Latest / The CMO Podcast with Fexingo: Marketing Leadership, Budgets, and Executive Strategy / How CMOs Are Using AI for Real-Time Ad Fraud Detection
Transcript
- Lucas: There is a number that keeps me up at night if I’m a CMO with a big digital budget. This year, ad fraud is projected to cost advertisers over a hundred billion dollars globally. Luna: That is an enormous leak in the bucket. And I wonder how much of it goes undetected because legacy verification tools just aren't fast enough. Lucas: That's exactly the problem — and it's where a growing number of marketing leaders are turning to AI for real-time detection. Not after the campaign runs, but in the milliseconds before an ad is served. They're essentially building a fraud filter into the bid stream. Luna: So not a post-mortem audit, but a bouncer at the door. How does that actually work at scale? Lucas: Let's look at one specific implementation. A large retail media network — think of one of the top five US retailers with its own ad platform — deployed a machine learning model that scores every bid request for signs of invalid traffic. The model looks at over two hundred signals in real time: IP reputation, device fingerprint mismatches, mouse movement patterns that are too uniform, even the timing of clicks relative to ad load. Luna: Two hundred signals in real time — that's a lot of compute. Did it actually move the needle? Lucas: Yes, and the numbers are compelling. Within the first quarter, they reduced fraudulent impressions by 35 percent. But more interestingly, they saw a 12 percent lift in measured conversion rates because the traffic that remained was actually human and actually engaged. So the efficiency gain came from both sides: less waste on bots, better data for attribution. Luna: That attribution piece is huge. If your data is polluted by bot traffic, every model you build on top of it — attribution, frequency capping, lookalike audiences — is built on sand. Lucas: Exactly. And that's the argument a lot of CMOs are now making to their CFOs. The investment in ai powered fraud detection isn't just a cost center; it's a data quality investment that improves everything downstream. One CMO I spoke with told me that after cleaning their traffic, their customer acquisition cost dropped by 18 percent because they stopped bidding on impressions that would never convert. Luna: I can see how that pays for itself. But I imagine there's a tension: if you filter too aggressively, you might block legitimate traffic from new devices or unusual browsing patterns. Lucas: That's a real risk, and early models definitely had that problem. The approach now is to use a probabilistic scoring system rather than a binary block. Instead of saying 'this is fraud, deny it,' the model assigns a fraud probability score. The CMO can then set a threshold based on their risk tolerance. For a brand awareness campaign, you might accept a slightly higher score to maximize reach. For a direct response campaign where every dollar has to convert, you set a lower threshold. Luna: So it becomes a tuning dial, not an on-off switch. That feels more practical for marketing teams who have to balance efficiency with scale. Lucas: Right. And the best part is that these models get smarter over time. Fraudsters adapt — they rotate IPs, they simulate human behavior — so the model retrains continuously. Some platforms are now using adversarial AI, where one model tries to generate fraudulent traffic and another tries to catch it. That arms race is happening in milliseconds. Luna: I want to go back to that retail media network example. Did they build the model in-house or buy it? Lucas: They used a combination. They licensed a base model from a third-party verification vendor — the kind that has been doing post-bid analysis for years — and then fine-tuned it on their own historical traffic data. That way they got the benefit of the vendor's large-scale training data plus the specificity of their own audience patterns. The total investment was about 1.5 million dollars, which sounds like a lot, but when you consider they were losing tens of millions to fraud, it paid back in under two quarters. Luna: That's a strong ROI case. And I think it's a good reminder that not every marketing AI use case has to be about creative generation or personalization. Some of the most impactful applications are in the plumbing — the infrastructure that makes everything else work. Lucas: Absolutely. And you know, speaking of the show being useful and ad-free — if these marketing conversations have sparked something you've actually used, the reason we can keep doing this without ads is listener support. Luna: Yeah, it's a small thing that makes a big difference for us. If today's episode helped you think about your own ad spend, buying us a coffee is a great way to say thanks. Lucas: It's buy me a coffee dot com slash fexingo. That's it — no subscription, no perks, just a way to keep the conversation going. And we really appreciate it. Luna: Alright, back to the plumbing. You mentioned that fraud