GuideGuides and tutorials
How Accurate Are AI Chatbots? How to Measure It
There is no single accuracy number for AI chatbots: it depends on what the bot reads, how closely it sticks to that, and what it does when the answer is not there. Here is the evidence, and a way to score your own.
The short answer
It depends on more than the AI model: what the chatbot reads, how strictly it sticks to that content, and what it does when the answer is not there. General chatbots guess. Chatbots that answer from your own content do better, yet a 2024 Stanford study found tools built that way still hallucinating 17% to 33% of the time on hard legal questions. Skip the vendor’s percentage: ask your own bot 50 real customer questions and score every answer.
Jake Moffatt’s grandmother died on 11 November 2022, and that day Moffatt went to Air Canada’s website to book a flight from Vancouver to Toronto. The website’s chatbot explained the airline’s bereavement fares and added that a traveller who had already flown could “submit your ticket for a reduced bereavement rate” within 90 days of the ticket being issued. The bereavement page that the chatbot itself linked to said the opposite: the policy did not apply once travel was complete.
Moffatt flew at full fare, applied for the refund and was turned down. At British Columbia’s Civil Resolution Tribunal, Air Canada argued it could not be held liable for what its chatbot said. The tribunal member called that “a remarkable submission” and wrote: “It makes no difference whether the information comes from a static page or a chatbot.” The airline was ordered to pay C$812.02 in damages, interest and fees (Moffatt v. Air Canada, 2024 BCCRT 149, issued 14 February 2024, checked 14 October 2026).
That is the real question behind “how accurate are AI chatbots”: will the bot on your site tell a customer something your business does not do, and what happens when it does not know? This guide covers what accuracy means for a support chatbot, why chatbots make things up, what independent tests found, and how to measure your own.
What “accurate” means for a support chatbot
For a general chatbot, accurate means true about the world. For the chatbot on your website it means something narrower and more useful: the answer matches what your business actually does. That breaks down into four tests.
- Correct. The facts are right: the price, the opening hours, the returns window, the areas you deliver to.
- Grounded in your content. The answer comes from your pages, files and written answers, not from whatever the model absorbed elsewhere on the internet. A sensible answer about returns in general is still wrong if it is not your returns policy.
- Complete. It includes the conditions that change the answer. “Yes, you can return it” is incomplete when the truth is “within 30 days, unused, with the receipt”.
- Honest about gaps. When your content does not cover the question, it says so and offers a person instead of filling the gap with a guess.
The fourth is the one people forget. A chatbot that says “I don’t have that information, shall I get someone?” has done what a sensible new employee would do. The Air Canada bot went wrong because it confidently stated a rule the airline did not have.
Why AI chatbots make things up
A language model writes the most likely next words, given everything before them. Most of the time the most likely words are also true. When the model does not know, the most likely words are still fluent and confident, and the result is a hallucination: an answer that sounds right and is not.
OpenAI’s researchers argued in September 2025 that the habit is partly taught. Models are trained and scored on tests that reward a right answer and give nothing for “I don’t know”, so guessing pays. Their example: asked for a stranger’s birthday, a model that guesses has a 1 in 365 chance of scoring, while admitting it does not know scores nothing. Over thousands of questions, the guesser looks better on the scoreboard (OpenAI, “Why language models hallucinate”, 5 September 2025, checked 14 October 2026).
The same post shows what saying “I don’t know” buys. On a test of short factual questions, an older model answered nearly everything and was wrong three times in four. A newer one declined about half the questions, got almost as many right, and was wrong far less often.
| Model | Declined to answer | Right | Wrong |
|---|---|---|---|
| OpenAI o4-mini | 1% | 24% | 75% |
| gpt-5-thinking-mini | 52% | 22% | 26% |
The two models were right about equally often. What separated them was what they did when they did not know. The post also says plainly that “accuracy will never reach 100%”, because some real-world questions are inherently unanswerable, and that hallucinations are not inevitable “because language models can abstain when uncertain”.
Does answering from your own content fix it?
Chatbots built to answer from a business’s own content use retrieval, often called RAG (retrieval-augmented generation). Before it answers, the chatbot searches your pages and documents for the passages that match the question, and the model writes its reply from those passages. That is a large improvement over a general chatbot, because the model is no longer answering from memory. It also leaves some errors behind, and two independent sources show why.
Even with the document in front of it, a model can add things
Vectara, a company that builds retrieval software, runs a public hallucination leaderboard. Each model is given more than 7,700 articles and asked to summarise each one “using only the facts presented in the document”; a separate checking model then compares every summary with its source. In the version last updated on 22 September 2026, the share of summaries containing something the source did not support ran from 1.8% for the best of more than 100 models to 24.2% for the worst (Vectara hallucination leaderboard, checked 14 October 2026).
