Here is a list of words that block advertisements on Reuters.com: dangerous, drug, fire, shot, insider. The words appear in stories about a dinosaur film, a medical breakthrough, the band Arcade Fire, a film production, and a business expansion announcement. These are the stories. The words are the blocklist triggers. The result is that 54% of Reuters news pages — pages that met the brand suitability standards of Integral Science, the verification firm commissioned to evaluate them — would be rejected by the keyword-based blocklists that a significant portion of Reuters's potential advertisers were running.
No one meant for this to happen. That is the essential characteristic of the phenomenon.
The machinery of brand safety — the blocklists, the keyword filters, the content categorization systems — was deployed to protect advertisers from association with harmful content. It has instead created a systematic defunding of the premium news environment that those same advertisers depend on for credibility, reach, and the particular quality of human attention that comes from journalism people actively seek out rather than content algorithmically delivered.
Phil Andraos, General Manager of Reuters Digital, described the situation to AdExchanger in April 2026 with notable restraint: "It is in everyone's interest, including the marketers, not to use this tool at this point." He was referring to keyword blocklists. The tool had been evaluated. It had been found wanting. The ghost in the machine had been named.
The IAS study, conducted in partnership with Reuters in June and September 2024, applied twelve industry-standard brand safety segments at a medium risk threshold to Reuters content — content that IAS's own semantic analysis had classified as low or medium risk, cleared for advertiser deals. The test was not run on obviously problematic content. It was run on content that had already been pre-approved.
Of that pre-approved content, 54% of Reuters news URLs and 13.5% of Reuters lifestyle URL impressions were flagged by the keyword blocklists currently in active use by advertisers in the auction.
The word "Paris" was added to blocklists during the Notre Dame fire. It subsequently blocked Paris Olympics coverage. The same word, a different story. The system could not distinguish between them.
One Fortune 500 advertiser was found to maintain a blocklist of nearly 2,000 terms, including common news vocabulary: Donald Trump, Rudy Giuliani. The list had not been meaningfully updated in years. Nobody at the company knew the full contents. Nobody had ever audited it against current editorial content. It existed as a static artifact, a bureaucratic object that continued to exercise power over billions of impressions long after the reasoning that produced it had expired.
Rich Raddon, CEO of Zefr, described the experience of sitting with brand teams and reviewing their blocklists together: "As you went through the list with them, you'd see names like 'Tom Cruise' or something, and it would be like, 'What the — oh, that was from 12 years ago when he jumped on a couch.'" The list as institutional memory. The list as archaeological record of PR crises that had been absorbed and resolved, preserved in the amber of the filter.
Jamie Credland, CEO of the World Media Group, put the problem in precise terms in a 2024 interview with The Media Leader: "The biggest challenge is advertisers or marketers who are shying away from election or war coverage, who are perhaps using keyword blocklists in an indiscriminate way and may not even be aware of it." The key phrase is "may not even be aware." Indiscriminate use is not the same as malicious use. The advertisers in question are not trying to defund journalism. They are trying to protect their brand. The consequence is the same as if they had intended it.
Scott Gatz, CEO of Q.Digital, was more direct in an AdExchanger podcast the same year: "I dare you to ask a marketer what's on their list — I bet they don't know." This is the structural condition of the market. Blocklists are maintained by media teams, agencies, verification vendors, and trading desks. The people whose budgets are affected by them rarely have direct access to the lists. The lists are treated as compliance infrastructure, not as editorial decisions — but they function as editorial decisions. They determine what is visible. They determine what is funded. They operate with the authority of a publisher's editorial policy but without any of the accountability.
Advertisers are making content decisions through the procedural fiction of technical compliance. They have outsourced editorial judgment to a keyword filter and forgotten that they have done so.
The word "unaware" appears in industry discourse not as a critique but as a finding. The system is populated by participants who do not understand what the system does. This is not an accident. It is a feature of the division of labor that programmatic advertising produced.
The most precise quantification of the economic damage came from a 2019 study by CHEQ and economist Roberto Cavazos of the Merrick School of Business at the University of Baltimore. CHEQ, it should be noted, sells an alternative to keyword blocklists. The conflict of interest is structural and should be acknowledged. The methodology, however, was straightforward: researchers applied a standard industry keyword blocklist of approximately 2,000 terms to 225 articles — the first fifteen uncontroversial or positive articles from each of fifteen major news sites, including the New York Times, the Guardian, CNN, and the Wall Street Journal. The articles were sampled on July 18, 2019. Fifty-seven percent were incorrectly flagged as brand-unsafe.
