{"channel":"llm","content":"Seen on the internet: https://a16z.com/geo-over-seo/\r\n\r\n<<< A new paradigm is emerging, one driven not by page rank, but by language models. We\u2019re entering Act II of search: Generative Engine Optimization (GEO). ...\r\n\r\nIt\u2019s no longer just about click-through rates, it\u2019s about reference rates: how often your brand or content is cited or used as a source in model-generated answers. In a world of AI-generated outputs, GEO means optimizing for what the model chooses to reference, not just whether or where you appear in traditional search. That shift is revamping how we define and measure brand visibility and performance.\r\n\r\nAlready, new platforms like Profound, Goodie, and Daydream enable brands to analyze how they appear in AI-generated responses, track sentiment across model outputs, and understand which publishers are shaping model behavior. These platforms work by fine-tuning models to mirror brand-relevant prompt language, strategically injecting top SEO keywords, and running synthetic queries at scale. The outputs are then organized into actionable dashboards that help marketing teams monitor visibility, messaging consistency, and competitive share of voice.\r\n\r\nCanada Goose used one such tool to gain insight into how LLMs referenced the brand \u2014 not just in terms of product features like warmth or waterproofing, but brand recognition itself. The takeaways were less about how users discovered Canada Goose, but whether the model spontaneously mentioned the brand at all, an indicator of unaided awareness in the AI era.\r\n\r\nThis kind of monitoring is becoming as important as traditional SEO dashboards. Tools like Ahrefs\u2019 Brand Radar now track brand mentions in AI Overviews, helping companies understand how they\u2019re framed and remembered by generative engines. Semrush also has a dedicated AI toolkit designed to help brands track perception across generative platforms, optimize content for AI visibility, and respond quickly to emerging mentions in LLM outputs, a sign that legacy SEO players are adapting to the GEO era. >>>\r\n\r\nThis is a mix of cargo-cult marketing, and pure bullshit. The theory of the these companies is fatally flawed because of a few factors:\r\n> the models use a dataset that is 9-12 months old. Whatever changes these companies make, won't show up immediately.\r\n> there are no \"traffic stats\". The traffic stats that they provide have to be fake.\r\n> the 30 \"companies\" listed include a lot that seem fake ( https://www.limy.ai and https://relixir.ai are two that are just \"somebody who started an idea at YCombinator 3 months ago, and don't actually have anything sellable yet). Key links like \"pricing\" and \"features\" don't exist. \r\n\r\n----\r\n\r\nhttps://davefriedman.substack.com/p/large-language-models-are-not-search\r\n\r\nUnfortunately, this rebuttal is also wrong.\r\n\r\n<<< LLMs are not indexes. They are statistical models of language, trained on enormous corpora to predict token sequences. There is no top 10 list inside GPT-4 or Claude. There is only a tangled web of parameter weights encoding the probability that, given a prompt, certain tokens will follow. Trying to optimize your brand\u2019s presence in that is like trying to guarantee your reflection in a kaleidoscope. ...\r\n\r\nWhat\u2019s more, the entire underlying substrate is profoundly unstable. Even minor prompt rephrasings can dramatically alter which brands get mentioned. Change the context window by 10 tokens, or adjust the system prompt\u2019s tone, and you might collapse entirely different parts of the model\u2019s probability distribution. >>>\r\n\r\nIf you want your \"ChatGPT ranking\" to be better, questions like << How does temperature, top-p sampling, and prompt framing alter our probabilistic surface area across different LLMs? >> don't actually matter.\r\n\r\nThe entire argument is flawed. It is just << this is too complex to understand, random means anything can happen, oogie-boogie >>.","created_at":"2025-07-06T18:45:41.272142","id":614,"llm_annotations":{},"parent_id":null,"processed_content":"

Seen on the internet: https://a16z.com/geo-over-seo/\r

\n

A new paradigm is emerging, one driven not by page rank, but by language models. We\u2019re entering Act II of search: Generative Engine Optimization (GEO). ...\r

\n

It\u2019s no longer just about click-through rates, it\u2019s about reference rates: how often your brand or content is cited or used as a source in model-generated answers. In a world of AI-generated outputs, GEO means optimizing for what the model chooses to reference, not just whether or where you appear in traditional search. That shift is revamping how we define and measure brand visibility and performance.\r

\n

Already, new platforms like Profound, Goodie, and Daydream enable brands to analyze how they appear in AI-generated responses, track sentiment across model outputs, and understand which publishers are shaping model behavior. These platforms work by fine-tuning models to mirror brand-relevant prompt language, strategically injecting top SEO keywords, and running synthetic queries at scale. The outputs are then organized into actionable dashboards that help marketing teams monitor visibility, messaging consistency, and competitive share of voice.\r

\n

Canada Goose used one such tool to gain insight into how LLMs referenced the brand \u2014 not just in terms of product features like warmth or waterproofing, but brand recognition itself. The takeaways were less about how users discovered Canada Goose, but whether the model spontaneously mentioned the brand at all, an indicator of unaided awareness in the AI era.\r

\n

This kind of monitoring is becoming as important as traditional SEO dashboards. Tools like Ahrefs\u2019 Brand Radar now track brand mentions in AI Overviews, helping companies understand how they\u2019re framed and remembered by generative engines. Semrush also has a dedicated AI toolkit designed to help brands track perception across generative platforms, optimize content for AI visibility, and respond quickly to emerging mentions in LLM outputs, a sign that legacy SEO players are adapting to the GEO era.

\n

This is a mix of cargo-cult marketing, and pure bullshit. The theory of the these companies is fatally flawed because of a few factors:\r

\n\n
\n

https://davefriedman.substack.com/p/large-language-models-are-not-search\r

\n

Unfortunately, this rebuttal is also wrong.\r

\n

LLMs are not indexes. They are statistical models of language, trained on enormous corpora to predict token sequences. There is no top 10 list inside GPT-4 or Claude. There is only a tangled web of parameter weights encoding the probability that, given a prompt, certain tokens will follow. Trying to optimize your brand\u2019s presence in that is like trying to guarantee your reflection in a kaleidoscope. ...\r

\n

What\u2019s more, the entire underlying substrate is profoundly unstable. Even minor prompt rephrasings can dramatically alter which brands get mentioned. Change the context window by 10 tokens, or adjust the system prompt\u2019s tone, and you might collapse entirely different parts of the model\u2019s probability distribution.

\n

If you want your \"ChatGPT ranking\" to be better, questions like How does temperature, top-p sampling, and prompt framing alter our probabilistic surface area across different LLMs? don't actually matter.\r

\n

The entire argument is flawed. It is just this is too complex to understand, random means anything can happen, oogie-boogie.

","quotes":[],"subject":"the opposite of wrong is still wrong"}