It measures English summaries, not customer answers, and its authors say it “does not definitively measure all the ways a model can hallucinate”. Two lessons still carry over: the model you choose makes a real difference, and no model on the list scored zero.
Tools described as “hallucination-free” were not
In 2024, researchers at Stanford’s RegLab and Institute for Human-Centered AI tested AI legal research tools from LexisNexis and Thomson Reuters, all built on retrieval. Legal research providers had described retrieval as “eliminating” or “avoiding” hallucinations, and one had promised “hallucination-free” citations. On more than 200 legal questions, the tools made fewer errors than GPT-4 on its own, but each still hallucinated between 17% and 33% of the time (Magesh and others, “Hallucination-Free?”, May 2024, and the Stanford HAI summary of 23 May 2024, both checked 14 October 2026).
The study also flagged answers that are right but “misgrounded”, citing a source that does not support them. Legal research is far harder than “what time do you open?”, so 17% is not a forecast for your shop. The pattern is what matters: retrieval lowers the error rate and leaves some behind, so you still need to test, and you still need a way through to a person.
How to measure your chatbot’s accuracy
No published benchmark can tell you how a bot will do on your questions with your content. A 50-question test can, and it takes an afternoon. It also works before you buy: run it on a free plan or trial with your own website loaded.
Collect 50 real questions. Take them from your inbox, chat history and phone notes: the ten you get most, some awkward wordings, questions with conditions, a few that need two facts at once, and five to ten your content does not answer at all.
Write the expected answer first. Before you ask the bot anything, write what a good member of staff would say from your current policy. Where your content has no answer, the expected answer is "I don't know, let me get you a person".
Ask the way a customer would. Start a new chat for each question, typos included, and copy each reply into a spreadsheet next to the expected answer.
Score every answer. Use the five scores in the table below. Ask a colleague to score ten of them too; where you disagree, sharpen the expected answer.
Count what matters. Correct answers plus correctly declined ones are your safe share. Then look hard at every wrong answer, because a confident wrong answer is the Air Canada problem.
Fix, then ask the same 50 again. Repeat the test after every change to your content, prices or AI model, and monthly regardless. Add new questions as they arrive, but keep the original 50 so the scores stay comparable.
The examples below come from a bike shop we made up for this article. The question is “Can I return a helmet I’ve worn once?”, and the shop’s policy is 30 days, unused items only.
| Score | What it means | Example | What to do |
|---|---|---|---|
| Correct | Matches the expected answer, conditions included, nothing added. | "Unused helmets can be returned within 30 days. A worn one can't, but we can check it for damage." | Nothing. Keep it in the set so you notice if it breaks. |
| Partly correct | Right direction, but a condition, step or number is missing or vague. | "Yes, returns are accepted within 30 days." It leaves out the condition that decides this case. | Put the condition next to the rule in your content, or write an exact answer. |
| Wrong | Contradicts your content, or invents a price, date, rule or service. | "Worn helmets can be returned within 90 days for store credit." No such rule exists. | Find the cause the same day: missing, outdated or conflicting content. Fix it and ask again. |
| Correctly declined | The answer is not in your content, and the bot says so and offers a person. | Asked to rebuild a 1970s gear hub, it says it does not have that information and offers the team. | Decide whether it should answer next time. If so, write the answer. |
| Declined, but the answer was there | Your content answers it, yet the bot said it did not know. | Asked "do you fix e-bikes?", it offers a person, although the services page lists e-bike repairs. | Reword the content the way customers ask, or add an exact question and answer. |
Worked example: say the bike shop’s 50 questions come back as 36 correct, 5 partly correct, 3 wrong, 4 correctly declined and 2 declined although the answer was there. The safe share is 40 of 50. The three wrong answers matter most, and here all three quoted last year’s service prices from an old PDF price list that was still loaded. Removing one file fixes them, and no vendor’s accuracy figure would have shown it. (The shop and its numbers are invented; they are not a benchmark.)
For the rest of a pre-launch check, from tone to edge cases, see our AI chatbot testing checklist.
How to raise your chatbot’s accuracy
- Cover every common question in your content. Most wrong and declined answers trace back to a gap. Your 50 questions double as a checklist: each one needs a clear answer somewhere the bot reads. Our guide to training a chatbot on your website content covers which pages to include.
- Write exact answers where the wording matters. Refund rules, warranty terms, prices for set services, anything a customer might quote back at you. Store these as exact questions and answers, so the bot uses your wording instead of its own.
