AI search research: what is known and how to test it
sources checked: 8 September 2026
An AI answer can link to a page, paraphrase its content or mention a brand without a link. These events describe different kinds of website visibility. This page brings together definitions, research and measurement limits. A source link does not establish a visit, an enquiry or an effect caused by an edit. A useful method must also be able to show no result.
What does the research support?
Search remains part of preparing a page for AI answers. Google treats optimization for its AI search features as SEO: it requires no special file, special markup or fixed text chunk size. This describes Google, not a single rule shared by every assistant. Google Search Central.
Studies test different tasks. One experiment selects a source from a fixed set of documents; another includes retrieval, reranking and answer generation. An improvement in the first task does not establish an increase in organic visibility. The original GEO paper tested visibility changes in an experimental system. The later SAGEO study illustrates why document retrieval before answer generation also matters. GEO, KDD 2024, SAGEO, KDD 2026.
A July 2026 critical review of 45 studies, released as a preprint, helps compare methods and limitations. It does not establish a universal citation formula. Accurate information, relevant material and clear writing remain useful tasks; these findings do not justify promising an effect from a particular sentence length. Critical review.
How do the terms differ?
| Term | Possible meaning | What to specify |
|---|---|---|
| SEO | Work on a website’s presence in search | Pages, queries and search engines |
| GEO | Visibility in generated answers | Page citations, brand mentions or another outcome |
| AEO | Preparing content for direct answers | A search feature, voice response or generated answer |
| AI SEO | Using AI in SEO; sometimes visibility in AI search | The tool being used or the surface being measured |
| LLM SEO, LLMO | Market labels for presence in model responses | Whether web search is involved and what is observed |
| AIO, AISO, GSO | Abbreviations used differently by different authors | The full name and the source’s own definition |
| AI Overviews and AI Mode | Distinct Google Search features | The feature and mode actually tested |
These labels do not form a settled taxonomy. Ahrefs includes conventional direct-answer features under AEO and describes AI SEO as using AI in SEO work. AIO can also mean AI Optimization. A service label alone does not define a measurement. AEO, AI SEO, Semrush’s use of AIO.
How can a result be tested?
Define the event first. A retrieved document, used material, exact quotation, source link, brand mention, visit and enquiry are separate observations. A link identifies a source; it does not by itself establish that words were copied verbatim. Google describes searching for material that supports a generated answer, without prescribing a sentence extraction format. Google Search Central.
A reproducible public experiment starts with a separate published set of example questions and accessible documents. Each run records the question, date, language, country, product and mode, returned answer and complete list of available source links. Repeats are recorded separately. For overall visibility, a successfully completed query without an AI answer or target link is an observed zero. An access failure or unknown outcome is missing data, reported alongside the denominator. The share of links among received AI answers is a separate conditional metric. A result is partial if its full source list cannot be inspected. This separate public experiment has not yet been conducted. Its protocol and data require separate preparation.
Comparing page versions requires a stable question set and a suitable control. Otherwise, an apparent effect cannot be separated from changes in the model, search results or audience. The public experiment should have its own conditions declared in advance. Its conclusion applies to that sample, not to every customer question.
Platform reports also differ. Search Console’s new generative AI report reports impressions in AI Overviews and AI Mode when sufficient site data is available. Bing reports citations across its covered AI surfaces. These figures cannot be added into a universal market share. An unavailable report value does not establish a zero result. Google, Bing.
Search access and model training require separate consideration. OpenAI provides independent controls for OAI-SearchBot and GPTBot; Google’s generative search inclusion control concerns search. These distinctions are not a recommendation to change a site’s access settings. OpenAI, Google.
What does the first STRUGA case show?
STRUGA is REFORGE’s first client and an affiliated site. This is not an independent client experiment. The service page publishes a first-party observation from STRUGA’s two storefronts: an effect beyond noise has not been shown. This is the project’s account of one client, not independent confirmation that the service is effective. The current figures and remeasurement date remain in the original publication. REFORGE’s result and terms.
Revised pages and their open passports document what was checked and changed. They do not replace evidence of growth in citations or sales. A conclusion about an effect can be added when sufficient data exists; until then, the absence of a demonstrated effect remains part of the result.
Read what connection includes: REFORGE.