Ads for Adobe, HBO Max, Siemens, Subaru, Capital One and Pepsi ran on recipe websites whose named owners and chefs appear to be AI-generated images, edited photos of real people or stock models, according to a report from DecryptAds, an ad tech data service operated by Svart Works Inc. The report, dated October 18, 2026, traces how roughly 400 such sites, and a wider pool of more than 2,100 related ad-funded domains, sell their inventory. It finds that two sales houses manage the advertising files for 96% of them, and that the partner list linked to one of those sales houses includes ad systems based in Russia and China.

Zach Edwards, a threat researcher who presented the findings on behalf of "our team at DecryptAds," announced the report on LinkedIn on October 20, 2026, describing it as the company's "first report, which we've been working on for months." The report's title is blunt: "AI Recipe Slop Networks Are Being Monetized Through Russian and Chinese Ad Tech." The body is more careful than the title, and the gap between the two matters for anyone reading it as a buyer, a publisher or a regulator.

In Short

DecryptAds says it found about 400 recipe sites run under AI-made or borrowed author identities, grew that into a tracked pool of more than 2,100 related ad-funded domains, and saw ads from brands including Adobe, Pepsi and Capital One on them. It matters to advertisers and publishers because nearly all of these sites hand their ad files to two sales houses, HBagency and Ezoic, and one of those partner lists includes Russian and Chinese ad systems that can receive visitor data when a page loads. The practical change DecryptAds and Edwards are pushing for is that sales houses require sites to disclose AI personas, though the report contains no spend figures and no response from any named company. For the one site where the report itemises the flagged ad systems, four of the nine are Turkish and three are Russian or Chinese.

How DecryptAds built the list

The research did not start from recipes. It started from ad tech companies. DecryptAds says it classifies nearly 300 ad systems as carrying "geo-risk" relationships, and the report breaks them down: 81 ad systems from Russia, 85 from China, 26 in secrecy havens such as Cyprus, and 104 in what it calls elevated-risk countries, naming the UAE, Turkey, Pakistan, Nigeria and Ukraine. By PPC Land's count, that is 296 systems. The company sorts them into four tiers: sanctions, adversary, secrecy haven and elevated.

From there, the team pulled every publisher that authorised at least one of those systems in its ads.txt file, the plain-text list a website publishes to declare which companies may sell its ad space. That pool was filtered with food and cooking keywords, then narrowed further with what the report describes as "other programmatic advertising data-based clustering techniques and technical fingerprints."

The threshold for inclusion was strict on paper. According to the report, a site counted as a true positive AI recipe website only when it showed both AI author photos and AI text indicators, "confirmed by manual review of at least three pages per site." About 400 sites cleared that bar. The analysis ran between July and October 2026.

The expansion came next. DecryptAds says its team built more than 190 technical fingerprints in a tool it calls Feed Builder, drawing on ads.txt and app-ads.txt account IDs, data collected by scanning page content, on-page account IDs, seller names and domains, DNS records, WHOIS registrar data and in-house clustering. Those fingerprints grew the pool to more than 2,100 made-for-advertising domains "owned by a wide range of sellers, many with ties to Morocco and Algeria." That distinction is easy to lose. The 2,100+ figure describes MFA domains linked to the confirmed operators, not 2,100 individually confirmed AI recipe sites.

One technique carries much of the weight. The report describes a low density accountID as one "mapped to a low number of publisher websites, thus making it more likely to be owned by the website owner and being a potential pivot point." A seller ID that appears on only a handful of domains tends to belong to whoever runs those domains, so it can link sites that look unrelated on the surface.

DecryptAds will not publish the fingerprints. "For obvious operational security reasons, DecryptAds will not be publishing the various technical fingerprints used to group these sites by their operator(s)," the report states, adding that organisations interested in the details can contact the company. That leaves the clustering claims, which run through the rest of the report, impossible to reproduce from the published material alone.

