What Is Agentic Search? How AI Agents Search, and What It Changes for B2B SaaS (2026)

In short: agentic search swaps one lookup for a loop of search, read, judge and search again, and the AI model runs the loop instead of the person. Classic one-shot RAG retrieves a fixed set of text once and answers in a single pass. On B2B SaaS buying questions the loop is already the default: in LoudFace's 14-day Peec AI read of 96 non-branded questions, ChatGPT searched the web in 96.8% of answers, Perplexity in 100% and Google AI Overviews in 96.5%. For a SaaS brand, that means the shortlist is built from whatever pages the agent pulls in that run.

What did the engines actually do on 96 B2B SaaS buying questions?

Read window and sample: the LoudFace project in Peec AI (our first-party data), 96 non-branded B2B SaaS agency-buying questions, 4 questions naming LoudFace dropped, one answer per question per engine per day from 2026-10-30 to 2026-11-12, which gives 1,344 answers per engine, pulled live on 2026-11-13. On these buyer questions the engines searched the web almost every time, so on commercial questions agentic search is the default, not an edge case. Each engine is reported on its own, because the three behave too differently to average.

Two things in that table matter more than the definition. Only ChatGPT shows a real fan-out, and it never searched with the buyer's own words. And the pages an engine reads are not the pages it credits: in our sample ChatGPT cited 13.5% of the pages it listed, Google AI Overviews 95.0%.

How does agentic search work?

Every product built on agentic search runs some version of the same loop. Mistral's definition is the cleanest: "Agentic Search introduces a multi-step retrieval loop for finding, inspecting, and verifying information across data sources." On arXiv, one survey describes models that "plan, retrieve, and reflect through multi-step interaction with search environments" (arXiv 2510.16724), and another says "a search agent iteratively plans and acts to gather information" and updates its plan after each step (arXiv 2508.05668).

In practice the loop has five moves:

1. Plan. The model breaks the question into sub-questions. The pricing example in Microsoft's Azure AI Search documentation assumes about three subqueries per query plan for its agentic retrieval.
2. Search. It writes its own queries and sends them to a search index or search provider, often several at once.
3. Read. It opens results, reads pages or chunks, and keeps what looks relevant.
4. Judge. It checks whether the evidence answers the question or leaves a gap.
5. Search again, or stop. If there is a gap, it writes a new query and repeats. If not, it answers, cites, or takes an action.

Step 5 is where the agent's reading steers its searching. In 'Agentic Search in the Wild', a SIGIR 2026 study of 14.44M search requests sent by agents to a research search API, "54% of newly introduced query terms appear in the accumulated evidence context". Most of the new words in the agent's next query come from what it has read so far, including pages from earlier steps. The same study found "over 90% of multi-turn sessions contain at most ten steps", so most loops are short.

Mistral and Microsoft use the term for searching a company's own documents. The marketing world uses it for consumer AI engines that search the open web. The mechanics are the same, and the second meaning is the one that decides whether buyers hear about your product.

Agentic search vs RAG vs traditional search: what's the difference?

Mistral puts the RAG side plainly: "Traditional, one-shot RAG" "retrieves a fixed set of text chunks and asks a model to answer in a single pass." The difference that matters for a brand is the middle row. In traditional search a person picks which result to click. In agentic search the model picks which pages enter the answer, and the person may never see the rest.

Which AI products run agentic search in 2026?

All the big assistants do, at different depths. Product names change fast here, so each line carries its source's date.

So is ChatGPT an agentic AI? Its search mode is agentic by this definition: it decides when to search, rewrites the question, and can search again. Deep research and ChatGPT Work go further, running longer loops and acting in a browser. For engine-by-engine detail, see where ChatGPT gets its information, how Perplexity works, what Google AI Mode is, how Gemini works and how Claude works.

How do an agent's searches differ from the buyer's question?

