Spam in Gemini Notebook lasted 3 days

On 19 September 2026 Charles Floate wrote on X that he had "hacked" AI Overview using Google's own resources: he put his text on a public Gemini Notebook page, Google's service for working with documents and notes, and that page made it into a generated search answer. By 22 September Search Engine Roundtable counted more than 12,000 such pages in the index: peptides, promo codes, adult content, links to apps. On 23 September Google removed the directory of public notebooks from search results and cleared out the spam. The window lasted about 3 days, and it is a very precise illustration of how holes on someone else's strong domain open and close in general.

What Gemini Notebook is and why its pages ended up in Google search

Gemini Notebook is the former NotebookLM, a Google service where you upload your own sources (texts, links, PDFs) and the model answers based on them, makes summaries and overviews. On 16 July 2026 Google renamed NotebookLM to Gemini Notebook, in late July 2026 moved it to notebook.google.com, and the old notebooklm.google.com now redirects to the new address (according to nlmtools).

The key detail: a notebook can be made public, and then it gets an ordinary URL on a Google domain. Which means search crawls and indexes it like any other page. On 21 September 2026 Gagan Ghotra publicly called this a new parasite SEO tactic: public notebooks get indexed, and people running black hat schemes take advantage of that.

How parasite SEO works and why a Google domain turned out to be a convenient host

Parasite SEO is when content is placed not on your own site but on someone else's strong one. In Google's guidelines for quality raters the wording is simple: content is placed on a host site mainly for the ranking signals the host has already accumulated. The classic hosts are big news outlets with their "partner content" sections, forums, classified ad sites, any platform with user generated content.

In this logic google.com is a host with the maximum level of trust. A page on it does not need to earn a reputation, it inherits one. Hence the speed: according to Search Engine Roundtable, by 22 September there were more than 12,000 pages in the index, that is an average of about 4,000 new pages a day over the 3 days the tactic lived in public.

There was also a completely mundane sign of how crude it all was. On 22 September 2026 Glenn Gabe showed that in search results one of these pages displayed the very prompt used to generate it. People were not even hiding the instruction the text was born from.

1. 19 September 2026Charles Floate shows a Gemini Notebook page in an AI Overview
2. 21 September 2026Gagan Ghotra calls it a parasite SEO tactic, Glenn Gabe confirms the scale
3. 22 September 2026Search Engine Roundtable finds more than 12,000 pages in the index
4. 23 September 2026Google removes the public notebook directory from search results and clears the spam

Why Google did not wipe 12,000 pages in an hour, although it could

Search Engine Roundtable pointed out the oddity directly: Google could have removed those pages from the index instantly, since the domain is its own, and did not do it. Three days of spam on your own domain, for a company that hands out quality guidelines to everyone, is expensive in reputation terms.

There are two explanations, and both are mundane. First: manual index edits by URL pattern are a decision somebody has to make and sign off on, not a button sitting next to an engineer on duty. Second: you first have to close the possibility itself, otherwise the cleanup is pointless, the pages will fill back up within a day. On 23 September the public notebook directory disappeared from search results, meaning they fixed the source, not just the consequence.

The practical takeaway for anyone watching competitors: a platform's reaction to spam is almost never instant, but it does come. You cannot plan your work around "Google does not notice", because they noticed 2 days after the first tweet.

The site reputation abuse policy: what is banned and how long Google took to enforce it

This story is not new, there is a ready made rule for it. Google announced the site reputation abuse policy in March 2024 together with two neighbouring ones: scaled content abuse (mass generation of pages without value) and expired domain abuse (buying dropped domains for their history). According to the breakdown by Digital Hitmen, the chronology looks like this:

• March 2024: the policy is announced;
• 5 and 6 May 2024: the first manual actions begin, that is roughly 2 months after the announcement;
• January 2025: the wording about parasite content is added to the quality rater guidelines;
• August 2025: algorithmic enforcement is strengthened by a spam update, roughly 17 months after the announcement.

