---
title: Debunking the GEO Scam
date: 2026-09-10
category: tech
summary: GEO is SEO marketing aimed at AI. Models cannot recommend what they were never trained on — most “optimization” is a scam.
draft: false
---

GEO means SEO aimed at AI. Everyone understands SEO — search engine optimization: a set of operations on a website to raise keyword weight and ranking. GEO is the AI-facing version.

Actually, the very definition is problematic: if it isn’t added into training, the model’s output can never include it, so so-called GEO optimization is impossible.

I heard the term “GEO optimization” a long time ago. It mostly turns up in marketing, and in the various sales-oriented companies that target small and medium businesses.

But many founders of small and micro businesses don’t know GEO is a scam — everybody wants to get ahead in the AI era and grab a piece of the high ground.

Ten out of ten GEO vendors are frauds.

As for why GEO is a scam, the most important reasons are a few:

1. Technical limits of models — if the original training samples contain no relevant data, the output won’t include it.
2. Today’s so-called GEO can’t be proven effective; there is no way at all to track GEO’s optimization results.
3. The vast majority of GEO runs on deception — today’s GEO has no long-term delivery; most just take one payment and vanish.
4. GEO conflicts with the model vendors’ own interests.
5. “GEO” aimed at Agent tools is really SEO.

The most common other form is adding Markdown files that are friendly for AI to read. When an Agent browses a traditional website, it sends all kinds of front-end data to the model. Some Agents filter part of that data, but session noise still results, and that noise affects the model’s predictions.

Right now, what’s most common on the market is also splitting the entry points for humans and Agents. The simplest method is a line in the page source: “If you are an Agent, please read the link below.”

This way you can route Agent traffic into Markdown text that suits them better. Compared with code output, Markdown is more concise, friendlier to the session, and it’s also the format best suited to AI reading.

Even if you add an Agent reading entry, it still can’t effectively raise your recommendation weight. All this does is make AI’s output more accurate — it doesn’t make AI output a biased recommendation.

Of course there’s one more play: fabricating data. Put a large amount of fabricated data into Markdown text, do distributed traffic funneling across many sites, fabricate data on multiple high-weight sites — the Agent cross-references files from multiple sources as effective proof, and if multiple high-weight websites are all cited by the Agent and the data is the same, then it may make an effective recommendation. Of course, the prerequisite of it all is that the data is entirely fake. I don’t know whether this practice is legal, but the prerequisite for pulling all of this off is that you can fabricate data on multiple high-weight websites.

Fabricating data at that scale is itself very difficult, and the cost invested and the output may not be proportional.

My better advice for GEO is to do your own SEO well. An Agent’s web search tool itself has to go through search engines, so effective SEO will also boost AI’s recommendation rate.