Hi everyone, this is Neo.

Today I want to talk about something that gave me chills.

Last September, SEO expert Lily Ray was on vacation in Austria when she casually asked Perplexity a question: what’s new in SEO and AI search lately?

Perplexity confidently told her that Google had just rolled out the “September 2025 Perspectives core algorithm update” — and gave her a detailed analysis.

Except — that update never existed.

Google hasn’t named core algorithm updates in years, and “Perspectives” has been a feature module inside the search results page for a long time. If an update on that scale had really shipped, her inbox would have been drowning in messages from colleagues before she even finished reading.

She checked the two sources Perplexity cited and was stunned: both links pointed to blogs run by SEO agencies, and the posts were clearly mass-produced by AI — they’d had the AI “imagine” a nonexistent algorithm update and then published it as legitimate news.

Perplexity read that “garbage content,” treated it as a reliable source, and fed it straight back to users.

This isn’t an isolated case. This is the new normal of the AI search era.

A 20-minute experiment fooled Google and ChatGPT

This February, BBC reporter Thomas Germain ran an even wilder experiment.

He spent 20 minutes writing a blog post on his personal website titled “The Tech Reporter Who Eats the Most Hot Dogs.”

In it, he invented a “2026 South Dakota International Hot Dog Eating Championship” — a contest that doesn’t exist. No sources, no citations. Pure fabrication.

But then the miracle happened:

• Within 24 hours, Google AI Overviews listed him as the “No. 1 tech reporter who eats the most hot dogs”
• ChatGPT started repeating his nonsense as fact to every user who asked
• Only Claude didn’t fall for it

One ordinary person, 20 minutes, one fake article — and the world’s most powerful AI search systems failed en masse.

Neo’s take: The scary part isn’t that “the AI got fooled.” It’s that AI search systems have an absurdly low bar for what counts as a credible source. A brand-new personal blog with zero authority can get cited by AI as an authoritative source. What does that mean? It means those of us building independent sites — whether we want to “exploit” this mechanism or “defend” against competitors weaponizing it against us — have to rethink the rules of the game.

Model collapse is slow motion; retrieval pollution is live

The industry has been debating a concept called “Model Collapse”:

In short, AI models train on data from the web, but more and more of the web is AI-generated. When the next generation of models trains on that “AI-generated content,” quality keeps degrading, until everything turns into mush.

The subtext of this theory: the problem develops slowly, on a training-cycle timescale, so we still have time to respond.

But that’s not what’s actually happening.

What Lily Ray ran into and what Germain ran into have nothing to do with model training. Those AI models weren’t retrained. The pollution spreads at crawler speed — meaning, in real time.

There’s a key distinction worth making:

| Problem type | Where it happens | Speed | Current status |