So the path from "we ban this" to "we catch this algorithmically" took Google a year and a half. That gap is exactly where all tactics of this kind live. And the same gap explains why a fresh hole in a new service gets closed manually and fast, while systemic schemes on old hosts drag on for years.

One more detail worth knowing: in November 2025 the European Commission opened an investigation into the enforcement of this very policy, as Search Engine Journal reported. The complaint is that legitimate publishers who run partner content get hit by the sanctions. Google is defending its position. In practice this means the wording of the policy is still going to move.

Is it worth climbing into windows like this: counting with the numbers from this story

Let me count honestly. The window lasted 3 days from the first public disclosure to the cleanup. In that time more than 12,000 pages made it into the index, and all of them were wiped in one move. Not one of them turned into an asset: not a subscriber, not an accumulated link profile, not a brand.

And now the second number. On 22 September 2026 Peec AI (Malte Landwehr) showed that the share of AI Overviews with external links had grown from almost 0% to more than 25%. A year ago a generated answer barely passed links outward, now every fourth one does. The chance of getting into a Google answer with your own site is higher today than it was, and it does not burn out in 3 days.

So my answer is this: you should watch windows like this, you should not climb into them with your main project. Three days of life against a domain you have invested several years in is not a deal.

At the same time, fighting for a place in the answer is getting trickier. On 22 September 2026 people noticed that some links inside AI Overviews lead not to sites but into AI Mode, the search chat mode inside Google. That same day Search Engine Land wrote that Google is testing a "Dive deeper" button in Discover: clicking it leads not to the publisher's article but to a generated overview of the topic inside Google. Both things shorten the user's path to the site rather than lengthen it.

How to spot a window like this before everyone else

You have to watch the index itself, not tweets. This whole story is visible through the site search operator on the service's domain: if a product has public pages, they will be there. The next question is what lives on them.

I keep a simple table for checks like this: a column for "service", a column for "the date public pages appeared on it", a column for "what a 5 minute site search turned up", a column for "date checked". Every 2 weeks I go down the list from the top. Without the table I forget half the services within a month.

• every 2 weeks, use the site operator to check new public sections of big services
• look at whether a public page carries noindex and whether it can be created without moderation
• record the date of the tactic's first public disclosure so you know how long the window has already been open
• do not carry the tactic over to your main domain, even if it works for a competitor
• check whether your material is being retold on someone else's pages by searching a unique phrase from your text in quotes

How to shut your site off from being scraped into other people's notebooks

The second half of this story concerns not the spammers but the people whose content gets taken. Along with the rename from NotebookLM to Gemini Notebook, the user agent of the fetcher that pulls pages on the user's command changed too: it is now called Google-GeminiNotebook. Search Engine Journal wrote that this is how your site can be loaded into someone else's notebook and retold with no attribution.

What to do about it, per the same Search Engine Journal piece:

1. Find requests with the user agent Google-GeminiNotebook in your server logs and see which pages are being pulled.
2. Write a disallow rule for that user agent in robots.txt, if you decide to shut it out.
3. Set up a block at the WAF level, that is the request filter sitting in front of your site, because it is more reliable than rules in robots.txt.
4. Decide deliberately: by shutting it out you also lose the chance of a mention, since a retelling can bring a person to you.

I would not shut everything off, I would split it: commercial pages and unique research are worth protecting, let the overview material get quoted.

What to do in an hour

1. Use the site operator on your own domain and make sure there are no other people's pages in the index from your open sections: profiles, comments, uploaded files. Any site that publishes without moderation can become a host for parasite SEO.
2. Take a unique phrase from your strongest piece, search it in quotes and see who is retelling your text on someone else's domain.
3. Open your logs and search for Google-GeminiNotebook, then decide whether to block it or not.
4. If you are investigating a traffic drop, remember that on 23 September 2026 Search Console again lost a day of data in the crawl stats report (for 15 September 2026): a hole in the chart does not always mean a problem on the site.

Of the 12,000 pages that caught a ride on Google's domain, nothing was left by 23 September 2026. Of those who spent those days calmly writing their own pages, no one was left with nothing.