Two sales houses, 96% of the sites

The most concrete finding concerns who manages the ads.txt files. Almost none of the publishers appear to run their own. According to the report, the managers declared across the 2,100+ tracked sites break down as follows:

  • hbagency[.]it: 1,631 websites (77.1%)
  • ezoic[.]ai: 407 websites (19.2%)
  • cultureg[.]eu: 34 websites (1.6%)
  • monetizemore[.]com: 11 websites (0.5%)
  • journeymv[.]com: 8 websites (0.4%)
  • monumetric[.]com: 5 websites (0.2%)
  • adpushup[.]com: 2 websites (0.1%)
  • No manager domain declared: 18 websites (0.9%)

HBagency and Ezoic together account for 96.3% of the declarations. The counts add up to 2,116 by PPC Land's calculation, slightly above "2,100+", because a handful of sites name two managers in the same file and the report counts them under each. Some files list account IDs for both HBagency and Ezoic while naming only one as manager. The report notes that the ads.txt 1.1 guide tells buyers to treat conflicting manager declarations as if no manager was declared, but also leaves interpretation to each buyer, so "it's unclear what the practical impact is for those sites." (Domain names in this article follow the report's practice of bracketing the dot.)

HBagency, an Italian company, provides ad tech support to "over 4,500 publishers," according to the report, and "appears to work with ~150 ad tech companies based on the accountIDs they list within client websites." Nine of those roughly 150 partners appear on DecryptAds' risk list.

The report adds an important qualifier. The MANAGERDOMAIN field "doesn't guarantee that all of the accountIDs within the file are owned or managed by this sales house." Publishers often add their own accounts. DecryptAds says that when the same account IDs recur across seemingly unrelated publishers sharing one manager domain, it assumes those accounts are "in some way controlled by that sales house." That is an inference, and the report labels it as one.

Ezoic is a publisher monetisation platform PPC Land has covered as recently as September, when the company reported cutting display impressions by about 25% to raise viewability to 87%. The report does not itemise which risk-listed systems appear on Ezoic-managed sites. Neither HBagency nor Ezoic is quoted, and the report contains no response from either company.

What one site's file shows

The report uses dailytastekitchen[.]com, an HBagency-managed recipe site, as its worked example. DecryptAds' risk panel flags 9 of the site's 153 ads.txt partners:

Tier Ad system Country
Sanctions betweendigital[.]com Russia
Adversary mintegral[.]com China
Adversary yeahmobi[.]com China
Secrecy haven markappmedia[.]site Cyprus
Elevated admatic[.]com[.]tr Turkey
Elevated monetixads[.]com United Arab Emirates
Elevated temedya[.]com Turkey
Elevated pixad[.]com[.]tr Turkey
Elevated rekmob[.]com Turkey

The figure caption says "most of them" are based in Turkey. By PPC Land's count, four of the nine are Turkish, which is the largest group but not a majority. One is Russian, two are Chinese, one sits in Cyprus and one in the UAE. That breakdown is the only itemised one in the report, and it does not match the emphasis of the headline: on this example, Turkish ad tech outnumbers Russian and Chinese ad tech combined.

The report is not hostile to sales houses as such. Outsourcing ads.txt management is "a totally normal process," it says, and companies doing that work "deserve credit as it's not easy work and the ecosystem is ripe with challenging fraud efforts." The problem it describes is distance. Sales houses carry long partner lists and push them to every client site, so "the publishers themselves don't always know the ad systems they are working with." The SSPs and sales houses, in turn, "may conduct some Know Your Customer (KYC) processes when the clients sign up" but "could just as well be largely 'checked out' otherwise."

What reaches an ad system when a page loads

The report's concern about Russian and Chinese systems rests on how programmatic auctions move data. When a page with programmatic ads loads, the visitor's IP address, device and browser details (the user agent string) and the page itself can be shared with numerous ad systems "in mere milliseconds," according to the report, along with third-party cookies where browsers still send them. Most of that exchange happens server-side, so even a careful client-side audit cannot see every recipient of the bidstream. The report points to advocacy for the Demand Chain Object and buyers.json, and to the SupplyChain Object, as attempts to close that gap.

The ads.txt file is the closest public proxy. DecryptAds uses the New York Times as a reference: its file lists roughly 30 DIRECT accounts across about 20 ad systems. Each account can be checked against the ad system's own sellers.json file, which confirms whether the seller is still authorised and whether it is a publisher or an intermediary. A listing is not proof of data flow. The report notes that a publisher may have switched off a partner on the server side and simply forgotten to update the file, and that some deals are regional and rarely visible from outside.

Edwards put the uncertainty plainly in his post. An authorised system can receive data on each page load and bid; if it wins, it serves an ad of its choosing. "None of that is guaranteed on a given visit, which makes it easy to ignore," he wrote.

