Your visitors search, read, refine, and search again. Most search box treat every question like the first one. Context is lost. The conversation resets. So they retype, rephrase, and start the whole thing over. Coveo conversational search solves this by carrying the conversation forward, grounding every answer in your content, keeping context across questions, and reasoning when no single piece of content has the answer. We'll show you how it works on Barca, a fictional marine manufacturer whose article, documentation and support content live across several systems. Let's start simple. A visitor asks what to check when maintaining a ballast tank. Coveo generates a grounded answer. Notice how it's pulling from separate articles to build that response. Every claim cites its source. Now a follow-up. The visitor asks, and what about in saltwater? Five words, no context repeated. The agent holds the conversation. It resolves the follow-up against the previous turn, retrieves for that resolved question, and generates a new answer from what comes back. And your rank results are still right there below. The follow-up carried the context. It didn't start a new search. Now the hard one. The visitor broadens the scope. Okay for ballast, but what about bolt and engine? I'm a complete beginner. There's no beginner friendly filter anywhere in this index. No tag, no attribute, nothing to match a keyword against. The agent has to reason across multiple articles to piece together a recommendation none of them gave on its own. It does and every step is cited. So how does this work? A question comes in and then it runs a loop. Understand the intent, reason about which content could answer it, retrieve the right content from across your sources down to the most relevant passages and evaluate what came back. If that isn't enough, it reasons again and searches again, each pass sharper than the last. Only when the content can support an answer does it handle passages to the model as grounded context and generate the answer. Every turn applies your permissions, your pipeline rules, and your filtering independently. And what the agent learned this turn is held in memory, so the next question builds on this one instead of starting over. That's retrieval with a reasoning loop behind it. Answering and following up is half the experience. The other half is trust. Every answer grounded, honest when the content runs out, and on brand. That's what the agent's rules are for. Here they are, exactly as the platform stores them for Barca Corporate. Every claim stays grounded in the retrieved content. No invented facts, products, or prices. When the articles don't answer, it says so and points to a person. Even the captain's voice is a rule, written once in the agent. Rules also let you tailor the experience. We give Barca Corporate one more. Start every answer with the captain's logline and end it with the safety reminder. Written on the agent, nothing on the site changes. Then the same first question: what to check when maintaining a ballast tank? Same articles, same facts. But the answer now opens with the logline and closes with the reminder: One rule written once and every answer follows it. In a few minutes, we've seen grounded answer with citations, follow ups that hold context without starting over, reasoning that pulls an answer together from several articles at once, and rules that Agent follows on every answer. Every one of them answered from the content Barca already had. And it's all set up through configuration in the admin console. No custom development.