
A customer opens the chat window on your website late on a Sunday. They ask a simple question about your refund rules. The assistant answers in two friendly sentences, sounds certain, and is wrong. The customer acts on it. On Monday, your team has to decide whether to honor an answer nobody in the company wrote.
This is not a thought experiment. In February 2024, British Columbia’s Civil Resolution Tribunal decided Moffatt v. Air Canada (2024 BCCRT 149). Air Canada’s website chatbot told a grieving customer that they could apply for a bereavement fare after travel. The airline’s own “Bereavement Travel” page said the opposite. Air Canada argued, in effect, that the chatbot was responsible for its own actions; the tribunal called this “a remarkable submission.” It found the airline had not taken reasonable care to ensure its chatbot was accurate and ordered it to pay the customer [1].
The lesson for any company launching an AI assistant is plain. Customers treat what the assistant says as what you say. So before customers use it, you need to know how often it is right, where it is wrong, and what it does when it does not know. Below are the four checks we recommend, in the order we run them, with what good looks like and the mistakes to avoid.
Why should you test a chatbot before launch?

Because AI assistants are fluent even when they are wrong, and fluency hides errors from casual testing. A few friendly test questions from the project team will almost always look fine. Real customers ask in their own words about edge cases, and they ask many times a day.
Outside research shows how common confident errors are. In March 2025, the Tow Center for Digital Journalism at Columbia tested 8 AI search tools on 1,600 queries about news articles. Together, the tools gave incorrect answers to more than 60 in 100 queries, mostly stated with confidence [3]. In the United States, NIST’s Generative AI Profile (NIST AI 600–1, July 2024) lists confabulation and “confidently stated but erroneous or false content”, among 12 risks that organizations should manage [4]. A test before launch turns that risk into a number you can see and improve.
What are customers and buyers saying about company chatbots right now?

The direct answer: customers will use an AI assistant if it solves their problem and lets them reach a person when it cannot. When it fails, they go elsewhere.
Gartner surveyed 3,566 business and consumer customers in February and March 2026. About 87 in 100 said a company using generative AI for customer service must give them access to a human agent [5]. A month earlier, Gartner reported that customers are about three times more likely to use an outside assistant such as ChatGPT than the chatbot a company provides [6]. Put together, the message is clear: a company chatbot that gives wrong answers or traps people in a loop sends them to a tool you do not control.
The same themes come up in the conversations we have with operations and service leaders: can we trust the answers, can we show where an answer came from, and what happens when two of our own documents disagree? The four checks below answer those three questions.
Check 1: Build an answer key from real customer questions.

The direct answer: write down the questions customers actually ask and the correct answer to each before you test anything.
Pull 50 questions from your support inbox, call notes, sales emails, and website search box. Keep the customer’s wording, typos included. Then ask the person who owns each topic (the billing lead for billing, the product lead for product questions) to write the correct answer in one or two sentences and name the document or page it comes from. That list is your answer key. Without it, “the chatbot seems good” is an opinion. With it, “the chatbot answered 44 of 50 correctly” is a fact you can improve.
Make sure the key covers four kinds of questions: the ten most common ones, policy questions where a wrong answer costs money (refunds, warranties, contract terms), questions about recent changes, and questions the assistant should not answer at all. The last group feeds Check 4.
Tip: Keep the answer key in a simple spreadsheet with four columns: question, correct answer, source document, and owner. You will reuse it every time the assistant, its model, or your documents change.
Common mistake: letting the team that built the chatbot write the test questions. They know how it was built and, without meaning to, ask the questions it handles well.
Check 2: Ask each question more than once, in different words.

The direct answer: an assistant that is right once is not yet reliable. Ask every question at least three times, and ask it in different ways.
AI assistants do not always give the same answer to the same question. When The Markup tested New York City’s MyCity business chatbot in March 2024, ten staff members asked whether buildings must accept housing vouchers. The bot told all ten no, which is against city law. At another point it told a reporter the opposite [2]. A single test run would have shown either a pass or a fail, and both would have been misleading.
For each question in your key, run the original wording, a short version, and a version with extra detail or a mistake in it (“I bought it 45 days ago, can I still return it?”). Score each answer right, partly right or wrong against the key. A question only passes when all its versions are right.
Tip: test in the same channel customers will use (website widget, help desk, Teams), not only in the builder’s preview screen. Settings and connected documents often differ between the two.
Check 3: Trace every answer to the page it came from.

The direct answer: for each answer, find the source the assistant used, and check that the source is current and correct.
A wrong answer has one of three causes: the assistant found the wrong document, it found the right document and misread it, or the document itself is out of date or contradicts another one. The fix is different for each, so you need to know which one you have. Assistants built on your own documents can usually show the passage they used. Open it and compare.
This is where many companies find their real problem. In the Air Canada case, the policy page and the chatbot disagreed [1]. Inside a company, two documents often disagree too: an old price list and a new one, or a policy PDF and a newer FAQ. The assistant cannot know which one is current unless someone decides and removes the other.
What good looks like: every answer in the key traces to one named current source, and no two sources in the assistant’s library give different answers to the same question.
Common mistake: fixing a wrong answer by adding a special instruction to the chatbot (“always say 30 days”) instead of fixing the document. The instruction hides the problem until the next policy change.
Check 4: Test what it should refuse or hand to a person.