The scale claim in the report is broad. DecryptAds says "there are more than 75,000 websites with programmatic ads which authorize ad tech companies, even those apparently based in China and/or Russia, to sell their inventory, resulting in the apparent transfer of user data to those companies." The sentence does not say how many of those 75,000 sites authorise Russian or Chinese systems specifically, as opposed to any ad tech company. The report's concluding section states the policy concern directly: "Companies based in China and Russia are under unique legal obligations by their authoritarian governments, and user data being shared to these companies creates unique risks for unsuspecting or unaware end users."

That argument has legal precedent behind it. In January 2025, EPIC and the Irish Council for Civil Liberties filed an FTC complaint over Google's real-time bidding, the first under the Protecting Americans' Data from Foreign Adversaries Act. In February 2026, a class action in California accused Google of sending browsing data through RTB to Pangle, MediaGo and Temu, relying in part on the Justice Department's Bulk Sensitive Data Rule, in force since April 8, 2025. And in October, Edwards' post noted that the Citizen Lab and Confiant had just shown targeted banner ads being used to deliver zero-click spyware. PPC Land covered that case, in which Citizen Lab said ads bought by Israeli firm Discover-Tech delivered Candiru spyware to at least 45 devices, and Google and IAB Europe had listed the vendor since 2020.

Borrowed faces and invented kitchens

Most of the report is examples. They are specific, and in several cases they involve real people.

The most prominent is addcakes[.]com, whose owner and author, "Elena," is described on the site as "a busy mom who wanted to connect with my roots through baking." According to DecryptAds, the site and its social profiles use five unique photos that resemble the actor Emily Blunt. Exact copies do not appear to exist publicly, so the report concludes that someone likely used an AI tool to create new versions of her cooking and in other settings. The footer links to an X account, @yassine_bourais, named "Add cakes," which X labels as based in Morocco, joined in February 2022 and connected via the Morocco Android app. The site's ads.txt names hbagency[.]it as manager.

A second case involves a private individual. Recipemaestro[.]com, according to the report, took a photo of a chef in Greenport, New York, published by The Suffolk Times in July 2022, and changed only the face, keeping the body, the pie and every other detail. DecryptAds connected the site to 18 other AI recipe websites through technical fingerprints; its ads.txt is managed by ezoic[.]ai. Its linked X account, @RecipeMaestro, shows "Account based in Algeria" and a join date of October 2023. A linked Facebook page called "Daily Dishes," with 75,000 followers, was created on October 7, 2023, and lists page managers in Algeria and the United States.

Others rely on stock imagery. The report matches the owner photo on miaspice[.]com to a Magnific stock image, and does the same for "Evelyn Barrett" on airfryerecipespot[.]com. On thingstobake[.]com, a reverse image search ties the photo of "Maggie Porter" to Dreamstime, Alamy and an Amazon cookbook cover, while her biography describes a woman "37 years old, living in a cozy little farmhouse just outside Asheville, North Carolina." On yumfamilyrecipes[.]com, the photo of "Emily Grace" also appears as "Patricia Menon" in a testimonial on an unrelated site.

Clustering shows up in the faces too. An AI image of "Sozane" (also spelled "Sozan") appears on yummyrecettes[.]com, yourrecipes[.]net and sozanrecipes[.]com, part of a cluster of about 20 sites. The yourrecipes[.]net about page welcomes visitors to SozanRecipes.com despite the different address. Two of those sites are managed by HBagency; sozanrecipes[.]com is managed by Ezoic.

Teams of chefs who do not appear to exist

Several sites go beyond a single persona. Tastyflavorhub[.]com presents six AI-generated chefs, each with a cooking specialty, and is connected to 40 other AI recipe sites. Simplefoodzone[.]com describes a "diverse team of culinary experts" and is linked to 34 others; the report's screenshots show VRBO ads on its team page. Spicedandbaked[.]com and easydinerecipes[.]com each present three-person teams, and both list the same founder name, Alicia Wilson, with the same job titles. Chefyoyo[.]com has four chefs whose quotes cite their own recipe counts; a related site, tastedessert[.]com, features an author claiming "678+ recipes."

