a much slower cycle. What we’re doing is immediate; it is real-time. So, let’s cut to the why: what is real-time about what we’re doing? 

Michael Wiegand: Yeah. So that’s a great question. 

Sending Signals to Ad Platforms 

Michael Wiegand: When we speak about real-time, we mean that as a user or prospective patient interacts with a lead form on your website, we are running formulas in the background to calculate their likelihood to convert based on their inputs. As soon as they hit submit, we’re turning that into an aggregated score that goes to the ad platforms to tell them how likely that user is to convert and their quality based on value. That is the real-time aspect: people input values, we run them against a formula, and we pass a score to the ad platforms. 

Aaron Burnett: Right. Okay. All right. So many questions derive from that. One is, all right, how do we get a score, this value? How is that score developed with our clients?  

Using CRM Data 

Michael Wiegand: We build a model comprised of elements captured through the lead form, such as age, gender, their current treatment method, and specificity around their condition. 

We gather that information and create a formula where we stack-rank the values of each component. Then we derive a coefficient or multiplier that allows us to coalesce all of that information into a single score. 

We aren’t passing the ad platform data about specific conditions. We do a calculation on the front end to see how a condition affects their conversion rate, and then we boil that down into an anonymized score. 

Aaron Burnett: How do we know how to characterize and value what’s captured in that lead form? 

Michael Wiegand: We typically run a regression analysis based on past leads. Our client gives us access to their Salesforce information, which contains anonymized information about patients who submitted lead forms in the past and the business outcome for those patients. 

We see if they became a qualified lead, a customer start, or delivered revenue. Based on that, we develop the model using those factors to see how often they translate to a qualified lead or revenue. 

Typically, we’ll do that with a look-back of at least a year’s worth of data, but we try to go back as far as we have data to develop that model. 

Aaron Burnett: Sure. Okay. All right. I’m going to repeat back then what I think I understand. We look back at least a year’s worth of historical data, and we evaluate that data using a series of analyses, regression and statistical analyses, to determine the factors that are predictive of a patient or customer who generates revenue. And then we go further and figure out: what’s the lifetime value of this patient or customer? And we use that data to develop a propensity score. Is that right? 

Michael Wiegand: That is correct. We find that not all factors gathered in the lead form are created equal. Differences in age or gender may not be big determining factors for some clients; it might be things like the type of insurance or the specific subset of the condition they have. 

So it’s really instructive in which factors in the model we should pay attention to at the time of the lead, that really helps us inform the score.  

What Score Gets Sent 

Aaron Burnett: So then we developed this propensity score. What’s the actual value that we’re passing to the platforms? Is it a number, one through 100? Is it something else? 

Michael Wiegand: In cases where we can attribute a lifetime value to a customer, we start there and use a reductive process to figure out how much that particular lead is worth at the time of submission. For example, if the lifetime value is $35,000, we use that as the base of the model and then use different factors to understand the potential value of that lead. 

In other cases, we build a propensity model that returns a binary zero or one: this person is or is not likely to convert. We use those in tandem, sending either the binary value or the score where we have access to revenue data.  

Privacy and HIPAA Compliance 

Aaron Burnett: We’re working in privacy-first industries. How is this compliant with regard to data privacy when sending values back to platforms? 

Michael Wiegand: The first thing to tackle is that even when clients take the risk of putting a client-side advertising pixel on their website, they should only send PHI back if they have a prescribed combination of information. 

That most often manifests as a condition and an IP address. Those two will get prospective clients in trouble if sent back to an ad platform. 

Aaron Burnett: Along with the name, email address, phone number, zip code. 

Michael Wiegand: Yeah. All of that kind of stuff. So anything that constitutes a way to identify the person, plus the condition they have. 

In the case of the propensity model, we make those determinations with a formula and boil everything down to a score before sending it out. We aren’t telling the platform what condition that person has, only the score and value. We do all of that using local and session storage on the individual’s machine. 

That storage gets flushed when they close their browser. We aren’t actually holding onto that information in a cookie or anything the pixels can read. We calculate the score using local storage, send the score out, and the information is flushed as soon as they leave the website. 

