English (EN)

← Back to Tech News

AIChatbotsSmall Business

How to Test Your Website Chatbot's Answers Before Customers Rely on Them

Before a chatbot answers customers on your website, test it. A simple written check small businesses can run, with lessons from New York City's AI chatbot.

A white paper speech bubble cutout on a bright pink background
A paper speech bubble. Illustration only, not a chatbot or business named in this article. Credit: Photo by Miguel Á. Padriñán via Pexels (Pexels License).

Short answer: Before a chatbot answers customers on your website, give it a written test: 20 to 30 real customer questions, each asked more than once, checked against your own prices, hours and policies. Fix or block every wrong answer, add a clear way to reach a person, and repeat the test every month and after any change.

Key points

  • AI chatbots can state wrong information with full confidence. The National Institute of Standards and Technology (NIST) lists this "confabulation" as a core risk of generative AI in its July 2024 Generative AI Profile.

  • The same question can get different answers. When The Markup tested New York City's business chatbot in 2024, the bot gave one reporter a correct answer, then gave ten staffers asking the same question the wrong one.

  • Problems can last long after launch. A Dec. 30, 2025 audit by the New York City Comptroller found the city's chatbot was still in beta and still giving inconsistent answers, more than two years after it went live.

  • Customers need a way out. The Consumer Financial Protection Bureau has documented complaints from people stuck in chatbot "doom loops" with no path to a human.

What to do this week: write down the 20 questions your customers ask most, ask your chatbot each one at least twice, and mark every answer that doesn't match what you would say.

Why test a chatbot that seems to work fine?

Because a bot that sounds right is not the same as a bot that is right. Generative AI chatbots write new sentences each time, so a reply can read smoothly and still be wrong.

NIST, the federal standards agency, defines confabulation as confidently stated but false content, the errors most people call "hallucinations." Its Generative AI Profile (NIST AI 600-1) says these errors are common in today's large language models, the technology behind most AI chat tools. For a small business, a wrong answer about a price, a refund window or opening hours lands on your counter, not the software company's.

What went wrong with New York City's business chatbot?

New York City launched an AI chatbot in September 2023 to answer business owners' questions, according to the Comptroller's audit. In March 2024, The Markup reported that the bot gave wrong answers about city rules on worker tips, accepting cash and housing vouchers. The Markup also found its answers changed from one try to the next.

A few days later, The Markup reported that the bot stayed online with a more prominent note calling it a beta product that may give inaccurate or incomplete responses, and telling users to double-check its information.

The Comptroller's Dec. 30, 2025 audit said the chatbot was still in beta testing and did not consistently give accurate answers. Of 70 users who left thumbs-up or thumbs-down feedback in July and August 2025, 50 (71.4%) were negative, out of more than 2,200 questions asked in those two months. The city's own reports for those months flagged inconsistent or inaccurate answers as issues to fix. The city's Office of Technology and Innovation disagreed with the audit's project-management findings, according to the report.

How do I test my chatbot before customers use it?

Use a short written test you can run again later. Here is a simple version for a small business:

  1. Collect the questions. Pull 20 to 30 real questions from your email, contact form, phone notes and reviews. Include prices, hours, service area, booking, refunds and "do you offer...?" questions.

  2. Write the right answers first. Next to each question, write what you or your best employee would say. That page is your answer key.

  3. Ask each question at least twice. Reword it the second time, the way a customer might. The Markup's test showed one try is not enough.

  4. Ask what it should not answer. Try questions about discounts you don't offer, services you don't sell, legal or medical questions, and other customers' orders. A good bot says it doesn't know and points to a person.

  5. Mark every miss. Wrong, made-up, outdated or inconsistent answers all count. Fix the bot's source material, block the topic, or turn the bot off for that question.

  6. Save the results. Keep the date, the question, the answer and what you changed, so you can see if a fix holds.

What should a chatbot do when it doesn't know?

It should say so and hand off to a person. Being stuck with a bot is a real complaint: the CFPB's June 6, 2023 report described customers caught in loops of unhelpful replies with no way to reach a human representative.

Lawyers at the firm Debevoise & Plimpton, writing on the firm's data blog in April 2024, suggested that companies tell customers when they are talking to a chatbot, consider giving them an option to reach a person, and test the bot both before launch and on an ongoing basis. For a small business that can be as simple as a line that says "This is an automated assistant," plus your phone number or a contact form link in every chat.

How often should I retest?

At least monthly, and every time something changes. Retest after you update prices, hours or policies, after you add new pages the bot reads, and after the chatbot vendor updates its AI model. New York City's experience shows that a bot can keep drifting long after launch.

Our view

Analysis: For a busy owner in Southwest Michigan, a chatbot is worth having only if it saves time without creating new cleanup. A one-hour test with a written answer key is the cheapest insurance you can buy. If the bot can't pass it, keep it off customer-facing pages until it can.

Frequently asked questions

Does a disclaimer like "answers may be inaccurate" solve the problem? It helps set expectations, but it does not make wrong answers right. New York City added a stronger beta warning in 2024, and the 2025 audit still found inconsistent answers. Testing and fixing come first.

How many test questions are enough? Start with the 20 to 30 questions customers ask most, because those cover most real conversations. Add any question that ever produced a complaint or a wrong answer.

Can I let the chatbot quote prices? Only if it reads from a price list you keep current, and you have tested every price it gives. If your prices change often, have the bot link to your price page or hand pricing questions to a person.

Should I tell customers they are talking to a bot? Yes. Saying so plainly sets expectations, and the Debevoise lawyers suggest it. Some states also have their own AI disclosure rules, so ask your attorney what applies to you.

For more, see our guides to writing a one-page brief so AI tools know your business, claiming your business profile in ChatGPT and choosing AI chat privacy settings for your business.


Disclosure: No company or organization named in this article paid for it, reviewed it or saw it before publication. Cybercave Technologies builds, hosts and manages websites and sets up AI/SI business automations for small businesses. This article is general information, not legal advice. Figures are as published by the sources linked here and were checked on Nov. 14, 2026. Not sponsored.

Sources

  1. PRIMARY: Office of the New York City Comptroller, "Audit Report on the New York City Office of Technology and Innovation's MyCity System," Dec. 30, 2025 (IT25PAR2002)
  2. PRIMARY: National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile" (NIST AI 600-1), July 2024
  3. PRIMARY: Consumer Financial Protection Bureau, "Chatbots in consumer finance," June 6, 2023
  4. The Markup, "NYC's AI Chatbot Tells Businesses to Break the Law," March 29, 2024
  5. The Markup, "Malfunctioning NYC AI Chatbot Still Active Despite Widespread Evidence It's Encouraging Illegal Behavior," April 2, 2024
  6. Debevoise & Plimpton Data Blog, "Mitigating AI Risks for Customer Service Chatbots," April 16, 2024
Email