Credentials nobody can check

The invented biographies often borrow prestige. Yummyandeasy[.]com names three different owners on one about page, "Sophia Harper," "Lina Williams" and "Emma Claudie," and lists Le Cordon Bleu training and work at Michelin-starred bistros in Paris and Casablanca without naming a single restaurant. "Lena Carter" of torecipes[.]com claims an education at the Culinary Institute of America and the University of California, Berkeley; the site belongs to a cluster of 20. Masonrecipes[.]com, in the same cluster, presents a "Pro Chef" who "attended culinary workshops at New York University." Onedaycooking[.]com lists studies at a Swiss culinary academy beside a likely AI image that includes a child, whom DecryptAds blurred "out of an abundance of concern that a minor's image may have been used."

Some sites assert a corporate owner. Thecrispycheff[.]com says it is run by "Godza LLC." The report states: "We haven't been able to find any indication that 'Godza LLC' is an actual registered business." Two others, thechefrecipe[.]com and recipebite[.]com, list the same phone number, 888-405-8721, alongside a Budd Lake, New Jersey address; both sit in a cluster of 15 sites managed by Ezoic.

The image errors are sometimes obvious. The chef's jacket on ketodietguide[.]net reads "Keto Gudie." On foodiefiber[.]com, the author photographs herself in a bathroom mirror that shows no reflection. "Whether AI or a sign of the supernatural, in this case our team believes the former is more likely," the report says. Mrrecipes[.]net presents a woman named Penelope as its author.

Fabricated credentials of this kind go to the heart of what Google groups under E-E-A-T, the experience, expertise, authoritativeness and trustworthiness framework its raters use. Edwards linked the report to Google's recent move. Two weeks before his post, he wrote, Google "updated their search policies to penalize publishers who use 'deceptive authorship information,' a positive step for slowing down AI slop websites. But far more needs to be done, and a lot of folks don't understand the layered risks from these sites." PPC Land's reporting describes the change somewhat differently. Google edited its "Creating helpful, reliable, people-first content" help page to say that invented writer names, AI-generated headshots and false credentials are "a form of deception." The edit was first noticed on October 2, 2026, and nothing PPC Land reviewed showed rankings changing as a result.

A recipe site serving malware

One finding is not about ads at all. During manual review, DecryptAds found that dailytastekitchen[.]com, whose author "Nora Haven" wears a chef's coat with garbled embroidery where a name should be, appeared to be compromised "for over a week." Loading almost any page triggered a ClickFix lure styled as a Cloudflare human-verification check, which told visitors to open Terminal and paste a command.

According to the report, the compromise resembled a campaign known as "EtherHiding," which stores its command-and-control instructions in a smart contract on the BNB Smart Chain and retrieves them when a page loads. The report says that campaign "has compromised thousands of WordPress sites." DecryptAds documented three domains in the chain:

  • mytds1029380[.]com, which initiated the injection, had 16 malicious detections on VirusTotal as of September 22, 2026.
  • senterprise2026[.]com, which injected the script, had zero detections when DecryptAds first found it on September 10, 2026, and 20 by September 22.
  • cloudeliveryenterprise[.]com, used in the lure, had 16 detections as of September 22.

The report frames this as a site compromise, not an ad-delivered attack, and as evidence that "some of the owners of these AI recipe networks may even not be paying attention to their own sites." Edwards' post compresses this to the site having "served a ClickFix malware lure styled as a Cloudflare check for over a week."

Morocco and Algeria

The geographic thread in the report comes from the platforms, not from ad tech data. X and Facebook now publish location information on accounts and pages, and DecryptAds followed the social links on site after site. Beyond addcakes[.]com (Morocco) and recipemaestro[.]com (Algeria), the report lists:

  • @amytastyblog, linked from amytasty[.]com: Algeria, joined January 2025
  • @yumfamilyrecipe, linked from yumfamilyrecipes[.]com: Algeria, joined January 2025
  • @thefamilycookin, linked from thefamilycooking[.]com: Algeria, joined January 2025
  • @5minMomRecipes, linked from tastydishy[.]com: Morocco, joined July 2023
  • @recipesbyjanet, linked from recipesbyjanet[.]com: Algeria, joined February 2025
  • "Thecrispychef" Facebook page, 15,000 followers, created August 17, 2024: page manager in Morocco

The report hedges this section heavily. "These platform location disclosures are inferred by X and Facebook rather than verified, and can be skewed by VPNs, travel or accounts changing hands, so we treat them as indicative rather than conclusive," it says. Its explanations range widely: the pattern could be "potentially indicative of a network of inauthentic coordinated behavior, or potentially an indication that someone may have attended a training course on how to run these types of websites." Then: "Or maybe it's simply an inexplicable coincidence."