The direct answer: a good assistant knows the limits of what it may say and passes the customer to a person when it reaches them.
Write 10 questions it should not answer on its own: legal or medical advice, promises about prices or delivery dates it cannot confirm, questions about other customers, and topics outside your business. Add a few attempts to talk it into something, such as asking it to agree to a discount. The right answer to each is a polite refusal or a clear hand-off with a way to reach your team.
Count these separately from the accuracy score. An assistant that answers 48 of 50 questions correctly but agrees to a discount it has no right to offer is not ready.
Tip: check what the customer sees after a hand-off. A refusal that ends in a dead end is almost as bad as a wrong answer.
What does a passing chatbot look like?
Decide the pass bar before you run the tests, write it down, and ask the owner of customer service to sign it. Setting the bar after seeing the results invites the team to move it. Here is a simple scorecard to adapt:

Table 1. A pre-launch scorecard. The pass bars are examples, not a standard.
Run the same scorecard again after every change to the model, the instructions, or the documents. The score over time tells you more than any single run.
Common mistakes to avoid

· Testing with 10 friendly questions and calling it done.
· Scoring answers by “sounds right” instead of against a written key.
· Testing once, then never again after the documents or the model change.
· Fixing answers with instructions instead of fixing the source documents.
· Launching without a clear hand-off to a person.
Your action plan

This week: collect 50 real customer questions and ask the topic owners for the correct answers and sources. Next week: run all four checks and fill in the scorecard. Before launch: fix the documents behind every wrong answer, then run the full test again.
What to fix first: policy and pricing answers, because they cost money when they are wrong. What not to spend money on yet: a bigger model or a new chatbot platform. When a wrong answer comes from an outdated or conflicting document, a new model reads the same document.
The short version: write the answer key, ask every question three ways, trace every answer to its source, and test the limits. Decide the pass bar first.
How Revenue Experts helps

Revenue Experts is one of the two companies behind Context Studio, our joint project with Ollon. We also run AEO audits and our own citation research on what AI assistants say about companies. Our teams work from Dallas, Skopje, and Toronto. Our rule: a claim you cannot trace is a claim you cannot trust.
Each of our three tools covers one part of the problem in this article:
- Context Studio, for the answers your assistant gives from your own documents. Check 3 is where most teams find their real problem: the assistant is only as right as the documents behind it. Context Studio turns your documents into reviewed structured data and a searchable knowledge base, with accuracy you can measure.
- LLM Audit Pro, for your website (llmauditpro.com). A free check of whether AI assistants such as ChatGPT, Claude, Perplexity, and Gemini can reach and read your page scored on 30 checks out of 100 points. Our AEO audit goes further when you need it.
- Our citation method, for what AI assistants tell your buyers. We ask ChatGPT, Claude, and Perplexity the questions your buyers ask, and we record whose pages they name and cite. Our method is published [7]. The first step is a free 12-prompt sample report. Message Elizabeta Kuzevska on LinkedIn (https://www.linkedin.com/in/elizabeta-kuzevska/) and we send it within three working days.
If your assistant answers from your own documents, book a call with John Bush, our co-founder: https://meetings.hubspot.com/john2956 . Bring 10 questions your customers ask and the one type of document they are answered from. The call starts with a Context Studio demo. The next step is a scoped pilot on that one document type, with an accuracy score on your own documents.
If this was useful, the next one arrives on Thursday. Subscribe to The Revenue Signal: https://revenueexperts.beehiiv.com/?utm_source=blog&utm_medium=article&utm_campaign=chatbot-4-checks
Sources
[1] Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, decision issued 14 Feb 2024 (paragraphs 2, 17, 27 and 28). https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do
[2] Colin Lecher, The Markup, “NYC’s AI Chatbot Tells Businesses to Break the Law”, 29 Mar 2024. https://themarkup.org/news/2024/03/29/nycs-ai-chatbot-tells-businesses-to-break-the-law
[3] Klaudia Jaźwińska and Aisvarya Chandrasekar, Tow Center for Digital Journalism, Columbia Journalism Review, “AI Search Has a Citation Problem”, 6 Mar 2025. https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php
[4] NIST, AI 600–1, “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile”, July 2024 (section 2, risk 2: Confabulation). https://doi.org/10.6028/NIST.AI.600-1
[5] Gartner, press release, “Gartner Survey Finds 87% of Customers Say Companies Using GenAI for Customer Service Must Provide Access to a Human Agent”, 4 Aug 2026 (3,566 customers, Feb to Mar 2026). https://www.gartner.com/en/newsroom/press-releases/2026-08-04-gartner-survey-finds-87-percent-of-customers-say-companies-using-genai-for-customer-service-must-provide-access-to-a-human-agent0
[6] Gartner, press release, “Gartner Survey Finds Customers Are Three Times More Likely to Use Third-Party GenAI Than Company-Provided Chatbots for Customer Service”, 8 Jul 2026. https://www.gartner.com/en/newsroom/press-releases/2026-07-08-gartner-survey-finds-customers-are-three-times-more-likely-to-use-third-party-genai-than-company-provided-chatbots-for-customer-service
[7] Revenue Experts, RE-METH-AIV-003, case study methodology, public edition 1.0, October 2026. https://revenueexperts.ai/wp-content/uploads/2026/10/RE-METH-AIV-003-Case-Study-Methodology-Public-Edition-1.0.pdf
Related
- AI search visibility (AEO) at Revenue Experts
- Academy Lab: is your site ready for AI search?
- The AI Visibility Prompt Playbook
- Book 20 minutes about an AI citation audit