Mad Science: Can AI + adtech 3x your homepage visits?

Our contextual AI test shows how enterprise advertising power can be put into the hands of SMBs.

The democratization of CTV technology has changed the game for small and local businesses, who now share ad space with the biggest, most storied national brands. But advertising can still be difficult for companies without major resources.

At Madhive, AI is always on our mind, and lately we've been wondering how local businesses — and the agencies and media sellers that support them — can get a bigger bang for their buck on CTV.

So we asked our innovation team to use AI to solve a data and efficiency challenge that scares many SMBs from advertising — specifically, how smaller businesses with very low amounts of data can feel confident that they’re hitting the right households for their product.

And like a true "Mad Scientist," we decided to test the AI model on ourselves first.

Read on to discover:

  • How we trained an AI to support smaller budgets 🤖💰
  • The results we achieved with our own CTV campaign 📊📺
  • How a local auto dealership is succeeding with our test 🚗🏆
  • What this could mean for the future of adtech 🔮🚀

The problem: Small budgets must hit the right households

CTV is a powerful way to advertise because it's engaging and emotionally impactful. But since viewers are passively watching, getting them to take action means advertisers need to repeat their message until it sticks.

The tricky part is figuring out which households to target. Big brands have tons of data to do this, but small businesses have always struggled without that same audience insight.

While tools like geotargeting, third-party data, and campaign optimization help, we saw an opportunity with AI to take targeting to the next level and give SMBs more confidence in where their budget goes.

The test

Last spring, we launched our own B2B CTV campaign so we could share real-time insights about targeting a very tricky audience — media pros and B2B marketers.

Our goal? To show off CTV's performance potential, and drive viewers to visit the Madhive website.

It was the perfect chance to test two new AI models we casually referred to as Lookalike+ and BlankSlate.

Both were designed to find smarter audience segments, focusing on households with a higher chance of converting — helping us optimize our spend and make every dollar count.

Here's how everything shook out.

Step 1: Using contextual data to make lookalikes smarter  

The takeaway

  • Traditional lookalike models struggle with distinguishing intent and timing.
  • Lookalike+ analyzed both on-site and external behavioral data to find high-intent households.
  • AI audiences delivered 62% higher CVR than third-party data audiences and a 38% lower CPA.

We started with a familiar approach: lookalike modeling. Typically, this method identifies potential customers by finding people with similar behaviors and demographics to existing customers.

However, traditional lookalike models miss a key factor: intent. They can't tell if someone is just an enthusiast or an actual buyer ready to act.

That’s why we developed Lookalike+, an AI model designed to analyze deeper signals, like online behaviors and interactions, to pinpoint people who show real buying intent.

For our CTV campaign, the AI tracked prior converters and examined external behaviors, like social media and search activity, to spot patterns suggesting they were ready to buy.

Unlike standard lookalike models that group users by static attributes, Lookalike+ evaluates individual households in real time, determining if they're a fit for our product right now. It’s a smarter, more dynamic way to target intent-driven audiences.

Lookalike+ results: Higher conversion, lower CPAs

After four months of testing Lookalike+ with our campaign, the results were telling.

We compared our AI audience to two other buckets: audiences created with first party data as well as third party data. While 3P-generated audiences provided the highest scale, and 1P-generated audiences provided the highest conversion, AI offered the best of both worlds.

  • The AI averaged a reach of about 75,000 uniques per week, which is 8x the size of our 1P reach.
  • The AI conversion rate was .026%, compared to the .014% average for the entire campaign. Unsurprisingly, this CVR was lower than our 1P bucket, but it was 62% higher than the CVR for the 3P bucket.
  • CPA for our AI was also 38% lower than human-generated segments, meaning the AI was more efficiently converting people and we were paying less per customer with those households.

Step 2: Solving a frequency problem  

The takeaway

  • Creating personalized media plans for each household is complex, but even standard retargeting improved conversion efficiency.
  • CVR went from one per 3,000 impressions to one per 1,000 impressions.

We knew finding the right households was just the start; we also needed to effectively saturate our message too.

Our goal was to use AI to create personalized media plans tailored to each household's unique contextual data. However, this proved more challenging than we anticipated.

Crafting a unique customer journey for every household is complex, given the numerous publishers and endless combinations of timing and frequency. Recognizing this, we decided to pause the personalized plans and switch to standard retargeting for households identified by our AI.

This pivot made our model much more efficient at converting users.

With retargeting, our AI improved its conversion rate from one for every 3,000 impressions to one for every 1,000. More ad exposure was crucial for achieving better CPA with our AI audiences.

While our personalized media plans still need testing, we're committed to finding a solution. As digital marketing evolves away from a "one-size-fits-all" approach, CTV needs to get smarter, and delivering tailored experiences at scale relies on advanced AI.

Step 3: Creating an audience with no data at all

The takeaway

  • Building on our initial AI success, we developed a model to identify potential customers without any prior audience data.
  • This model performed comparably or better than Lookalike+, achieving a 14% lower CPA in tests.

At this point, we set out to tackle an even bigger challenge: identifying potential customers without any prior audience data.

What if you knew absolutely nothing about your audience, beyond assumptions or anecdotal evidence? This is a scenario that many small businesses know all too well.

So we created a model that doesn't rely on previous converters, which we called BlankSlate. The model was designed to analyze our product and its market to find new customers.

Surprisingly, this no-seed model often performed just as well or even better than our previous model, Lookalike+. In fact, BlankSlate had a CPA that was 14% lower in tests, proving it could identify the right customers without relying on past converters.

This offers huge potential for both small and large brands alike. With BlankSlate, you don't need to have a richly defined profile of your potential customers or market fit. You also don't need to be limited to just the core understanding of your customers, either.

Ultimately, this allows advertisers to move past biases and assumptions, allowing for more effective audience targeting.

Step 4: Bringing Lookalike+ to our clients

Now it was time to bring our AI model to our clients, the broadcasters and agencies serving local advertisers and SMBs.

These advertisers need to deliver results in the most efficient way possible – both in terms of the budget spent, but also the internal man hours required to hit their goals.  

A local car dealership boosts lower funnel actions

One of our first test clients was Voice Media Group, and their customer Valley Chevy, a local car dealership. Valley Chevy was primarily interested in homepage visits, and lower funnel actions that suggest deeper consideration.

The AI scored for the auto dealership, resulting in 3.5x more homepage visits than the baseline, showcasing its ability to attract potential customers. Lower funnel actions, such as vehicle comparisons and dealership location searches, also saw notable improvements at 1.3x and 1.7x, respectively.

But the best outcome of all is a happy customer; Valley Chevy decided to continue using Lookalike+ and even increased their spend for AI.

Looking ahead: How AI can change local advertising

So what does all of this mean for adtech? How can AI actually change how we do business? Here's a rundown of what we learned from our test:

  • AI-driven models could redefine advertising strategies, especially for businesses with tighter margins and low volumes of data.
  • Advertisers may increasingly adopt AI for customer acquisition and conversion optimization, making CTV a more effective performance marketing channel.
  • AI can help companies adjust in real-time to changing market conditions and target audiences effectively.

Ultimately, we're most excited about how this can make an impact for local advertisers looking to get a stronger hold in the CTV advertising game.

If you're interested in helping us test Lookalike+ or BlankSlate for your CTV campaigns, click here to get in touch.