The new Adalytics report this time focuses on publishers working with UGC content and placing ads on pages with "unusual" content.

Report Summary

The Adalytics report is a detailed study focusing on the issues and vulnerabilities of modern AI systems used to ensure brand safety. The report primarily focuses on the effectiveness of these technologies in protecting advertisers from placing their ads on inappropriate content.

The study revealed that despite the use of tools like DoubleVerify and Integral Ad Science (IAS), ads from many well-known brands, including Procter & Gamble, Microsoft, IKEA, Mercedes-Benz, and others, appeared on pages with blatantly inappropriate content.

The report does not aim to accuse these publishers of wrongdoing. It is no surprise that UrbanDictionary.com contains inappropriate pages—it's a slang dictionary. The same goes for Genius, which reproduces song lyrics, or Tumblr, a free-flowing platform with diverse images. On Fandom, for example, ads appeared on pages containing offensive words or descriptions of unusual sexual practices. These words could even be in the URL. However, these pages were rated as brand-safe by IAS, as seen in the source code, despite containing adult content.

Separate concerns are raised about opaque Retail Media Networks operating on a walled garden principle, such as Amazon Advertising, Walmart Connect, Target Roundel, ShopRite, Petco, and Best Buy. These companies are now actively expanding in the advertising market but do not offer brand safety services, which have already become standard for advertisers in the open internet.

Scale Issue

Adalytics founder Krzysztof Franasek began researching these media sellers after a global media department head of a major advertiser asked to review brand safety standards, he writes in the report's introduction. The advertiser received full assurance from their DSP and verifiers that their placements occurred on appropriate pages.

In response to the Adalytics report, DoubleVerify (DV) stated that the study was misinterpreted and required corrections. DV notes that the report does not consider important details such as advertiser campaign settings and strategies, which may include exceptions for certain publishers. DV also notes that some examples in the report are based on a misinterpretation of DV's code, which relates to services provided to publishers, not advertisers. The company emphasizes that their content classification system works accurately and correctly and that the Adalytics report is an example of analysis that distorts the real state of affairs.

"The results are manipulated, artificial, or not widely spread," says DoubleVerify in response to the report.

In other words, DoubleVerify claims that Adalytics deliberately selects pages with inappropriate content and creates conditions for ads to be displayed on these pages. This is done to present them in a negative light in the report.

Fandom is one of the largest sites on the internet. An advertiser buying ads on the open internet will get a significant share of Fandom in any campaign. In most cases, the advertiser will not encounter what Adalytics documents in its report, say DV and Fandom executives.

A Fandom representative contacted by Digiday journalists said that the company has not yet reviewed the full Adalytics report. However, they noted that the few screenshots they saw "highlight an industry-wide problem." The representative also emphasized that the examples include content from "old, extremely low-traffic wikis, so it was not flagged by our current moderation systems or Google AdManager, which monitors our active wikis."

One of the largest internet sites with 50 million pages of user-generated content, Fandom reported that less than 0.08% of content was deemed inappropriate. The site also employs various safety measures, including three industry vendors and an internal team that promptly reviews and removes flagged content.

"We do not endorse the placement of dangerous and harmful materials on our platform—such content is prohibited by our rules, and we will not allow it," says Fandom. "Ensuring brand safety on our platform is of the utmost importance to us, and we take these issues very seriously. Although these individual cases were not a widespread problem, we have added additional safety measures to proactively disable ads on low-traffic wikis."

On the other hand, DoubleVerify's response demonstrates the core of the problem. Both DV and IAS claim in their marketing materials and other sources that they provide 100% brand safety. However, it's clear that this is not the case.

"No client has expressed concern about the accuracy of our content categories," states DoubleVerify in their blog.

IAS declined to comment until they had fully reviewed the report.

Artificial Intelligence and the Speed Issue

Major brand safety companies are increasingly using AI to detect and prevent violations. However, advertisers, agencies, and other brand safety experts note that the report raises new questions about the effectiveness of these technologies.

Advertisers surveyed by Digiday and AdExchanger suspected that AI was not as effective as it was portrayed. However, they were still surprised to see that even with easy identification of harmful content, systems were missing it.

They also noted inconsistencies in how AI classifies web pages by risk level. The Adalytics report showed that pages with harmful content on Wikipedia were flagged as low-risk, while pages on The Washington Post and Reuters were classified as medium or high risk, despite lacking such content.

"Brand safety is an issue, and those who truly face this issue are the brands paying for it," said one source. "We are an industry that errs by believing we can develop tools quickly enough to solve all problems."

Jay Friedman, CEO of Goodway Group, believes that the number and severity of examples indicate that brand safety technologies are not effective enough to justify the time and money invested in them. Like other agencies, advertisers, and tech experts, he notes the need for greater transparency to better understand the problem and find an optimal solution. This includes more detailed reporting on every aspect of an advertising campaign so that everyone can make decisions based on the same data.

"The argument 'We can't tell you how it works because then the bad guys will know too' is no longer valid," says Friedman. "These vendors make billions of dollars a year from advertisers and should provide working technology with a transparent mechanism."

Constantly scanning web pages is a complex process, notes Joseph Turow, a professor of media systems and industry at the Annenberg School for Communication at the University of Pennsylvania. Turow and other scholars wonder how often bots browse a site and detect issues or check pages before ads are served. If companies can use AI for contextual targeting, they should be able to ensure reliable brand safety and do so in real time. However, while blocking swear words itself is not challenging, the speed required to analyze each page in nanoseconds before serving ads presents a more complex challenge.

Hidden Threat

The report also addresses advertisers placing ads on Retail Media Networks (RMN).

Reviewing the screenshots of ads in the report, you can see how ads leading to marketplaces were displayed alongside harmful content, such as Fruity Pebbles from Target or Greenies dog treats from Petco. In an AdExchanger experiment, they launched a browser in incognito mode and received ads on pages with undesirable content for organic milk Horizon from ShopRite, Lunchables from Walmart, and electronics from well-known brands from Best Buy.

Note that the latest Adalytics report does not mention Reddit, despite its extensive brand-unsafe content and larger scale compared to Fandom. This is because Reddit is a closed system where third-party JavaScript is prohibited. Adalytics cannot analyze these log files as they are inaccessible.

Meanwhile, DV and IAS are actively partnering with Reddit, Snapchat, TikTok, Meta, YouTube, and Amazon, offering brand safety solutions. In their presentations, they claim that these solutions guarantee 100% safe placements for advertisers.

This brings us to the main conclusion: you should not fully rely on ad tech vendors' claims of 100% safety, especially if you do not have the means to verify it yourself.

Other related materials

• Colossal substitution: Colossus SSP forges cookies in ad requests
• Advertisers paid for ads on Forbes.com, but some ads ended up on a fake site
• Criteo and Microsoft team up to grow Retail Media
• Huge increase in fraudulent video ads in recent months
• Medical companies share patient data with tech companies
• What are Financial Media Networks?