Aaron Burnett: Okay. So we’re not sharing anything that’s even adjacent to PHI. We share a single number. It’s a dollar value or a numerical score. We don’t pass anything else related to that prospect or patient. 

Michael Wiegand: Precisely.  

Why Platforms Need Value Signals 

Aaron Burnett: Okay, so a couple of things. Why is it important? Why is it valuable to send a propensity score? Why is this such a critical mechanism for us? 

Michael Wiegand: If you aren’t sending a propensity score, the paid platform algorithms will try to find how many people just fill out your lead form in general. That, divorced from any sort of value, doesn’t speak to the quality of the leads that Google and Meta can find. 

They may deliver a lot of lead form fills, but those leads might be poor quality overall and not worth your sales team’s time. This propensity score helps us understand the potential value of the person immediately after they fill out the form so we can optimize for leads worth pursuing. 

It also allows us to do maximum value bidding and return on ad spend (ROAS) bidding. Since we’re sending a dollar value, we can bid to a ROAS figure, which is very powerful. 

Aaron Burnett: All right, so you mentioned ad platforms. This is an important topic. Talk to me about the way ad platforms currently optimize. What’s different today about how ad platforms might have optimized two or three years ago? And why is a propensity score or some signal important in that context? 

Michael Wiegand: The way they bid has changed so much. They’re using more AI signals to inform their algorithm, but many of the factors feeding that algorithm are signals they have about users based on other sites in the ad network. 

Other things they intrinsically know about the user who might be logged into their platform, like Meta. And so, divorced from any sort of propensity model signal, they’re making all kinds of assumptions about what they think is important to someone who fills out a lead form on your site. 

They make assumptions based on demographic and psychographic information rather than the specific health condition, current treatment, or insurance type. The propensity score we’re building incorporates those signals that the ad platforms don’t have out of the box. 

Aaron Burnett: Typically, a marketing agency, an in-house advertising team, is going to optimize to the thing that can be most easily measured by the ad platform, by analytics. We’re doing something different. What we’re optimizing for can’t actually be measured by an analytics platform or even by an ad platform. 

Describe the difference between this approach and the approach that’s enabled by just the tools that you get from ad platforms and analytics. 

Michael Wiegand: Out of the box, ad platforms are incentivized to have you buy more clicks at a higher cost per click. That moves the needle for their shareholders, but it often runs counter to what you are trying to do as an advertiser. 

You’re trying to get higher-quality leads and optimize the cost to get them. Default tools are designed to maximize clicks and conversions, but not necessarily quality or efficiency. Our propensity model tries to put that power back in the hands of the marketer. 

Aaron Burnett: The other critical thing we’re solving for, which seemed to be intractable, is that the moment of actual business value delivery, the real thing you’re trying to optimize for, is a new customer start or a patient who actually shows up for an appointment and becomes a patient of your hospital system or your clinic or that sort of thing. 

And there is a significant elapsed time between the conversion event that an ad platform or analytics measures, someone completed a form, and the moment of value creation as well. Talk a little bit about that and, in particular, maybe talk about how extreme that can be in a medical device context. 

Michael Wiegand: In our research, we’ve found that the time to business value can be upwards of 180 days, sometimes a year or more, depending on the device and insurance hurdles. When communicating value to an ad platform, you have a limited window to do that. 

Typically, you only have about 90 days to send back offline signals from the time an ad click happens. If your sales cycle is 180 days, you can only send so much back. The propensity model accurately predicts business value at the time of lead creation within a tolerable margin of error. 

That signal we send at lead creation is close to the business value we’ll realize in that year-long sales cycle, allowing us to predict the future so we’re not worried about the signal over time. 

Implementation: Server-Side Tagging 

Aaron Burnett: We’ve skipped over the actual mechanics. How do we operationalize this? Once we develop a propensity score, where and how is it fed back to the ad platforms? 

Michael Wiegand: Regression modeling and testing can happen in a spreadsheet or through Python analyses to generate the formula applied to each factor. We put the formula into action using a tag management solution. 

We use Google Tag Manager or Tealium to read the user’s input on the lead form and apply the formula in real-time to build the score. 

We perform the calculations on local or session storage and then egress that score to the ad platform along with your conversion pixel when the user hits the thank-you page. 

