Google’s Spam Update Targets What Nobody Read

Google’s September 2026 spam update hits mass-produced content nobody reviewed, not AI itself. Five signs your traffic drop is spam-related.

Google’s September 2026 spam update isn’t aimed at AI content. It’s aimed at mass production nobody read before publishing. The same week, Google put “low effort” into writing for the first time.

Below: three dates, Google’s four attributes, five signs your drop is spam-related, and the one thing to do this week.

The September spam update is still rolling out and the forums are full of one question: “Our traffic halved overnight. What happened?” Marie Haynes, who tracks these updates closely, thinks this one targets mass-produced AI content. Most site owners in the middle of a drop don’t know where it’s coming from. In the last two weeks we’ve had site owners ask us the same thing.

The headline you’ll read everywhere is “Google cracks down on AI content.” That’s the wrong takeaway. Back in August I wrote that Google doesn’t penalize AI content. This week, in the middle of the rollout, Google updated two docs and almost wrote the rest of that sentence down. It isn’t about AI. It’s about nobody reading what AI wrote.

Grab a coffee.

The thesis

Google now defines “low effort” in its own documentation: auto-generating pages from feeds, and generating large amounts of text with AI without manual review. A spam update shaking exactly this kind of site in the same week may not be a coincidence. The problem isn’t “we used AI.” It’s “we published what AI produced and nobody read it.”

1. What happened? Three dates, one story

Sept. 24: Google launched the September 2026 spam update, the fourth spam update of the year. It’s global, applies to all languages and should take about two weeks to roll out. Google didn’t name a target or add a new policy.

Sept. 30: A second wave hit. Many site owners told Glenn Gabe they saw big drops on exactly that day, and according to Search Engine Roundtable’s round-up of the second wave, almost every volatility tool from Semrush to Sistrix flagged the same date.

Oct. 1: Google updated two documents together:

• The guidance on AI-generated content gained this sentence: it’s “critical” to manually fact-check and review all AI-generated content for accuracy and trustworthiness before publishing. Two reminders sit next to it: generative models don’t retrieve facts, they predict likely word sequences, and output can hallucinate.
• The helpful, reliable, people-first content doc gained two new sections: a definition of main content, and the four attributes raters look for in it.

Google’s official note is modest: it aligned the docs with presentations it uses at developer events. Barry Schwartz points out in his Search Engine Land piece that “critical” reads more like an instruction than a suggestion. I agree. When Google writes “critical” in a sentence, it wants us to read that sentence.

2. Main content and four attributes

Google defines main content as everything that directly serves the page’s purpose. Not only text: tools like calculators, user reviews, information inside tabs and page titles all count.

Raters look at four things in main content:

There’s a nice note under talent or skill: not everything needs expertise. Someone describing their experience of shoveling snow doesn’t have to be an expert. First-hand experience can beat a title.

There’s a new warning too: made-up author profiles. Presenting content as written by an expert with an AI-generated headshot, a made-up name or a fake title is, in Google’s words, a form of deception. And deception makes a page untrustworthy to readers and to Google’s automated quality systems alike.

Marie Haynes, in her line-by-line video on the new doc, makes the key point: these questions aren’t a scorecard. Google’s machine learning systems try to imitate the “helpful content” described in the doc. Ticking boxes one by one won’t do it. You have to look at the page as a whole.

3. Is your drop about the spam update? Five signs

Not every traffic loss comes from a spam update. AI answers are cutting everyone’s clicks, plugins can break URL structures, servers can block AI bots. Take the spam update seriously if two or more of these fit:

1. The dates match. In Search Console, the drop started around Sept. 25-27 or around Sept. 30.
2. The loss sits in one section. Mass-produced blog, city or product-variation pages fell, while hand-written pages held.
3. There’s a manual action. Check “Security & Manual Actions” in Search Console.
4. Folders are dropping out of the index. The “Crawled, currently not indexed” count in the Pages report jumped.
5. The authors aren’t real. The site has author profiles with AI-generated photos and nobody behind them.

One difference with spam updates, in my view: the “wait for the next update” strategy that sometimes works after core updates doesn’t really work here. As long as the problem content stays live, the signal stays too.

We wrote up how we read these five signs together with Search Console data on our traffic loss recovery plan page.

4. What to do

• Take inventory. List the pages you mass-produced with AI. How many, when, and who reviewed them?
• Prune. Remove or merge pages with no traffic and no links. Noindex the ones in between. As Marie Haynes puts it: if you have hundreds of pages that wouldn’t exist if not for search, you may need to remove most of them.
• Add effort to what stays. First-hand data, real experience, screenshots, original analysis.
• Make authors real. Real name, real photo, real track record.
• Make pre-publication review a process. Don’t leave it to one person’s mood.

From the field: we ran the same test on ourselves

When I read the new “deceptive authorship” clause, the first thing we did was check our own site. Stradiji’s About page shows two people: me, and Emre Ercan, who has been on our team since 2013. Both with real names, real titles and real histories. Our blog bylines link to real people.

Here’s how the rule played out inside the newsletter where I first covered this. We run every piece of AI-assisted content through three gates before it goes live. Gate one is brand voice: does it sound like me, are the terms right? Gate two is an E-E-A-T checklist: named author, sourced numbers, correct dates. Gate three is an originality score: what does this say that isn’t already on the internet? If there’s no first-hand data, we don’t invent it. We leave a gap in the draft and ask.

In the first draft of that issue, the AI wrote an opening line: “this week the same question landed in my inbox.” It sounded perfectly natural. But there was no such data. The sentence bounced at gate two and was deleted. That’s exactly what Google means by “manual review is critical.”

Outside the U.S., the pressure is sharper. In Turkey, the number of sites publishing dozens of AI posts a week is growing fast, and I expect the impact of this update to show up more clearly in non-English search over the coming weeks.

The one thing to do this week

Open Search Console. Set the Performance report to the last three months and mark Sept. 24 and Sept. 30. Then sort pages by lost clicks. If most of the loss comes from a section you mass-produced with AI, you’ve probably found your cause.

If you can’t read the picture, or you’re not sure whether it’s spam, AI answers or a technical error, you’re not alone. We built a four-stage plan for sudden traffic drops, and the first 48 hours are covered free on the traffic loss recovery plan page.

Closing

Producing content with AI isn’t a crime. I wrote this post with AI too. The difference is whether a human reads what comes out, fixes it and signs it. This week Google wrote that down more plainly than ever.

So, a question: on your team, who reads AI-generated content before it goes live?

This post is adapted from the lead story of The SEOs Diners Club #241, my weekly newsletter on SEO, GEO and AI search.