AI Engine Optimization // Data Analysis

LLM vs. LLM+RAG

* Early versions of LLMs were based on next-word prediction, which is different from search

* More and more, AI chat and search use RAG+LLM

* RAG (retrieval augmented generation) starts with a search (retrieval), then summarizes/reforms the search as an answer (generation)

* Perplexity has built its own search engine, CoPilot frequently uses Bing, and OpenAI is building SearchGPT - so AI Search is moving towards RAG+LLM more and more

* Therefore, optimizing for AI Search starts with optimizing for search using traditional SEO strategies + is an evolution of new optimization strategies



AI Chat Traffic Analysis

* Method: We looked at anonymized traffic across several sites to assess the composition and growth rate of traffic and conversions from AI chat

* We are seeing a roughly 68% increase in traffic coming directly from ChatGPT since August 2024

* 87% of traffic is coming from ChatGPT, 7% from Perplexity, 4% Gemini

* Roughly 90% of visits go directly to the homepage and not to specific landing pages

* This is of note because a large % of times when publishers appear in AI chat, specific landing pages (not their homepage) are cited in answers

* This highlights that these numbers are likely undercounting the full impact of appearing in AI chat because it’s common for users to see an answer in AI chat, then open a new tab and go directly to that site, which means that visit is not attributed to AI chat

Keyword Research > Question Research

* "Question research" is the new keyword research

* Rather than finding keywords, instead we need to find all the ways people ask questions for our product

* We may target thousands of keywords, but questions vary more than keywords, thus there are millions of variations of questions

* Keyword research can use known data from Google and other keyword tools

* There are no tools to find what questions people ask - new strategies will need to be made to source how people ask questions

* Based on the questions people ask, we create landing pages and content to target questions similar to SEO



Good Content for Questions

* "Good content" in Google SEO is comprehensive and answers the questions users have for that page via the presence of TF-IDF terms, sub-topics, and embeddings

* A landing page that targets a topic in AI Search needs to first understand the thousands of questions users have for that topic, then answer as many of them as possible



Citation Optimization

* RAG+LLM performs a search, looks at multiple pages, summarizes them, and cites its sources

* Companies can optimize for being cited using SEO strategies similar

* Forbes and NerdWallet can try to optimize for being a citation for "best credit cards" across many variations of credit cards



SERP Tracking > AI Answer Tracking

* SEO tracks single positions for keywords (e.g. I rank #5 for "best credit card")

* AI Answer Tracking is a distribution or frequency across surfaces, question variants, and question runs