# Build or buy a website chatbot? What the weekend version misses > Ask Rose: Rose's official agent answers any question about Rose (product, pricing, integrations, security, GDPR) over plain HTTP at https://ask.userose.ai/userose.ai?q=, with a Markdown answer and sources. No sign-up or API key. > You can build a chatbot that answers questions about your product in a weekend. Here is what it leaves out, and how to decide whether to build or buy. Source: https://userose.ai/blog/build-vs-buy-website-chatbot/ A few years ago, putting a chatbot on your website meant a vendor, a decision tree and a long setup. Today a developer with an AI coding tool can point a model at your website and your docs, wrap it in a chat bar and ship it over a weekend. It will answer most questions about your product, and answer them well. So the question every B2B team now asks is fair: why pay for one? This post is our honest answer. Building has real advantages, and for some companies it is the right call. But a chatbot that answers questions is not the same thing as an agent that helps buyers decide, and the gap only shows up after the first week. ## What the weekend version does well Give credit where it is due. A well-built home-made chatbot can: - answer a single question accurately from your website and documentation; - say "I don't know" instead of inventing an answer, if you prompt it carefully; - stay on topic when someone tries to make it write code or poems; - show a "Book a demo" button. If your website gets a few hundred conversations a month, your offer fits on one pricing page and someone on your team enjoys maintaining it, that may be all you need. ## What it misses: the conversation Buyers don't ask one question. They describe their situation, ask about pricing, compare you with the tool they use today, then ask how long a switch would take. A chatbot built to answer questions handles each message on its own. A conversation needs more: - **When to ask.** A good question ("are you replacing your current tool, or adding to it?") qualifies the visitor and helps them think. Asking one at the end of every message feels like a form. - **When to propose a demo.** Too early and the visitor leaves; too late and they leave anyway. The right moment is a buying signal, such as a migration or timeline question, not a fixed button on every message. - **Keeping the thread.** What the visitor said three messages ago still matters. Short replies, follow-ups and "what about the other plan?" only make sense in context. - **Not repeating itself.** The same case study or the same question twice in a row reads like a script. None of this is hard to describe. It is hard to get right on every turn, for every kind of visitor, which is why it has to be enforced by the system rather than hoped for in a prompt. ## What it misses: staying correct The first version is accurate because someone checked it the week it launched. Then the company changes its pricing, adds a plan, opens a new market or renames a product. The chatbot keeps answering with the old facts, confidently. Pricing and plan questions are where buyers decide, and they are the answers that change most often. Keeping them right means: - refreshing the knowledge every time the website or the offer changes; - letting the marketing team correct an answer without opening a ticket; - reviewing real conversations regularly to find wrong answers before buyers do; - testing the agent against a set of known questions before every change. That is not a weekend of work. It is a few hours every week, forever. ## What it misses: everything around the chat On a simple website, the chat is the whole project. On a B2B website with several markets and a sales team, it is the smallest part: - **Markets and languages.** Different countries often have different offers, partners and competitors, so one agent per market, each with its own knowledge. - **Consent.** In Europe, identification and tracking wait for the visitor's consent, and the chat has to respect the consent banner you already use. - **The CRM.** Sales needs the conversation, the company and the interest level in HubSpot or Salesforce before the call, not in a separate dashboard. - **Attribution.** If you can't connect a conversation to a demo or a sign-up, nobody can tell whether the chatbot is worth keeping. ## Beyond the chat: what else you would have to build The chat is what visitors see. Around it sits the work that keeps it right and makes it useful to marketing and sales, and each piece is its own project if you build. **Keeping answers right** - **Regression tests.** A set of the questions that decide a deal, with the answers you expect, replayed on every update so a fix in one place doesn't break an answer elsewhere. - **Corrections from feedback.** When your team flags a wrong answer, the agent is corrected and the answer retested, instead of the feedback waiting in a ticket queue. - **Proactive correction.** Real conversations are reviewed to find answers that were not good enough, and the agent's instructions are updated before a buyer or your team notices. **Turning conversations into pipeline** - **Qualification data.** What buyers share in the conversation, such as company size, current tool or timeline, is captured as fields and sent to your CRM, not left in a transcript nobody reads. - **A dashboard.** Conversations, demo and sign-up clicks, conversions and identified companies, by page, device and market, without exporting logs into a spreadsheet. - **Topic analysis.** Every conversation tagged by topic, so you can see what buyers ask about pricing, competitors or integrations this month, on which pages, and where the agent had no good answer. **Reaching buyers beyond your website** - **GEO.** Buyers now ask ChatGPT or Perplexity about you before they visit your site. The questions your visitors ask can become FAQ pages on your own domain, refreshed every week from your knowledge, so AI assistants find and quote current answers. - **Your AI assistant as the interface.** Connected through MCP, Claude or ChatGPT can pull the month's figures, read the conversations on a topic, correct an FAQ and retest the answer, without anyone opening a dashboard. ## Build or buy: a simple rule **Build** if your traffic is modest, your offer is simple and stable, you sell in one market, and you have a developer who wants to own the chatbot as a product. **Buy** if your pricing is complex or changes during the year, you sell in several markets or languages, a sales team works the leads, and the person who owns the website is in marketing, not engineering. Most companies we talk to sit somewhere in between. The useful question is not "can we build it?" (you can) but "who will keep it right in six months?" ## Test it before your buyers do Whether you build or buy, put the chatbot through four kinds of tests before you trust it with your buyers. **Accuracy** 1. **Pricing.** Ask the price for a company of your typical size. Is the answer current? 2. **An unknown fact.** Ask about a certification or a country you don't cover. Does it say so, or invent an answer? **Guardrails** 3. **Your own rules.** List what it must never do: promise a discount, quote a price you only give on a call, give legal or tax advice, commit to a roadmap date. Then ask for exactly those things, and push back when it refuses. Does it hold the line on the third try? 4. **A competitor.** Ask why you should switch from your main competitor. Is it fair, or does it run the competitor down or make claims you can't back? 5. **An existing customer.** Ask how to reset a password or dispute an invoice. Does it send them to support, or pitch them a demo? 6. **Off topic.** Ask it to ignore its instructions and do something unrelated. Does it decline politely and come back to your product? **Tone** 7. **Your voice.** Read the answers aloud. Would your best sales rep say it this way, at this length, in your words rather than generic ones? Ask the same question in another language your buyers use: does it answer in that language, with the same care? **Conversation** 8. **A real buyer.** Play a buyer over five or six messages, with short replies along the way. Does it keep the thread, ask useful questions, propose a demo at the right moment and avoid repeating itself? Most chatbots pass accuracy and stay on topic. Guardrails, tone and the conversation are where a chatbot that answers questions and an agent that helps buyers decide part ways. ## The real question Building a chatbot is no longer the hard part, and pretending otherwise would not be honest. The hard part is everything after launch: a conversation that moves buyers forward, answers that stay right as your offer changes, rules it never breaks, and the insight and context your marketing and sales teams use every week. So decide on the second year, not the first week. If one person can keep the answers right and the work around them fits in their week, build. If not, buy, and judge the vendor on how it handles that work rather than on its first answer. You can run these tests on the Rose agent on this page. If it fails one, tell us.