The world of political advertising is still a little old school.
First off, their fundraising methods still seem to be heavily reliant on generalized calls and texts. At least, that’s my assumption based on the number of desperate messages that politicians – and sometimes even their dogs – have sent me.
TV is still the most popular ad channel among politicians, surpassing more contemporary options like social or podcasting, according to Vai Gupta, chief product officer of IQM, a DSP and DMP that works specifically with advertisers in highly regulated markets. (That said, a large portion of the TV ads are CTV; it’s not all retro.)
One barrier to adoption for political advertisers is the breadth of restrictions and regulations that they’re beholden to, like campaign disclaimers and financial disclosures. TikTok, for instance, allows no paid political content. Users are free to share their own beliefs, so long as they “remember not to mix it with paid ad products like Spark Ads and Promote,” per its Creator guidelines.
Still, despite the challenges, digital media is winning out and tipping the scales based primarily on targeting capabilities that reduce waste and can vastly increase reach with the right voter segments.
At least, that’s the pitch from IQM.
Building out granular political segments, simply put, “is hard,” said Gupta. On Wednesday, IQM launched a new tool called Custom Voter Audience meant to bring greater targetability to political spend.
If you build it
As the name suggests, Custom Voter Audiences lets political advertisers more easily target certain swaths of the voting population.
Under the previous system, developing custom audiences required manually checking and unchecking boxes in a massive data file to build a specific segment – say, registered Republicans who have moved in the last six months and voted in the past three cycles.
Now, with the Custom Voter Audiences tool, which is an AI agent built for the use case, advertisers can query the data directly in plain language and churn out the same results much more quickly. Claude is “the backbone” of the tool, chosen for the strength of its search mechanism, Gupta said.
Tinkering within spreadsheets to find the right targeting parameters and scale is exactly the type of task at which an LLM – forgive the pun – excels.
All of the models are enterprise-grade and run in house, so there’s no chance of data leakage, Gupta added. The LLM effectively converts the user’s audience restrictions into a query, which it then migrates to a data clean room where the audience is built and assigned an anonymous identifier. The actual data never goes into the LLM itself, said Gupta, but rather lives solely in the clean room.
Nobody’s perfect?
IQM partners with a number of political data providers, including i360, Aristotle and L2, which aggregate and sell household-level voter data to advertisers. The political data brokers build their data sets from voter files and registrations, past campaign or election participation records and, sometimes, info like previous donations.
Along with the voter data, IQM brings in more general data sources, such as LiveRamp audiences. And having this large buffet of providers is important for meeting custom audience requests, Gupta said.
For instance, a politician might want to reach Ivy League graduates who are passionate about clean energy solutions. Through Custom Voter Audiences, an advertiser could build a model that finds the intersection of the two categories – pulled from existing third-party segments – and creates a new audience to go out and target programmatically.
But in spite of all of the guardrails, the AI can still make mistakes, Gupta noted. Sometimes, it generates inaccurate assumptions about which traits correlate with others. He noted one recent campaign in which the AI predicted that older individuals who are highly educated would have strong opinions on whether or not to build data centers. In reality, he said, the people most opinionated about data centers were those living in the nearby neighborhoods.
IQM’s human advisors caught and corrected the mistake, he said.
The tool has been in beta testing for about a month with IQM’s top five customers, he said, and the back-end team has been continually adjusting guardrails to determine more accurate audience correlations.
Now that the tool is in the hands of IQM’s wider audience, it’s more important than ever to put a stop to AI hallucinations before they start.
