You've read the argument that AI search is drowning in slop.
DecryptAds confirmed about 400 AI recipe sites by hand and expanded them to more than 2,100 made-for-advertising domains, in a report built around ad fraud. We pulled Ahrefs AI citation counts for six sites it names. They add up to 209 sampled citations. Perplexity gave 110 and Copilot 63. Google's AI Overviews gave 2 and Gemini none.
My read is that the trust problem sits with specific engines.
The report landed on October 20, 2026. DecryptAds is an advertising supply chain intelligence firm, and it wrote about these sites as an ad-buying problem.
Its team checked at least three pages on each of the 400 sites for AI author photos and AI text, then used more than 190 technical fingerprints to find the rest. The analysis covers July to October 2026. Zach Edwards, who posts as @thezedwards, described the focus on X: "The research is focused on AI slop recipe and cooking websites, the fake chefs who supposedly run them, and the Russian and Chinese ad tech companies integrated into the sites."
The fake chefs are the part that stays with you. The report says recipemaestro.com appears to have taken a 2022 Suffolk Times photo of a real chef and changed only the face. It also lists ads from brands including Adobe, HBO Max, Siemens, Subaru, Capital One and Pepsi running on these sites. AdExchanger covered the ad side and asked the sales houses for comment. HBAgency told it that every publisher it onboards is screened by an independent third-party fraud provider.
The report also names the sales houses it says manage the inventory. It puts hbagency.it at 1,631 of the sites, or 77.1 percent, ezoic.ai at 407 and cultureg.eu at 34. Edwards told AdExchanger that the report is not accusing any one company, and that attributing the sites to a single sales house is hard to do with complete accuracy. That caution applies to the share figures above as well.
Recipe publishers reacted too. Inspired Taste, a recipe publisher, posted on X that Google is the biggest promoter of such recipes through AI Mode and AI Overviews. The post reached few people. I read it as a reaction and nothing more. The ad coverage skips the question of where the citations land.
What the citation counts show
On October 21, 2026, we pulled Ahrefs' AI responses report in subdomains mode for six domains from the report. Ahrefs builds those counts from its own prompt sample, not from every AI answer. A second pull at publish time returned the same numbers. The six are recipessmile.com, recipemaestro.com, addcakes.com, yummyandeasy.com, dailytastekitchen.com and thingstobake.com.
Together they drew 209 citations. Perplexity had 110, Copilot 63, AI Mode 17, ChatGPT 5, AI Overviews 2 and Gemini 0.

Two of the six sites account for nearly all of it, and two drew no citations at all. Six hand-picked domains cannot stand for a network of 2,100. A repeat pull a week from now would show how stable the split is.
So the engines do not agree about these sites, at least in this sample. Two readings fit the Google numbers. Google may filter these networks at the citation layer. Or its AI answers may write slop-style recipes and cite nothing, a gap a citation count cannot see. The counts cannot tell those apart, and they cannot test Inspired Taste's claim about generated recipes at all.
I think the second reading deserves as much suspicion as the first. AI Mode cited these sites 17 times. Google is not clean in this sample either.
Citations are a diagnostic, not the outcome I care about. They show where a bad source has a hold on an answer, and here that is Perplexity and Copilot. A team that tracks one engine will read a different story from the same recipes than a team that tracks five. That gap is the finding. I would ask any slop report to name the engine. One blended number for "AI search" would have hidden the split in these six sites.
Free newsletter
The Sherpa Brief
One email a week on what changed in AI search, what it means, and what to do about it. Written for marketers, SEOs, and publishers.
Free. Unsubscribe in one click.