Nothing in that section connects the North African operators to the Russian or Chinese ad systems in the headline. The two findings run in parallel. One oddity does point east: valerierecipes[.]com loads a Chinese-language verification screen, asking visitors to wait while the request is checked, before showing a recipe site attributed to "Valerie Harper." The report does not explain it.

The advertisers on the page

The report names Adobe, HBO Max, Siemens, Subaru, The Wall Street Journal, Capital One, Sam's Club, LinkedIn, Google Workspace, Tumi, Pepsi, Anker and VRBO as brands whose ads it found on these sites. Edwards' post called a subset of them "enterprise advertisers." The screenshots add others, including Square, TeamViewer, Brother, Carbonite, PetSmart and Acorn Stairlifts.

The placements are specific. An HBO Max MLB banner overlaps the navigation on thecrispycheff[.]com. Two Siemens ads cover half the author photo on simplerecipes[.]me. Snackinstyle[.]com shows banner ads between every paragraph of its about page, including three different Adobe creatives and two Anker Solix banners. A Pepsi ad sits on a thechefrecipe[.]com article about a "Flat Tire Trick," behind a pop-up reading "Please view this ad to continue." Sam's Club appears under a Spanish-language Actimel ad on tastygreat[.]com, a site DecryptAds connects to 70 others, the largest cluster in the report.

Ad density is a theme. Simplemealmagic[.]com carries three banner ads and two video ads on one view. On many sites, banners load over the author photo itself.

What the screenshots cannot show is how those ads were bought. The report does not identify which demand-side platforms, exchanges or deals delivered them, whether they arrived through Russian, Chinese or other risk-listed systems, how many impressions ran, or what they cost. Programmatic ads vary by visitor, location and moment, so a screenshot records one serving, not a buying pattern. None of the named advertisers is quoted.

Dr. Augustine Fou of FouAnalytics commented beneath Edwards' post that he had "seen a bunch of brand new domains in live campaign data in FouAnalytics too." He added that "none/few of these sites are known by humans, so virtually all of the traffic is artificial." That is his assessment; the DecryptAds report makes no traffic-quality claim and publishes no traffic figures.

What the report does not establish

Read as a whole, the report is stronger on deception than on money. Several limits stand out.

First, there are no spend figures. DecryptAds documents authorisations, ownership signals and ads on screen, but not how much advertiser money reached the sites, the sales houses or any risk-listed ad system.

Second, the risk list is wider than the headline. Of the 296 systems DecryptAds tracks, 166 are Russian or Chinese, by PPC Land's calculation from the report's figures. The other 130 sit in Cyprus and in elevated-risk countries that include Turkey, the UAE, Pakistan, Nigeria and Ukraine. The initial filter caught publishers authorising any of them. The report does not state how many of the 400 confirmed sites, or of the 2,100+ tracked domains, authorise a Russian or Chinese system specifically.

Third, image provenance is probabilistic. "We say 'likely' in places because we can't prove how any one image was made, only that it shows the artifacts, reuse patterns and stock-photo matches typical of these sites," the report states.

Fourth, the company has a commercial interest in the findings. DecryptAds sells access to the data behind the report through accounts, an API and an MCP interface, and the report refers readers to its Geo Risk page, publisher pages and Feed Builder throughout. The fingerprints that underpin the clusters are withheld and offered on request. The report also invites any company or publisher it names to submit evidence for corrections or comment.

Fifth, the dates differ slightly. The report page is dated October 18, 2026, a Sunday. Edwards announced it as newly released in his LinkedIn post of October 20, the same day AdExchanger covered it, according to a comment he left under the post.

The report itself acknowledges the grey area it is working in. "We would be remiss not to mention that anonymity and pseudonymity matter online and plenty of political and sensitive-topic publishers can't publish under real names. But many AI slop operators appear to be aggressively exploiting that grey area."

Why it matters for the market

AI recipe sites are not a new concern for buyers. In July 2025, DoubleVerify published a report warning that AI recipe sites threaten brand safety, naming sites such as InsanelyGoodRecipes.com and citing survey data that 83% of consumers think AI-generated content should be legally required to carry a label. That same month, IAS identified AI-generated slop sites as a major ad quality threat. In March 2026, DoubleVerify's Fraud Lab exposed AutoBait, a network of more than 200 MFA domains using language models to produce clickbait slideshows, and in April IAS opened a beta for pre-bid avoidance of low-quality AI content.