detection improves attribution — can you walk through that chain a bit more? Lucas: Sure. Imagine you run a campaign and your analytics show a thousand conversions. But two hundred of those came from bots that never actually visited your site — they just clicked on an ad and triggered a fake conversion event. If your attribution model gives credit to the last click, you're now giving credit to fake traffic. That skews your understanding of which channels work, which creative resonates, and even which audiences to target. Luna: So fraud doesn't just waste money, it distorts your strategy. Lucas: Exactly. A study from the Association of National Advertisers found that brands that implemented pre-bid fraud detection saw their measured return on ad spend improve by an average of 20 percent. Not because they spent less, but because the data they used to optimize was actually real. Luna: Twenty percent is a massive swing. And I imagine it also changes how you negotiate with publishers and ad networks. Lucas: It does. CMOs are starting to include fraud rate guarantees in their contracts — they'll say we won't pay for impressions that score above a certain fraud probability. And some are even asking for refunds when post-campaign audits show high fraud rates. That shifts the incentive for publishers to clean up their own inventory. Luna: I've heard that some large platforms have been slow to adopt this because it might reduce their reported inventory volume. Is that a factor? Lucas: There's definitely a tension. An ad network that filters out 30 percent of its traffic as suspicious suddenly has less inventory to sell. But the smart ones realize that quality inventory commands higher CPMs. If a publisher can prove that their traffic is 95 percent human and engaged, they can charge a premium. So it becomes a competitive advantage. Luna: Let's talk about the practical side for a CMO who's not at a retail media giant. What's the entry point? Lucas: The simplest step is to audit your current campaign performance through a fraud detection lens. Run a pilot with a verification vendor on one major campaign and see what percentage of your traffic is flagged as invalid. Most CMOs are shocked — numbers between 15 and 30 percent are common. Once you see the data, the business case for investing in real-time detection writes itself. Luna: And they don't have to build a model. Many demand-side platforms now offer built-in fraud scoring as an optional layer. Lucas: Right. The Trade Desk, for example, has its own fraud detection algorithms. Amazon Ads uses machine learning to filter invalid traffic. So even if you're buying programmatically, you can often toggle a fraud filter on. The key is to ask your platform partners exactly what they're doing and what metrics they report. Luna: I want to push on one thing: how do you distinguish between a sophisticated bot and a real human who just happens to behave like a bot? I'm thinking of people who use ad blockers or browse in private mode. Lucas: That's a great question, and it's the hardest problem in the space. Modern models use behavioral biometrics — things like how you move your mouse, how you scroll, the pressure patterns on a touch screen. Bots tend to have perfectly linear mouse movements or clicks that occur at exactly regular intervals. Humans have micro-hesitations, acceleration, and deceleration. Also, real humans typically have a longer session duration and a variety of page interactions. Luna: So it's not just about flagging a single click, it's about analyzing a session's worth of behavior. Lucas: Exactly. And the models are getting good enough that false positive rates are down to around 1 to 2 percent. That's low enough that the efficiency gain far outweighs the small amount of legitimate traffic you might block. Luna: One last angle: how does this intersect with privacy regulations? If you're analyzing user behavior in real time, does that raise any flags? Lucas: It's a valid concern, but most fraud detection doesn't rely on personally identifiable information. It looks at device fingerprints, IP addresses, and behavioral patterns — which are generally considered not PII under most regulations. The key is to ensure your vendor is transparent about their data practices and that you have proper consent mechanisms for any data that could be considered personal. Luna: So it's possible to fight fraud without compromising privacy. That's a good story to tell the legal team. Lucas: Absolutely. And I think that's where the opportunity lies for CMOs right now. The technology is mature enough to deploy, the ROI is clear, and the privacy landscape actually supports it. It's one of those rare moments where doing the right thing for your data quality also aligns with protecting user privacy. Luna: And it makes your marketing team smarter — better data, better decisions, better results. Lucas: Exactly. So if you're a marketing leader and you haven't looked at real-time fraud detection yet, start with an audit. You might be surprised at what you find — and how much you can save.