Aaron Burnett: Right. And we’re doing all of that server side. So in the case of Tealium, we’re going to use their event data framework. You can do this using Tag Manager, but there’s an important distinction here around using Tag Manager. You don’t want to use Tag Manager in the default way that Google describes. 

Describe how we use Tag Manager in a privacy-first context versus the default implementation. 

Michael Wiegand: We stand up an instance of server-side Google Tag Manager on cloud infrastructure that our client controls, such as AWS or Azure. This private instance becomes a resting point for data before it is sent to ad platforms. 

This allows us to cleanse anything that would constitute PHI. We ensure we aren’t sending out IP addresses or URLs that could reveal a user’s health condition. We redact and anonymize all of that information before it reaches the ad platform. 

Aaron Burnett: Yeah. So we end up, and this is not well known or well understood, from what we can tell, using Google Tag Manager as an appliance. We’re not using it on Google Cloud infrastructure, so there’s zero data sharing with Google. We have our own data pipeline that is consistent with the infrastructure on which we deploy Google Tag Manager as an appliance, and that makes it HIPAA compliant. 

Michael Wiegand: That is correct. 

The Takeaway for Healthcare and MedTech Marketers 

Aaron Burnett: All right, so let’s talk about results. We have implemented propensity modeling for clients. Give me a sense for how well it’s worked and the kind of results that we’ve driven. 

Michael Wiegand: We have a case study on the Wheelhouse website that I encourage listeners to check out. We’re seeing a lot of success with value-based bidding against the propensity model score in Performance Max (PMax) campaigns. 

We’re seeing that Google can algorithmically reach clients with a number of different ad formats, and that marries really well with the propensity score. So we can meet people at the quality of the lead, but also meet people where they’re at and most likely to look for us on the ad platforms. 

Marrying the propensity score with Google’s ability to reach clients with different ad formats works really well. We found success with PMax and Advantage+ Shopping Campaigns (ASC) on Meta, where there is great synergy between the value score and the ad formats. 

I think that’s where we’re finding the biggest pockets of success with this sort of approach. 

Aaron Burnett: And so I know we implemented for a large med tech client, and over the course of about six months, we increased their new customer starts by about 50% using propensity modeling, which is non-trivial, particularly at scale. 

Michael Wiegand: Yes, absolutely. It’s important to test this approach. We recommend setting up an A/B test running half the campaigns on a max-conversions model versus a value model to understand where the signal is most valuable. 

Aaron Burnett: Is there a future propensity modeling that you want to hold forth on? 

Michael Wiegand: Regarding the future, modeling like Marketing Mix Modeling (MMM) will be critical for evaluating incremental lift. Better leads often convert immediately, but for longer sales cycles, you’ll want to implement MMM alongside this to corroborate results and ensure statistical modeling holds weight against the entire program. 

In a lot of cases, you will get better leads and they will convert immediately to customers. But I think the long game, and how you realize actual results, especially for those longer sales cycles, you’re going to want to implement something like MMM alongside this to corroborate those results, to make sure the back testing you’ve done and the statistical modeling holds weight once you’re running this up against an entire program. 

You might be sending these scores back to a handful of platforms and not to others. You may be sending it to all platforms, and it’s important to understand how each of those platforms and campaigns aligns into a business result for your clients. It might not just be one ad that they interacted with that drove the result. 

It’s a multi-touch process, and that will be the real factor in understanding the impact of the score over time. 

Aaron Burnett: That’s very well said. It’s been a great conversation. I appreciate it. 

Michael Wiegand: Cool. No problem. 

Aaron Burnett: Real-time propensity modeling solves for the critical gap between the need for immediate signal for optimization and the reality that the actual moment of conversion may take months, and it does this in a privacy-compliant way. We know it works. We’ve got the receipts, and we’re excited to share the approach with you today. 

Don’t hesitate to reach out if you’d like to talk or need more detailed information on what we shared with you. And if you enjoyed today’s discussion, be sure to subscribe to the podcast. Leave us a review and share this episode with anyone on your team looking to build smarter, fully compliant marketing campaigns. 

I’m Aaron Burnett, and I’ll see you next time on The Digital Clinic. 

Sponsored by Wheelhouse DMG