Google deploys Gemini AI to combat ad fraud with 40% reduction

Google's advertising division quietly implemented artificial intelligence tools powered by large language models to detect and eliminate invalid traffic, achieving substantial improvements in ad fraud prevention according to an announcement made on August 12, 2025.

According to researchers at Google, the company has been testing multimodal large language models for approximately 18 months to identify deceptive and disruptive advertising techniques across mobile applications and web platforms. The effort has led to measurable reductions in invalid traffic since late 2023.

The announcement revealed that Google's ad traffic quality division collaborated with experts from Google Research and Google DeepMind to develop systems capable of navigating websites and mobile applications while mimicking human user behavior. These AI systems can understand application functions, interact with navigation tools, observe advertisements, capture screenshots, and identify deceptive practices by examining rendering instructions.

Technical Implementation Details

The Gemini-powered detection system operates by identifying instances where advertisement buttons should be visible but are not displayed to users. Paired with machine learning tools and traditional backend data processing, the technology can detect accidental clicks, hidden advertisements that register impressions without user visibility, and disruptive out-of-context advertisements that force unwanted user interactions.

According to Per Bjorke, Google's director of product management for ad traffic quality, the system routes detected problematic ad experiences through human review processes to ensure accurate policy enforcement. "If you [as an advertiser] spend $100,000, and $10,000 goes to invalid traffic, that's $10,000 wasted that you could have spent on actually acquiring users," Bjorke explained to ADWEEK.

A pilot program conducted between December 2023 and October 2024 demonstrated the effectiveness of these AI-powered detection methods. The pilot resulted in a 40% reduction in mobile invalid traffic stemming from what Google classifies as "deceptive or disruptive" advertisements, which includes various hidden advertisements and unexpected pop-up advertisements that violate placement policies.

Though the process has not reached full automation, Google reports that the pilot program improved both speed and accuracy of enforcement compared to previous detection methods.

Industry Context and Scope

Invalid traffic represents a significant challenge across the digital advertising ecosystem. Research conducted by Pixalate during the first quarter of 2025 found that global invalid traffic rates reached 18% on web platforms and 31% on mobile applications, based on evaluation of more than 100 billion programmatic impressions worldwide.

This problem extends beyond financial waste for advertisers. Bjorke highlighted that invalid traffic also damages publishers, as "any dollar paid to bad actors is a dollar that should have gone to a good publisher." The issue affects end users as well, with Google receiving complaints about emergency services being disrupted by pop-up advertisements, including one instance where a user's attempt to dial 911 was interrupted.

The company has documented multiple cases of sophisticated fraud operations targeting its platforms. Google battled significant ad fraud schemes throughout 2025, including the removal of 352 applications associated with an operation that generated 1.5 billion daily bid requests at its peak. In February, the company eliminated nearly 200 applications from the Play Store involved in a separate fraud operation, and in July announced legal action against China-based hackers over a fraud operation that infected more than 10 million Android devices with malware.

Performance Improvements and Future Outlook

Google's ad safety team published findings last week indicating they achieved a 10,000-fold reduction in the amount of AI training data needed to identify policy-violating advertisements, suggesting their detection techniques are becoming increasingly efficient.

The technology builds upon Google's existing sophisticated multi-layered systems for identifying and filtering invalid activity. The company employs both General Invalid Traffic (GIVT) and Sophisticated Invalid Traffic (SIVT) detection techniques to comply with Media Rating Council guidelines.

Bjorke emphasized the ongoing nature of the challenge, describing it as "always an adversarial kind of 'tit for tat' type of game" where advances in detection prompt behavioral changes from bad actors. However, he noted that AI tools can respond more quickly to these developments and catch fraudulent operations faster than traditional data analysis methods.

"It's not guaranteed foolproof to anyone bypassing it, but compared to data analysis, this is more robust in these types of use cases," Bjorke said.

Impact on Digital Advertising Ecosystem

The implementation reflects broader industry concerns about invalid traffic affecting advertising budgets and platform integrity. European anti-fraud programmes prevented €3.45 billion in losses during 2023, demonstrating the global scale of fraud prevention efforts.

Google's announcement comes during a period of increased scrutiny regarding ad fraud detection capabilities. Recent investigations revealed systematic failures in bot detection across multiple verification companies, highlighting the need for improved detection technologies.

