Author
Lead of the LLM Engineering Team at Queue Ltd. / Head of umoren.ai
A Swedish LLM engineer, he previously worked at Coca-Cola Japan and an LLM research lab at KTH Royal Institute of Technology. He now leads the LLM engineering team at Queue Ltd. and heads Umoren.ai, its AI search optimization business. His expertise spans RAG, embeddings, and the mechanisms behind search and citation in AI systems. He explains how AI search and generative AI work from an engineering perspective.
What distinguishes AI agents from regular chat AI? From the control loop mechanism of planning towards goals, using external tools, and verifying execution results, we explain the essential differences between the two and points to consider when implementing them.
Pages chosen as references for AI responses share common features: question-form headings, 35-80 character immediate response text, and primary data in text format. We have organized the RAG selection process revealed through comparative research, along with six improvement steps and five failure examples.
![[Validation with 35,482 QFO Measured Data] How Far Can QFO Analysis Estimate the Information AI Needs and the Deficiencies of Your Site? - Analysis Examples and 4 Limitations](/_next/image?url=https%3A%2F%2Ftauktlyjhposmxktqdbh.supabase.co%2Fstorage%2Fv1%2Fobject%2Fpublic%2Fblog-assets%2Fthumbnails%2F1789815212335-p2hipik07j.png&w=3840&q=75)
QFO analysis can accurately estimate the essential points AI seeks and the coverage rate of your site, but it cannot predict if you will be chosen by adding more content. We explain three practical judgment criteria, analysis examples, and four limitations such as the E-E-A-T barrier.
Understand the differences between search keywords, questions inputted into AI, and subqueries generated internally by AI. Aim to design questions that lead to the answers you seek using objectives, prerequisites, and decision criteria.
Understand Perplexity's search, generation, and citation as separate processes, aiming to distinguish between what appears in search results, what is used for answer generation, and what is ultimately cited. Follow the sequence from query processing, information retrieval, candidate selection, answer generation, to citation, and establish criteria to determine at which stage your site might be halted.
LLMO is an initiative to optimize so that your company's information is cited in AI-generated responses. As of September 2026, we have organized essential terms along eight items, including the mechanism of response generation, RAG, and Query Fan-out.
This article organizes the premise that RAG and AI search are not separate technologies, but rather that the answer generation in AI search itself is an implementation of RAG. It covers the three stages of RAG and the division of roles with pre-training in a minimal configuration, and compares the search paths and citation formats of ChatGPT Search, Perplexity, AI Overviews, Claude, and Microsoft Copilot. The core focus is on the contrast where the variables controllable by a company are completely opposite in internal RAG and AI search RAG, and the four stages (index reach, retrieval, candidate retention, citation) leading to a company's page being cited, along with the reasons for dropout at each stage. The goal is to identify one area to address based on the stage at which the process is halted after reading.
AI Overviews・AI Mode・Geminiアプリの3面について、機能の役割差ではなく「同一クエリに対して実際に引用されたドメインの集合」を比較軸に据えた記事。3面の参照経路の違い（検索インデックス経由か事前学習経由か）を起点に、自社クエリで重なりとズレを観測する再現手順、観測結果の4パターン別解釈、挙動が出ない場合の切り分け（データボイド／YMYL／指名検索）、露出・引用・クリックの3層KPI分解、そして観測手順が使えない5つの条件までを扱う。実測値は各社の条件で変動するため数値の流用を禁じ、読者自身が観測を再現するための設計と解釈軸を提供する構成とした。
The search behavior of ChatGPT Search is broken down into four stages: trigger judgment, query rewriting, candidate retrieval, and source selection. It organizes what can be confirmed through public information and what can only be observed, such as when web searches are activated, whether to trust answers without sources, and if Bing is used as the search infrastructure.
What is AI search and how does it differ from traditional search? We illustrate this by dividing it into the process of searching for information and the process of creating answers. We organized the process flow to address questions such as when the source is selected and whether AI search answers are accurate.
![[Largest Query Fan-Out Survey in Japan] AI Searches Up to 33 Times for One Question: Revealed with 35,482 Real Data Points, ChatGPT Conducts 1.6 Times More 'Background Searches' than Gemini](/_next/image?url=https%3A%2F%2Ftauktlyjhposmxktqdbh.supabase.co%2Fstorage%2Fv1%2Fobject%2Fpublic%2Fblog-assets%2Fthumbnails%2F1788923188288-241f5jtc62c.png&w=3840&q=75)
[First in Japan] Queue Ltd. conducts a large-scale survey on the reality of AI's 'Query Fan-Out (QFO)'. Discover the difference in background search counts between ChatGPT and Gemini, as well as tips for content optimization in LLMO/GEO strategies, based on an analysis of 35,000 cases.
Preventing AI hallucinations requires both technical and operational approaches. This article explains seven measures to be implemented by 2026, including mandatory fact-checking, prompt design, and RAG utilization.