Why You Should Do Away With Fancy Models and Target the Old-Fashioned Way

Every cycle, down-ballot campaigns across the country get access to a new set of voter targeting models. They’re handed down from the party or the data vendor, and almost immediately they get adopted on the campaign as the main way to prioritize voters. 

There’s nothing wrong with them in general. They’re built to help you narrow down your list, and that’s something that needs to be done. The trouble starts when they get used ahead of hard signals, like which party’s primary a voter actually pulled a ballot in. 

Additionally, down-ballot candidates kind of need to do their own thing, prioritizing their own needs and building the organization around that. Unfortunately, these models were made for the top of the ticket, so a candidate in a D+10 district (whose actual problem is recruiting volunteers and training an organization) ends up calling the people a statewide campaign would want called for voter ID and persuasion. For the most part, that candidate (and the party) would get much more mileage out of pulling a list of everyone who voted in their party’s primary this year, thanking them for showing up, inviting them to an event and repeating that through Election Day. D+10 is closer to what I’d call a “long-term development” target and no model is going to tell you that. 

What’s Actually Inside These Things

I’ve used these pre-built models on campaigns and I’ve also built my own. 

Oftentimes, they use data beyond the voter file. This “aftermarket consumer data” can be very useful. It’s fair to say that a Guns & Ammo subscriber probably leans right and is pro-Second Amendment. There are also presence-of-children models built off kid-related purchases at Target and similar stores, which is useful information when targeting parents, even though it doesn’t point at either party. 

Beyond that, they often include things like property values, income, education level and the like. These have far less to do with party affiliation and they carry very little weight in the final score. When I work a score backwards to see what’s really driving it, I often find that primary participation is doing most of the heavy lifting. But, if primary participation is what’s driving the model, you could just as easily work with primary participation directly. It’s sitting right there in your file. 

The Old-Fashioned Way

Working with the voter file is easier than people think. We have access to primary vote participation data. So, if we know that a person votes in a Republican primary, it’s pretty easy to just treat them like they are a Republican. Same for voting in Democrat primaries. If they don’t show up for primaries at all, they’re more likely to be an independent. 

That basic principle gets you the right answer about the overwhelming majority of people in your file, and it uses the strongest partisan signal available in public data. A primary ballot is a decision the voter made, on the record, about which party they feel they’re a part of. Layer in general election turnout and you have your five buckets (Hard Rep, Soft Rep, Ind, Soft Dem,

Hard Dem), which you can use to target your voter list. This is good enough for the vast majority of the outreach your campaign will do. 

How This Changes Your Strategy

Before you GOTV or even persuade voters, you will still have to ID everybody in your independent universe anyway. So, a model telling you an individual voter is 62 percent likely to lean your way doesn’t save you the time of knocking, because 62 percent is not a firm ID. In a competitive race you need to reach those people regardless, and if you’re walking the block anyway, you might as well knock. 

What overlaying a model really does is let you cut down or expand your outreach list to match your budget. That’s genuinely useful in certain situations. If your independent universe is too big and expensive to mail, dropping the ones modeled as hard opposition is a perfectly reasonable way to make your final cuts. If you’re prioritizing volunteer or donor prospecting calls, you’ll probably want the most likely prospects at the top, and a model gives you a column to sort by. If you have some extra room in your budget, you can add soft supporters to expand your call universe. 

It Really Wasn’t Built for Your Campaign 

You shouldn’t be confident you’re following your best possible strategy if you’re using a model built for someone else. That’s particularly true if you’re more moderate (or extreme) than the average member of your party. 

A down-ballot candidate in a competitive district probably needs to talk to far more people in their district than the ones a statewide candidate might. 

The law of large numbers is what makes this work at a statewide scale. Miss some people, GOTV a few of the wrong ones by accident, and it comes out in the wash across millions of voters. For an individual candidate in a district of 50,000, it doesn’t work the same way. You have to be a lot more thorough than a model allows. 

A Word About AI 

Up until now, I haven’t been talking about LLMs and the things people are calling artificial intelligence today. I want to take a minute on that, because some vendors are going to start bolting AI into their offerings over the next couple of cycles. A few will have the LLMs ranking voters and applying scores directly.

I’ve tried it so you don’t have to. I gave Claude, Gemini and ChatGPT lists of voters, along with some of their aftermarket data, and asked each to score how likely those voters were to support us. What I didn’t tell them is that we’d already contacted a large number of those people and IDed them by hand, so I could test their answers. The results were abysmal. Close enough to random that it looked like the models were just making things up. 

Build Your Own Strategy 

So be careful over the next few cycles. Not every model is AI slop, and the traditional ones provided by the Republican National Committee and Democratic National Committee have some real practical applications for down-ballot candidates. But when something new shows up promising to score your file, ask what it was tested against and consider testing it yourself before you build your whole program on top of it. 

So, open your file. Build out your outreach buckets with your own hands using primary history and turnout.

Caitlin Huxley bridges data and strategy to help moderates win close races, leveraging 15 years of experience. She’s the author of Ancient Wisdom for Modern Campaigns: Lessons from Sun Tzu’s Art of War.