The multimodal approach represents a departure from traditional detection methods that relied primarily on data pattern analysis. By incorporating visual understanding and behavioral mimicry, the system can identify deceptive practices that might evade conventional detection algorithms.

Regulatory and Market Implications

Google's investment in AI-powered fraud detection aligns with broader regulatory pressures and industry standards. The Media Rating Council has established comprehensive standards for identifying and filtering invalid traffic, requiring accredited companies to demonstrate effective detection capabilities through independent audits.

The company's announcement on the @adsliaison Twitter account on August 12, 2025, emphasized their collaboration with research divisions to develop "LLM-powered solutions that allow us to more precisely identify ad placements that are generating interactions from non-genuine users."

According to Google's official blog post, published simultaneously with the announcement, invalid traffic "wastes ad budgets, siphons revenue from publishers and erodes trust." The company positioned their new AI applications as providing "faster and stronger protections by analyzing app and web content, ad placements and user interactions."

The blog post noted that beyond the 40% reduction in problematic practices, the company continues running "extensive automated and manual checks to ensure advertisers aren't charged for IVT, even if an ad serves."

Google expressed confidence that AI investments will continue proving effective against invalid traffic and fraud operations. The company framed their detection capabilities as part of "two decades of work to defend against threats and uphold the integrity of the digital ad ecosystem."

For marketing professionals utilizing Google's advertising platforms, these developments suggest improved protection against wasted ad spending and enhanced confidence in campaign performance metrics. The implementation represents significant technical advancement in combating fraudulent activity that has plagued digital advertising since its inception.

The announcement demonstrates how large language models are expanding beyond content generation into practical applications for platform security and user protection. As advertising fraud schemes become increasingly sophisticated, particularly with AI-generated deception techniques, Google's multimodal detection approach may establish new industry standards for fraud prevention.

Timeline

• May 2017: IAB Tech Lab launches ads.txt standard to prevent unauthorized advertisement inventory sales
• Late 2018: DoubleVerify identifies first major ads.txt exploitation scheme
• December 2023: Google announces Gemini AI model as most capable multimodal system
• December 2023 - October 2024: Google conducts pilot program testing AI-powered invalid traffic detection
• February 2025: Google removes nearly 200 apps from Play Store involved in ad fraud operation
• March 2025: Adalytics report reveals systemic failures in bot detection across digital advertising ecosystem
• July 2025: Google announces changes to MRC accredited metrics affecting invalid impression counting
• July 2025: Google announces legal action against China-based hackers over fraud operation affecting 10+ million Android devices
• August 12, 2025: Google publicly announces AI-powered invalid traffic detection results achieving 40% reduction in deceptive practices

PPC Land explains

Invalid Traffic (IVT): Invalid traffic encompasses any advertising activity that fails to originate from genuine human users with authentic interest in the advertised content. This broad category includes both intentional fraud schemes and unintentional issues such as accidental clicks, bot-generated interactions, and automated crawling activity. According to industry standards established by the Media Rating Council, IVT represents traffic that does not meet quality or completeness criteria for legitimate advertising measurement. The classification system divides invalid traffic into General Invalid Traffic (GIVT), which includes routine bot activity and known crawlers, and Sophisticated Invalid Traffic (SIVT), encompassing more complex fraud schemes designed to evade detection systems.

Multimodal Large Language Models: These artificial intelligence systems represent advanced machine learning architectures capable of processing and understanding multiple types of data simultaneously, including text, images, video, audio, and code. Google's implementation in fraud detection leverages these models' ability to navigate websites and mobile applications while mimicking human behavior patterns. The multimodal approach enables the system to analyze visual elements, understand user interface components, and identify discrepancies between intended and actual advertisement presentation. This comprehensive analysis capability allows detection of sophisticated deception techniques that traditional single-modality systems might miss.

Ad Traffic Quality: This specialized division within Google's advertising operations focuses on maintaining the integrity and authenticity of advertising interactions across the company's platforms. The team employs various detection methods, human review processes, and automated systems to identify and eliminate fraudulent activity before it impacts advertisers or publishers. Their responsibilities include developing new detection technologies, collaborating with industry organizations on standards development, and ensuring compliance with Media Rating Council guidelines. The division works closely with Google Research and Google DeepMind to implement cutting-edge technologies for fraud prevention.