The most important benchmark came in July 2026. Analysis from TAG, the ANA and Fiducia put AI slop at 1.3% to 2.4% of open web programmatic spend in the first quarter of 2026, across 11,552 slop domains, and found that slop inventory was graded premium more than 70% of the time once measurability was accounted for. In that analysis, slop inventory showed lower invalid traffic and higher viewability than clean supply, which is why standard quality checks rated it well. The DecryptAds report adds a different axis: not whether the inventory performs, but who is behind it and which ad systems sit in its supply chain.

The recipe niche has its own history on PPC Land. The report's phrase "Frankenstein recipes" echoes the term food bloggers Adam and Joanne Gallagher of Inspired Taste used when warning in December 2025 that AI recipes risk ruining holiday meals, and again in February 2026 when they compared their key lime pie with an AI Mode version. Human recipe publishers, the report argues, are the ones losing budget: AI recipe sites "have been diverting advertiser budgets from many legacy publishers and wasting ad spends."

The structural point concerns ads.txt itself. The file was designed to stop domain spoofing by listing authorised sellers, not to vet the publisher behind the domain. An IAB Europe study published in August 2025 found that 72.64% of 2,054 European news publishers hosted ads.txt files, and roughly 79% of valid lines matched sellers.json entries. Proposals to replace it, such as the adagents.json format debated earlier this year, focus on describing commercial relationships more richly. None of them asks whether "Elena" is real. Who, then, checks the chef? When a sales house pushes one partner list to thousands of client sites, the authorisation layer becomes a distribution mechanism, and the only checkpoint left is whatever onboarding the sales house performs.

Google's search side has moved in the same direction. The September 2026 spam update began rolling out on September 24 with a window of up to two weeks, and the helpful content guidance on fabricated authors followed. Whether search demotion reduces the ad revenue of sites that, according to the report, aggressively spread their content on social platforms such as Pinterest, is a separate question the report does not address.

What DecryptAds wants changed

The report's recommendations are short. Every site using AI personas, AI authors or AI owner photos "should be required to disclose that usage in some transparent fashion, on their website." SSPs and sales houses, it argues, "could easily change their rules to require this simple disclosure on publisher websites." On ad tech partners, it cautions sales houses to "be careful with the ad systems they partner with," and warns that any publisher authorising Chinese or Russian systems in its ads.txt or app-ads.txt files "is likely to get on the public's radar by doing so."

Edwards went further on timing. "Any site using AI personas should have to disclose it and the sales houses monetizing them could require that tomorrow as part of KYC requirements," he wrote.

The report closes by stating that "many questions remain unanswered" about these sites and their ad tech partnerships, and promises updates if the company learns more. On the moral and legal weight of what it found, it declines to rule: "We'll leave what this means legally, ethically and politically to folks smarter than us, but our instinct is that there are significant red flags here."

Timeline

Summary

Who: DecryptAds, an ad tech data service operated by Svart Works Inc., and Zach Edwards, a threat researcher who announced its findings. The report names sales houses HBagency and Ezoic, nine risk-listed ad systems including Russia's betweendigital[.]com and China's mintegral[.]com and yeahmobi[.]com, and advertisers including Adobe, HBO Max, Siemens, Subaru, Capital One and Pepsi. None of the named companies responded in the report.

What: A report identifying about 400 AI recipe websites with AI-generated, face-swapped or stock-photo authors, expanded by fingerprinting into more than 2,100 related made-for-advertising domains. HBagency manages ads.txt for 1,631 of them and Ezoic for 407. The report documents fabricated culinary credentials, a ClickFix malware lure on one site and social accounts that X and Facebook place in Morocco and Algeria, but provides no spend figures.

When: The analysis ran from July to October 2026. The report is dated October 18, 2026, and Edwards announced it on October 20, 2026.

Where: The sites operate on the open web and promote themselves on X, Facebook and Pinterest. The risk-listed ad systems are based in Russia, China, Cyprus, Turkey, the UAE and other jurisdictions; HBagency is Italian.

Why: Ads.txt authorisations pushed by sales houses can route visitor data and advertiser money to ad systems that publishers may not know about, including firms subject to Russian or Chinese state data demands. AI slop as a whole already takes an estimated 1.3% to 2.4% of open web programmatic spend, according to earlier industry analysis, and DecryptAds argues that sales houses could require AI persona disclosure as part of their onboarding checks.