Google Built an AI to Catch AI Slop. The Search Loophole Is Closing.

We're using AI to scale content. Google is using AI to catch it. That's the whole game now.

And yes, every other AI company is working on something similar. They're committed to building the best experience and recommendations for their paying users, not for the companies trying to game the system.

What the paper actually says

The research is titled Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse. Two things matter from it, and neither requires you to care about the machine learning underneath.

First, it works at the cluster level. The system doesn't just flag a bad page, it identifies whole networks of coordinated accounts that share the same synthetic patterns and terminates them together. If a high share of a network is generating templated AI content, the network goes, not just one page.

Second, it's built to keep up. The whole point of the design is that when spammers switch to a new generative model, Google can adapt its detector fast instead of retraining from scratch. The cat-and-mouse loop got a lot shorter on Google's side.

The paper is technically about online video platforms and coordinated abuse. But the direction is obvious, and it won't stay confined to video or to classic Google search. Wherever there's an index, there's going to be a filter.

We've seen this movie before

Here's the part most people are getting wrong right now.

Yes, it's easy to rank in AI search today. Yes, you can spin up 2,000 pages and watch some of them land. It feels like a cheat code.

It was a cheat code in Google's search results around 2000 too. Keyword stuffing worked. Doorway pages worked. Link farms worked.

Then the countermeasures came. And every business built on the cheat code got wiped out by those updates.

We are at exactly that stage with AI search. Early, messy, easy to game, and most importantly: temporary.

What this looks like in practice

I see this shift in our own work. Around January 2026, you could still get a page to perform from a single prompt. Generate, publish, get cited. That window is closing.

What works now is the unglamorous part: getting close to the business. I talk to the people inside the company who work with customers every day and actually solve their problems. A lot of it also comes from Reddit and forums, where people describe their problem in their own words, before anyone has packaged it into a tidy keyword.

The goal is to land exactly at the point where someone has a real problem in your space and knows they have it. Then you answer it with something only that proximity gives you: real expertise, the lessons you learn from doing the work, your own data and statistics, a genuine point of view.

Those pages get noticeably more citations and impressions in both AI search and Google. The templated ones increasingly get nothing. The stuff a model can produce from two prompts is exactly the stuff the detectors are getting good at catching, and exactly the stuff every competitor can produce too.

So what do you actually do

The move isn't "how many pages can I generate this quarter." The move is the hard work nobody wants to hear:

• Fewer pages, more substance. One page built on real customer insight beats fifty built on a prompt.
• Get close to the business. The questions worth ranking for come from the people solving customer problems, and from where those customers already talk, not from a keyword tool alone.
• Publish what only you have. Your data, your numbers, your lessons learned. That's the part a model can't template its way into.

The companies that win the next five years of search — blue links and AI — won't be the ones who scaled slop the fastest. They'll be the ones who were still there after the cleanup.

Sources: Google Research: Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse · Search Engine Journal: Google Research Shows How AI Spam Can Be Detected