Deceptive or Disruptive Advertisements: This category encompasses advertising implementations that violate platform policies through misleading presentation or forced user interaction. Deceptive advertisements include hidden ads that register impressions without user visibility, misleading button placements, and advertisements disguised as legitimate content elements. Disruptive advertisements force unwanted user interactions through unexpected pop-ups, interference with navigation, or obstruction of intended user actions. These practices not only waste advertiser budgets but also create negative user experiences that can damage platform credibility and user trust.

Detection System: Google's comprehensive fraud prevention infrastructure combines artificial intelligence, machine learning algorithms, and human review processes to identify and eliminate invalid traffic. The system operates through multiple layers, beginning with automated screening using pattern recognition and behavioral analysis. Suspected violations undergo additional AI-powered evaluation using multimodal models that can assess visual elements and user interaction patterns. Finally, human reviewers examine flagged content to ensure accurate policy enforcement and minimize false positive detections while maintaining system effectiveness.

Programmatic Advertising: This automated approach to digital advertising utilizes real-time bidding systems and algorithmic decision-making to purchase and place advertisements across websites and mobile applications. The programmatic ecosystem involves multiple participants including advertisers, publishers, supply-side platforms, demand-side platforms, and ad exchanges. Invalid traffic particularly impacts programmatic advertising because automated systems may not detect sophisticated fraud schemes that human oversight might identify. The scale and speed of programmatic transactions create opportunities for fraudulent actors to exploit system vulnerabilities.

Bot Traffic: Automated software programs designed to simulate human web browsing behavior for various purposes, including legitimate functions like search engine crawling and malicious activities such as click fraud. Sophisticated bot networks can mimic genuine user interactions through randomized clicking patterns, varied IP addresses, and realistic device characteristics. These systems pose significant challenges for fraud detection because they continuously evolve to circumvent security measures. Bot traffic can artificially inflate advertisement metrics, waste advertiser budgets, and distort campaign performance analysis.

Mobile Applications: Software programs designed for smartphones and tablets that have become primary targets for advertising fraud due to their popularity and technical vulnerabilities. Mobile environments present unique challenges for fraud detection because of diverse device types, operating systems, and user interaction patterns. Fraudulent mobile applications may contain hidden advertisement spaces, generate fake user interactions, or redirect legitimate traffic to fraudulent destinations. The mobile advertising ecosystem's complexity creates opportunities for sophisticated fraud schemes that exploit technical limitations in detection systems.

Artificial Intelligence: The underlying technology powering Google's enhanced fraud detection capabilities, encompassing machine learning algorithms, neural networks, and automated decision-making systems. AI enables rapid analysis of vast data volumes, pattern recognition across multiple data types, and adaptive learning from new fraud techniques. The implementation allows real-time evaluation of advertising placements, user behavior analysis, and identification of anomalies that indicate fraudulent activity. AI systems can process information faster than human reviewers while maintaining consistency in detection criteria and policy enforcement.

Google Research and Google DeepMind: These research divisions within Google focus on advancing artificial intelligence capabilities and developing innovative applications for various challenges including fraud detection. Google Research emphasizes practical AI applications and algorithm development, while Google DeepMind specializes in advanced machine learning research and artificial general intelligence development. Their collaboration with Google's advertising division demonstrates the company's commitment to applying cutting-edge research to real-world problems. The partnership enables rapid deployment of experimental technologies and integration of academic research findings into commercial fraud prevention systems.

Summary

Who: Google's ad traffic quality division, working with Google Research and Google DeepMind, developed and implemented the new detection system affecting advertisers, publishers, and users across Google's advertising platforms.

What: Implementation of multimodal large language models designed to identify invalid traffic, deceptive advertisements, and disruptive ad serving practices, resulting in a 40% reduction in problematic mobile advertising during the pilot program.

When: The testing period ran from late 2023 through October 2024, with the announcement made publicly on August 12, 2025, following approximately 18 months of development and testing.

Where: The system operates across Google's advertising ecosystem, including mobile applications and web platforms, protecting advertisers globally while specifically targeting deceptive practices in mobile environments where invalid traffic rates can exceed 30%.

Why: Invalid traffic wastes advertiser budgets, diverts revenue from legitimate publishers, and creates poor user experiences, with global rates reaching 18% on web and 31% on mobile applications according to industry research, prompting Google to develop more sophisticated detection methods to maintain ecosystem integrity.