Seen on the internet: https://a16z.com/geo-over-seo/
A new paradigm°hard word is emerging, one driven°hard word not by page rank°hard word, but by language models. We’re°hard word entering Act II°hard word of search°hard word: Generative°hard word Engine Optimization°hard word (GEO°hard word). ...
It’s no longer just about click-through°hard word rates, it’s about reference°hard word rates: how often your brand°hard word or content°hard word is cited°hard word or used as a source in model-generated°hard word answers. In a world°hard word of AI-generated°hard word outputs°hard word, GEO°hard word means optimizing°hard word for what the model chooses to reference°hard word, not just whether or where you appear in traditional°hard word search°hard word. That shift°hard word is revamping°hard word how we define°hard word and measure brand°hard word visibility°hard word and performance°hard word.
Already, new platforms°hard word like Profound°hard word, Goodie°hard word, and Daydream°hard word enable°hard word brands°hard word to analyze°hard word how they appear in AI-generated°hard word responses°hard word, track sentiment°hard word across model outputs°hard word, and understand which publishers°hard word are shaping°hard word model behavior. These platforms°hard word work by fine-tuning°hard word models to mirror brand-relevant°hard word prompt°hard word language, strategically°hard word injecting°hard word top SEO°hard word keywords°hard word, and running synthetic°hard word queries°hard word at scale. The outputs°hard word are then°hard word organized°hard word into actionable°hard word dashboards°hard word that help marketing teams monitor°hard word visibility°hard word, messaging°hard word consistency°hard word, and competitive°hard word share of voice.
Canada Goose used one such tool to gain°hard word insight°hard word into how LLMs°hard word referenced°hard word the brand°hard word — not just in terms of product°hard word features°hard word like warmth°hard word or waterproofing°hard word, but brand°hard word recognition°hard word itself. The takeaways°hard word were less about how users°hard word discovered Canada Goose, but whether the model spontaneously°hard word mentioned the brand°hard word at all, an indicator°hard word of unaided°hard word awareness°hard word in the AI°hard word era°hard word.
This kind of monitoring°hard word is becoming as important as traditional°hard word SEO°hard word dashboards°hard word. Tools like Ahrefs°hard word’ Brand°hard word Radar°hard word now track brand°hard word mentions in AI°hard word Overviews°hard word, helping companies understand how they’re°hard word framed°hard word and remembered by generative°hard word engines. Semrush°hard word also has a dedicated°hard word AI°hard word toolkit°hard word designed to help brands°hard word track perception°hard word across generative°hard word platforms°hard word, optimize°hard word content°hard word for AI°hard word visibility°hard word, and respond°hard word quickly to emerging mentions in LLM°hard word outputs°hard word, a sign that legacy°hard word SEO°hard word players°hard word are adapting°hard word to the GEO°hard word era°hard word.
This is a mix°hard word of cargo-cult°hard word marketing, and pure°hard word bullshit°hard word. The theory of the these companies is fatally°hard word flawed°hard word because of a few factors°hard word:
- the models use a dataset°hard word that is 9-12 months old. Whatever°hard word changes these companies make, won't show up immediately.
- there are no "traffic°hard word stats°hard word". The traffic°hard word stats°hard word that they provide°hard word have to be fake°hard word.
- the 30 "companies" listed include°hard word a lot that seem fake°hard word ⚙️ https://www.limy.ai and https://relixir.ai are two that are just "somebody°hard word who started an idea at YCombinator°hard word 3 months ago, and don't actually have anything°hard word sellable°hard word yet°hard word. Key links°hard word like "pricing°hard word" and "features°hard word" don't exist.
----
https://davefriedman.substack.com/p/large-language-models-are-not-search
Unfortunately°hard word, this rebuttal°hard word is also wrong.
LLMs°hard word are not indexes°hard word. They are statistical°hard word models of language, trained on enormous corpora°hard word to predict°hard word token°hard word sequences°hard word. There is no top 10 list inside GPT°hard word-4 or Claude°hard word. There is only a tangled°hard word web°hard word of parameter°hard word weights encoding°hard word the probability°hard word that, given a prompt°hard word, certain tokens°hard word will follow. Trying to optimize°hard word your brand°hard word’s presence in that is like trying to guarantee°hard word your reflection°hard word in a kaleidoscope°hard word. ...
What’s more, the entire°hard word underlying°hard word substrate°hard word is profoundly°hard word unstable°hard word. Even minor°hard word prompt°hard word rephrasings°hard word can dramatically°hard word alter°hard word which brands°hard word get mentioned. Change the context°hard word window by 10 tokens°hard word, or adjust the system prompt°hard word’s tone, and you might collapse entirely different parts of the model’s probability°hard word distribution.
If you want your "ChatGPT°hard word ranking°hard word" to be better, questions like How does temperature, top-p°hard word sampling°hard word, and prompt°hard word framing°hard word alter°hard word our probabilistic°hard word surface area across different LLMs°hard word? don't actually matter.
The entire°hard word argument is flawed°hard word. It is just this is too complex to understand, random means anything°hard word can happen, oogie-boogie°hard word.