Five things, in this order: the system decides to search, it can get into your site, it retrieves you among thousands of pages, it places you among the few the model will read, and what it reads from you is useful for building the answer. A website worked on at each of those stages has better chances of being found, chosen and used by an AI.
The order matters because each stage depends on the one before it. In the test bed from Yonsei and Konkuk universities (KDD 2026), 5.8% of the pages tested —all chosen from among the ten that reach the model— dropped to rank 11 in reranking after being rewritten, and were left out. This article lays out that path and sends each stage to the article where we explain it. Each study measures one stage with its own test system, so the map does not promise results: it says what each stage does and what favors it.
It depends on the question. The documentation for Claude's web search tool (Anthropic) says it searches when the request depends on information that is current, changing or outside its training —among other cases, information about specific organizations, people or products that might have changed— and that it answers directly for stable knowledge, such as established facts or math. That description applies to the developer tool: whoever integrates it can steer it to search more or less readily.
A University of Toronto study (Chen, Wang, Chen and Koudas, EDBT/ICDT 2026) showed the other side with GPT-4o. For very well-known entities, what the model retrieves works as confirmation of what it already knew. For niche entities, the answer depends much more on what it finds. That is why what is specific to your business is where what you publish weighs the most; we explain it in Why a Hardware Store Shouldn't Compete With Wikipedia.
Before reading anything, the system has to be able to get in. This stage is not part of Nostos's analysis; we include it because without it the others do not happen.
OpenAI separates its crawlers by purpose. According to its documentation, sites that exclude OAI-SearchBot are not shown in ChatGPT search answers, although they can appear as navigational links, and it recommends allowing it in the robots.txt file. GPTBot, the one used to train models, is a separate setting: you can allow search and decline training.
Google explains in its guide that, to be a supporting link in AI Overviews and AI Mode, a page has to be indexed and eligible to appear with a snippet, and that there are no additional technical requirements or special markup. Among the practices it mentions is allowing crawling in robots.txt and in the CDN or hosting.
The IAB Tech Lab published in May 2026, for public comment, a guide for deciding which bots and crawlers to allow. It does not recommend blocking or allowing any one in particular: it argues that an all-or-nothing approach is not appropriate, and it proposes evaluating each crawler separately and observing before blocking.
If you want to see what happens to a machine that gets in but cannot read what you published, we cover it in Is Your Website Ready to Be Found?.
By comparing the words of the question with those of each field on your page: title, description, headings and markup. In the same Yonsei and Konkuk study, improving only those fields made a page appear among the top 20 candidates in 71% of queries, against a 58% starting point. Rewriting only the body left it at 53%.
The system is a purpose-built test bed —a word-based retriever, a reranker and a generator, over about 171,000 pages— and it does not reproduce any particular commercial engine. We explain it in Can Optimizing Your Text for AI Make You Invisible? and the practical side is in How Do You Write a Title and Description So an AI Retrieves Your Page?.
By being the one that best answers what was asked. After retrieval, a reranker picks the few pages handed to the model. In the Yonsei and Konkuk study, that reranker favored content that answered what the query was looking for, rewarded having the answer at the start of the page, and penalized content that widened the topic beyond that.
A University of California, Berkeley study (Wan, Wallace and Klein, ACL 2024) reached something similar with a different method. It pitted paragraphs with opposite answers against each other on controversial questions: adding scientific references or sounding more objective had a neutral or negative effect, and changes that brought the text closer to the question improved its win rate. We develop it in Why Does an AI Care More About Relevance Than Style?.
With what the body of your page says. In the Yonsei and Konkuk study, the vast majority of citations in the final answer came from the body text and not from the structural fields: the fields help you get found and the body helps you get cited.
On how to organize that body there is a preprint —a paper posted on arXiv, not yet peer-reviewed— from the universities of Tokyo, Tsukuba and Hiroshima and the National Institute of Informatics (Yu et al.). The authors reorganized the structure of 200 articles —heading hierarchy, paragraph length, lists and tables, emphasis— while preserving their meaning, and report that the citation rate went from 45.0% to 52.8% across six generative engines, 17.3% more. The detail is in How Do You Structure Your Content So an AI Cites It?.
That is five checks. None guarantees you will appear; each gives your site better chances at its stage.
Get your GEO score across eight academically-backed criteria and find out where your site stands with AI engines.
Run the analysis at nostosgeo.app