The agent searches in its own vocabulary. On our 96 buying questions, none of ChatGPT's 2,315 queries repeated the buyer's question word for word. 25.3% of its queries added a year the buyer never typed, and 36.7% added GEO or generative-engine wording the buyer never used. Its median query ran 11 words against 9 for the question (LoudFace Peec data, same window).

Here is one real run, from 2026-10-30. The buyer asked ChatGPT: "Best agency for getting mentioned in Perplexity and ChatGPT". ChatGPT searched three times:

• "best generative engine optimization GEO agencies ChatGPT Perplexity 2026 case studies"
• "agency AI search visibility optimization ChatGPT Perplexity clients case studies GEO"
• "top GEO agencies generative engine optimization independent reviews 2025 2026"

In our Peec data, that run listed 30 pages and cited none of them. The buyer said "getting mentioned". The agent searched for "GEO", "case studies" and "independent reviews". A page that answers the buyer's phrasing but never uses the agent's words can miss the retrieval set entirely.

Perplexity behaves differently. In the same data, its one visible query was the buyer's question verbatim in 99.7% of cases. Google AI Overviews shows no queries at all, though Google says AI Overviews and AI Mode "may use a 'query fan-out' technique".

Our data also changes how you should track this:

• The fan-out is not stable. In our data, on the same question on consecutive days, ChatGPT ran an identical set of queries in 10 of 565 pairs (1.8%). Tracking one prompt once tells you about one run.
• Published fan-out rates disagree, and the prompt mix is a likely reason. Ahrefs found that "Only 14.4% of the 66.6 million ChatGPT responses contained fan-out queries", averaging about 1.6 each, across mixed logged-out prompts. Peec AI's research on 5 million fan-outs reported "ChatGPT: 2.1 fan-outs per prompt" and 1.4 for Perplexity. On commercial buying questions, ChatGPT showed queries in 67.7% of our answers. In our sample, buying questions showed queries far more often than the mixed logged-out prompts Ahrefs measured, though the two studies did not capture fan-out the same way.

The academic logs point the same way. 'Agentic Search in the Wild' found "Agents instead phrase queries as full constraint-bearing questions", and 'A Picture of Agentic Search' (arXiv 2602.17518) found "approximately 16% of organic queries contain spelling errors" against about 5% of agent queries. Agents write cleaner queries that carry more constraints than people's do. On length the evidence is mixed: 'A Picture of Agentic Search' found agent and human queries of comparable length on most of its benchmarks, while ChatGPT's queries in our data ran longer than the buyer's question. More on this in fan-out queries.

Why does an agent read pages it never cites?

Because listing and citing are separate steps, and each engine weights them differently. In our data the median searched answer listed 15 pages and cited 2 on ChatGPT, listed 10 and cited 3 on Perplexity, and listed 6 and cited 6 on Google AI Overviews. In the same data, 23.7% of ChatGPT's searched answers cited nothing at all.

Our own traced answers show what this means for a brand. First, an engine can only name a brand that sits on a page it retrieved in that run; we were named only when one of our pages was in the sources, while other agencies were named from third-party pages (we wrote this up in AI engines name brands they read). Second, when a brand is left out, usually no retrieved page mentions it at all, so the engine had nothing to name.

So the work comes in order. First get into the retrieved set for the queries the agent writes, through your own pages and through the third-party lists it pulls. Then be the page it chooses to cite, which on ChatGPT means beating 13 or so other pages in the same answer.

Do agents check a vendor's own website?

ChatGPT does, as a separate hop. 17.7% of ChatGPT's queries in our window were aimed at a named vendor's own site, phrased like "[Agency] official website", in 19.3% of the answers that showed queries. In the clearest run we traced, for "Agencies like Animalz but with AI search optimization", ChatGPT ran two category searches to build a shortlist, then five queries of the form "[Agency] official website" plus a GEO or AEO term to check what each shortlisted agency says about itself. It listed 82 pages and cited 1.

