Agentic Search
Agentic Search refers to a search process in which an AI program directs the research itself: it formulates its own search queries, reads the results, notices gaps, and keeps searching until it can compile an answer. Instead of a list of hits, you get a result with source citations.
An ordinary search engine takes exactly one step. You type in words, and it shows a list of links. Everything else is manual work: reading, comparing, searching again if something is missing. Agentic Search shifts precisely this manual work into a computer program. The program breaks the question down into sub-questions itself, searches multiple times in succession, reads the pages it finds, and after each round decides whether it knows enough. In the end, it delivers a fully formulated answer instead of a list of links. The word “agentic” comes from the English word agent and here means: the program acts independently in multiple steps, without anyone dictating each individual step.
What classic keyword search cannot do
Many real questions cannot be answered with a single search query. One example: “Which German carmaker most recently lowered its electric vehicle targets for 2030, and by how much?” This question contains several hidden sub-questions. First you need to know which manufacturers even stated targets. Then you have to check for each one individually whether and when it changed them.
It is precisely these multi-step questions that explain why this technique is currently getting so much attention. Language models alone don’t help reliably here. A language model is a program that has learned to form plausible sentences from huge amounts of text. Its knowledge comes from training and is therefore outdated as soon as something changes. If it lacks a piece of information, it may, in doubt, invent something that sounds good. Experts call this hallucination.
Access to current sources defuses this problem but does not eliminate it. The model can misunderstand a source or favor a poor source. The practical benefit lies mainly in verifiability: good systems indicate, for every statement, where it came from. So you can check for yourself whether the claim is correct.
The loop of searching, reading, and following up
At its core, the same loop always runs. First, the system plans: it considers which sub-questions the task contains. Then it formulates search queries for these and sends them to a search engine or a database. The results are retrieved and read, often just the most important sections. The system then assesses what is still missing and starts the next round.
This loop can run five times or fifty times. This is exactly what distinguishes Agentic Search from the simpler variant RAG, short for Retrieval-Augmented Generation. With RAG, a search is performed once, what is found is appended to the question, and the model answers. Agentic Search is the version with feedback: the intermediate result determines what is searched for next.
The price for this is time and money. Every pass costs computing power, and the provider pays for that per request. A thorough piece of research can therefore take minutes instead of seconds. Moreover, errors can snowball: anyone who goes down the wrong track in round two ends up searching off-topic in round three. That’s why providers limit the number of rounds and build in checking steps.
From Deep Research buttons to enterprise research
The technique is most visible in chatbots. Features with names like “Deep Research” or “Research” in ChatGPT, Gemini, or Perplexity work on this principle. You ask a question, wait a few minutes, and receive a structured report with footnotes. Coding tools use this too: they search through an unfamiliar codebase until they have found the spot causing a bug.
In companies, the same loop runs on internal data. A lawyer has hundreds of contracts searched for a particular clause. An analyst has quarterly reports compared. For financial news, this is an important business area because time savings here are directly worth money.
For readers of news sites, there is also a second side to this. When AI systems read web pages and summarize the answer themselves, nobody clicks through to the source anymore. Publishers are therefore in dispute with AI providers over payment and access rights. Incidentally, a common misconception is that such systems “surf” the internet like a human would. Mostly, they simply call search engines and interfaces and process the text.