The most common false-positive triggers were the words dead, injury, lesbian, death, gun, sex, shots, alcohol. These are not fringe terms. These are terms that appear in the majority of substantive news coverage. Seventy-three percent of safe LGBTQ news content was blocked — a figure that illuminates the mechanism's particular violence toward coverage of marginalized communities, where the keywords that describe identity are also the keywords that trigger filters.
The word "lesbian" was flagged in a positive article about a same-sex marriage equality ruling. The system could not distinguish between coverage and harm. It was not designed to.
When the 57% overblocking rate was applied to the approximately $12 billion spent annually on US online news advertising, the result was approximately $2.8 billion in annual publisher revenue lost to false-positive brand safety filtering. This figure was cited widely in industry publications. It was not independently replicated. It was not peer-reviewed. It has not been updated since 2019. The number floats in the discourse as an artifact — useful, approximate, unchallenged because it is convenient to all parties who cite it.
The structural consequence is this: advertisers using extensive keyword blocklists pay more for scarcer inventory while missing the highest-performing audiences. The scarcity is artificial. The premium is real. The quality of attention in a news environment — loyal, purposeful, self-directed — is systematically avoided in favor of content environments that pass the filter because they are too generic to trigger it. The MFA (Made For Advertising) site ecosystem thrives precisely because it is too inert to offend. The journalism dies quietly in the blocked inventory folder.
The industry knows this. The industry has developed alternatives. The alternatives are gaining adoption but slowly.
IAS Context Control uses semantic NLP analysis — full-page natural language processing, entity recognition, 90+ sentiment signals, a proprietary knowledge graph — to understand meaning rather than match keywords. Reuters tested the system and found that replacing seven top-populating keywords with Context Control segments reclaimed 58% of previously blocked content. The same seven words. A different mode of analysis. More than half of the blocked inventory returned.
Other publishers have found similar results. Newsweek achieved 98% brand-safe content classification with Mobian's platform versus 63% under traditional keyword filters. Immediate Media saw a 50% increase in ad delivery using Mantis. News UK unlocked 20% more brand-safe inventory with Illuma. The numbers are consistent: the keyword approach is leaving between 20% and 40% of viable inventory on the table.
DoubleVerify announced updates to its News Accelerator program in March 2025, including AI-powered keyword optimization to identify stale terms and the ability to apply keyword avoidance only to unclassified URLs, leaving pre-approved content untouched. Zefr received the first MRC accreditation for content-level brand safety and suitability labeling on YouTube in March 2026 — a milestone in the industry's move from URL-level to content-level analysis.
GARM, the Global Alliance for Responsible Media, has established an eleven-category taxonomy with tiered risk levels — a move from binary safe/unsafe to suitability tiers, acknowledging that appropriate brand environment is a matter of degree and category fit rather than presence or absence of certain words. This is the conceptual shift that the industry needs: from defensive filtering to contextual matching. It is happening. It is not happening fast enough.
There is a structural tension that the market has not resolved. Verification firms — DoubleVerify, IAS, CHEQ — generate revenue when content is flagged. The more content flagged, the more intervention required, the more the vendor's services are engaged. Publishers allege, and Business Insider reported in 2024, that verification firms are "incentivized to block as much content as possible because that drives their business." The verification firm that clears most content generates less intervention. The firm that flags aggressively generates more workflow. This is not a conspiracy. It is an incentive structure.
The firms dispute this characterization. IAS points out that Context Control specifically is designed to unlock rather than block. The data shows that keywords now account for only 0.25% of blocked impressions on the buy-side open web through IAS's network — a figure that suggests the industry has already partially moved away from keyword blocking at the technical level, even if the cultural and procedural shift has lagged.
The attention economy runs on the premise that human attention is a resource to be extracted and sold. The brand safety infrastructure was built to protect the value of that extraction — to ensure that the attention environment surrounding an ad does not degrade the ad's effect. This is a legitimate concern. The problem is that the mechanism for protecting attention has become a mechanism for defunding the production of the attention that needs protecting. You cannot advertise effectively against journalism that has been financially destroyed by your own brand safety filters.
The ghost in the machine is not a malfunction. It is an emergent property of a system designed by risk-averse participants, maintained by vendors with misaligned incentives, operating at a scale that no individual can audit, and justified by a procedural fiction that says the machinery is neutral when it is in fact making editorial decisions on behalf of brands that have no idea they have delegated that authority. Fifty-four percent of Reuters. A 2,000-word list that someone compiled once and no one has touched since. Seven keywords that could unlock 58% of what the system is blocking. The tools work. The tools are being used incorrectly. These two facts coexist without resolving.