I recently spoke with Jesse Dwyer of Perplexity about SEO and AI search, focusing on what SEOs should prioritize when optimizing for AI search. His insights provided valuable guidance for publishers and SEOs today. ### AI Search Today Jesse emphasized that personalization is fundamentally transforming AI search. He noted, “The biggest and simplest point about AEO (Answer Engine Optimization) vs SEO is that it’s no longer a zero-sum game. Two users with the same query can receive different answers if the AI tool incorporates personal memory in its context window (as with Perplexity, ChatGPT). ” This shift stems from differences in indexing technology, explaining the divergence between traditional Geographic Engine Optimization (GEO) and AEO. Nonetheless, most traditional SEO best practices still apply. Dwyer’s core message is that search visibility is no longer about consistent results across users. AI answers are shaped by personal context, meaning two users may see distinctly different responses and underlying content sources for the same query. Although the infrastructure remains a classic search index—Perplexity AI employs a PageRank-like system evaluating site popularity and relevance—what is retrieved differs markedly from traditional search. When I asked if classic search typically displays the same top sites per query while AI search generates varied answers due to conversational context, Jesse confirmed this distinction. ### Sub-document Processing: The Key Difference in AI Search Jesse further explained the pivotal technology behind AI search: the contrast between whole-document and sub-document processing. Traditional search engines index entire webpages, scoring and ranking them as single units.

AI tools using this model—like ChatGPT’s web search—fetch the top 10–50 documents and then use a large language model (LLM) to summarize results. This approach, sometimes dubbed “4 Bing searches in a trenchcoat, ” still relies on algorithmic search rather than true AI retrieval. This method aligns with GEO or Generative Engine Optimization, still dependent on whole-page scoring. The emerging AI-first approach—referred to as “sub-document processing”—indexes and retrieves highly specific, granular snippets instead of full documents. In AI terms, a snippet consists of approximately 5–7 tokens (2–4 words) converted into numerical representations by the transformer architecture underpinning models like GPT. Upon a query, the system pulls roughly 130, 000 tokens’ worth of the most relevant snippets (around 26, 000 snippets) into the model’s context window—the total capacity of most current LLMs—with the goal of saturating this window with pertinent information. Saturating the context window minimizes the model’s tendency to “hallucinate” or fabricate information, enabling it to produce more accurate answers. This sub-document indexing is where the industry is heading, justifying the term AEO. Furthermore, personalization arises because the LLM can incorporate extensive data about the user into the full context window—much beyond a traditional Google user profile—producing uniquely tailored results. A company like Perplexity differentiates itself by optimizing the pipeline between the index and snippet retrieval using methods such as compute modulation, query reformulation, and proprietary models, which collectively enhance snippet relevance and answer quality. While less critical for SEOs, these advancements also explain why Perplexity’s search API is notably advanced for developers integrating search in other products. ### Summary of Indexing Approaches Dwyer highlighted the contrast between: - **Whole-document indexing:** Ranking and retrieving entire web pages, with AI summarizing the final set. - **Sub-document indexing:** Storing and retrieving meaning as fine-grained fragments, with AI reasoning directly over these snippets. The latter approach improves answer accuracy by fully saturating the model’s context window with relevant data, limiting fabrication, and producing richer responses. In closing, Jesse emphasized that “modulating compute, ” reformulating queries, and proprietary indexing models form their “secret sauce” for superior snippet retrieval and AI search quality. **Featured Image** by Shutterstock/Summit Art Creations