Definition
Agentic search is search carried out by an AI agent that plans its own queries, reads and compares pages, uses tools and keeps going until a task is finished. A classic AI search answer responds to one prompt; an agentic search session may run dozens of searches, open product pages, check availability and hand back a shortlist, a booking or a filled-in form.
By October 2026 the pattern is built into mainstream products. Google's AI Mode books restaurants in the US and, since April 2026, in eight more markets including the UK, Canada and India; in August 2026 it started hotel checkout for US users through partners such as Booking.com, Expedia, Hilton and Marriott. OpenAI moved its agent features into ChatGPT Work in July 2026, whose cloud browser reads pages, clicks buttons and fills forms on public websites, and it retired the standalone ChatGPT Atlas browser on August 9, 2026. Perplexity Comet has run on iOS, Android, Mac and Windows since March 2026. Deep research modes in ChatGPT, Gemini, Claude and Perplexity apply the same loop to long, cited reports.
Under the hood, the agent breaks the goal into sub-questions (a query fan-out repeated over several rounds), fetches pages, extracts facts and decides whether it has enough to act. Those fetches leave traces in server logs: ChatGPT-User, Claude-User and Perplexity-User for live retrieval, Google-Agent for agents running on Google infrastructure, and cryptographically signed requests from ChatGPT's cloud browser.
For GEO, the unit of visibility moves from a single cited answer to inclusion in the agent's working set: the pages it reads, the shortlist it builds and the option it finally picks. Agents favor pages where the facts they need sit in the HTML: prices, specifications, stock, return policies, opening hours, API limits. A vendor that hides pricing behind a demo form, or renders specs only with client-side JavaScript, tends to drop out of the comparison. This overlaps with agent experience optimization and, once money changes hands, with agentic commerce.
Measurement is harder than for a chat answer. The user may never see the sources the agent consulted, and the eventual visit can arrive as direct traffic or not at all. Practical signals are task-based prompt tests (give an agent a real buyer task and record which brands reach the shortlist), agent and fetcher requests in your logs, and conversions that follow spikes in those fetches. Our explainer on what agentic search is covers the practical side, and agentic search optimization covers the tactics.
Examples of Agentic Search
- A buyer asks ChatGPT Work to compare three payroll tools for a 40-person company. The cloud browser opens each pricing page, records per-employee costs and leaves out the vendor whose pricing is only available on a sales call.
- A UK user asks AI Mode for a dog-friendly Italian restaurant in Shoreditch for Saturday at 7 p.m. and gets bookable options through partners such as OpenTable and TheFork without opening a restaurant website.
- A coding agent searches docs, GitHub issues and changelogs for a breaking change before patching a project. The library with a clear migration guide stays; the one with a vague changelog gets swapped out.
- A GEO team scripts 30 real purchase tasks each month across Comet, ChatGPT and AI Mode and tracks how often its product makes the agent's final shortlist.
