AI search

Most AI search products combine a language model with a retrieval step. When a user asks a question, the system decides whether to answer from what the model learned in training or to search the web, pull selected pages and write an answer that draws on them, often with links. Products differ in how they decide, which sources they favor and how they display citations. Chat assistants such as ChatGPT, Claude, Gemini, Perplexity and Grok offer this to users directly, and search engines have added AI-written summaries to results pages; see AI Overviews.

Two consequences follow. Answers are generated, so the same question can return different wording and different sources on different runs. And the user may not click through, so the answer itself is the exposure.

A hypothetical example: a user asks an assistant which vendors are used for a certain kind of service. The assistant writes a short comparison and names a few organizations, citing two industry publications and a review site. Neither the organizations' own sites nor their press releases appear in the answer.

When AI search answers questions about an organization, it acts as a new intermediary between the organization and its audience. Its answers are built from what is published about the organization, including news coverage, owned pages and third-party sources. Being covered in the press does not by itself mean being present in AI answers, and PeakMetrics measures those as separate things. The free AI Perceptions check shows a snapshot for a brand from a subset of providers, and the AI Perceptions reports cover the topic in full. The glossary entries on AI visibility and generative engine optimization explain the measurement and the practice.

Because answers vary between runs, one observation is a snapshot and not a trend. Meaningful measurement requires repeated sampling with a stated method, providers, date and number of answers. It also requires attention to scope: results for reputational questions may not apply to product or shopping questions.

A useful test set covers the questions a real audience would ask about the organization: what it does, who runs it, how it compares with alternatives and whether specific controversies are true. Use the same questions over time and across several assistants so results are comparable. Record the date, the providers and the number of answers collected for each question.

AI search is not a replacement for traditional search everywhere; many people use both. It also does not mean web pages stop mattering. The pages that assistants cite are the sources of the answers, which raises the importance of accurate, well-sourced pages. And an assistant's answer is not an authoritative record: it can contain errors, as described under AI hallucination.