That second hop is the moment your own site has to answer the agent's question in the agent's words. If your homepage and service pages describe what you do in different vocabulary from the category search, the agent finds the page and still has little to quote.

Who visits your site on an agent's behalf?

Three kinds of bots, and they follow different rules.

The user fetchers are the agents in action: a person asked a question and the assistant went to read your page right then. Cloudflare found that in July 2025, OpenAI's "ChatGPT-User bot is responsible for nearly three quarters of the request traffic" among user-action crawlers. Those visits sit in your server logs. When an AI user fetcher requests a URL that returns a 404, it is telling you what a buyer's agent expected to find on your site. Our server log playbook covers how to read them.

What does agentic search do to clicks?

Fewer people click through, and the clicks that remain lean toward the brands the AI cites. In March 2025, Pew Research Center tracked 900 US adults and found that users who saw a Google AI summary "clicked on a traditional search result link in 8% of all visits", against 15% without one; a link inside the summary got a click in 1% of visits. Across 300,000 keywords, Ahrefs found an AI Overview "correlates with a 58% lower average clickthrough rate for the top-ranking page". The cited side gains: Seer Interactive found brands cited in an AI Overview get "35% more organic clicks and 91% more paid clicks" than when they are not cited. These studies cover Google's AI Overviews only, but for a SaaS team the lesson carries: being the page the answer cites now counts for more than holding the top organic slot beneath it.

What is agentic search optimization?

Agentic search optimization is the work of making a brand findable and citable for AI agents that search on a buyer's behalf. The industry also calls it GEO (generative engine optimization), AEO (answer engine optimization) or AI search optimization. In LoudFace's Peec data, ChatGPT's own queries added GEO or generative-engine wording the buyer never used in 36.7% of cases, and AEO wording in 9.7%, so the agent's vocabulary leans GEO.

It is not a separate discipline from SEO. Google's own AI optimization guide says "optimizing for generative AI search is optimizing for the search experience, and thus still SEO", and that Google Search needs no special AI text files. llms.txt is still "A proposal to standardise on using an /llms.txt file" and has no standing as a standard (more in llms.txt explained). At LoudFace we treat agentic search optimization as part of our existing SEO and AEO program, where SEO, AEO and GEO run as one piece of work.

What changes is the emphasis. Because the agent writes its own queries and then checks your site in a second hop, the work shifts toward the words it searches with and the pages it retrieves and lands on.

What should a B2B SaaS team change?

1. Write for the agent's query as well as the buyer's prompt. Use the words the engines search with in titles, H1s and headings: GEO, AI search optimization, the year. Check them against real fan-out queries rather than guessing.
2. Answer the question in the first lines. Agents search with full questions and quote the top of a page. A self-contained short answer and a clear title are the surface engines lift from.
3. Make your own site pass the vendor check. Your homepage, service and pricing pages should state what you do, for whom and with what proof, in category vocabulary, because ChatGPT's second hop lands there.
4. Get onto the third-party pages agents retrieve. For generic category questions the agent builds its shortlist from lists other people wrote. Your own page alone does not put you in those answers.
5. Read your logs for user fetchers. ChatGPT-User, Perplexity-User and Claude-User hits show which pages agents read for real buyers, and their 404s show pages you should have.
6. Measure per engine, over many runs. ChatGPT, Perplexity and Google AI Overviews search, read and cite differently, and ChatGPT's fan-out rarely repeats. One blended score hides the engine you are losing on.
7. Watch Google AI Overviews first for new pages. It sits on Google's live index and picks up a new page fastest; earning a steady citation slot takes weeks, and owning a competitive prompt takes months.

An agent builds its shortlist fresh on every run from the pages it retrieves, so a brand that is missing from those pages for the agent's own queries is missing from the answer.

For a baseline, our free AI visibility audit checks where your brand stands across ChatGPT, Claude, Gemini and Perplexity searches against your top competitors, and you get one fix you can implement within a week.