- Keep one current version of every fact. Old price lists, last year’s PDF, a seasonal page nobody took down: retrieval will find them. Update or remove them, and have the bot re-read your site after any change to prices or policies. The Air Canada bot contradicted a page on its own website.
- Read what it could not answer, every week. The questions your chatbot could not answer are your customers telling you what your website is missing.
- Set up the hand-over, then test it. Make sure “I don’t know” is followed by an offer of a person, and that someone on your team sees the request, including out of hours. Our guide to chatbot to human hand-overs shows what a good one looks like.
- Test again after changing the model. Vectara’s leaderboard shows how far models differ on the same task. A newer or cheaper model may do better or worse on your questions, and your 50 questions will show which.
For the technical side, from retrieval to prompts that keep answers tied to the source, see our guide to preventing AI hallucinations in customer support.
What to keep a person for
Some conversations belong with a person even when a bot could produce a correct-sounding answer:
- Exceptions to policy. A refund outside the window, a goodwill gesture, a special price. The Air Canada case began with a refund question.
- Complaints and upset customers, who want to be heard by someone who can act.
- Questions about one customer’s order or account, unless the bot is connected to your own system and can only look up what you allow.
- Advice with real consequences: legal, medical, financial or safety questions.
- Anyone who asks for a person. Make it easy, and never make them argue with the bot.
A workable rule: if a wrong answer would cost more than a five-minute conversation, send it to a person.
Where EzyConn fits
The EzyConn AI agent is built around the behaviour this guide asks for. It answers from the content you give it: your website (enter the address, and keep the pages you want), files in PDF, Word, plain text, Markdown or CSV, text you write, and exact questions and answers, which are used first, word for word. When you change prices or policies on your site, “read it all again” re-reads it in one click.
When your content does not cover a question, it says so and offers a person. The default reply, which you can change, is: “I don’t have that specific information right now. Would you like me to connect you with a support agent who can help?” If the customer says yes, a live agent from your team is brought in with the whole conversation, in the EzyConn inbox or in Slack. When a teammate answers a question it could not, that reply waits as a draft answer for you to approve, and the Unanswered screen ranks the topics it answers least: the weekly habit above, built in.
You also choose the AI model. GPT is included, or you can add your own Claude or Gemini key and switch in a click, then run your 50 questions again to compare. The EzyConn AI agent page explains how it finds and writes an answer, and how EzyConn works shows the hand-over to your team step by step. The Free plan includes the AI agent and needs no card, so you can start free, load your own content and put your 50 questions to it before you pay for anything.
AI chatbot accuracy: common questions
How accurate are AI chatbots?
There is no single figure, because accuracy depends on the questions and on what the chatbot reads. General chatbots answering from memory guess more often. Chatbots that answer from a business's own content do better but still make mistakes: a 2024 Stanford study found that retrieval-based legal research tools hallucinated between 17% and 33% of the time. The number worth knowing is your own, from 50 real customer questions.
What is a chatbot hallucination?
An answer that sounds confident but is not supported by the facts: an invented refund rule, a price that does not exist, a service you do not offer. It happens because a language model writes the most likely next words, and when it does not know, the most likely words are still fluent. Answering from your own content makes it rarer, and an honest "I don't know" keeps a gap from becoming a wrong answer.
What is the hallucination rate of AI chatbots?
It varies a great deal by model and by task. On Vectara's public leaderboard, last updated on 22 September 2026, models asked to summarise a document using only its facts added unsupported content to between 1.8% and 24.2% of their summaries (checked 14 October 2026). That is a summarising test, not customer support, so read it as a comparison between models rather than a forecast for your chatbot.
Can you stop a chatbot from hallucinating completely?
Not with today's technology. OpenAI's researchers wrote in September 2025 that accuracy will never reach 100%, because some real-world questions are inherently unanswerable. What you control is everything around the model: answers drawn from your own content, exact answers where wording matters, current prices and policies, an honest "I don't know" when the content runs out, and a quick hand-over to a person.
How do I measure my chatbot's accuracy?
Collect 50 real customer questions, including some your content does not answer. Write the expected answer for each before you ask the bot. Ask each one in a fresh chat, then score the reply as correct, partly correct, wrong, correctly declined, or declined although the answer was there. Fix the wrong answers first, and ask the same 50 again after every change to your content or your AI model.
Are AI chatbots reliable for customer service?
For common, factual questions answered from your own current content, they can be, and they answer at any hour. Exceptions to policy, complaints, questions about one customer's account and anything that needs judgement belong with a person. A reliable setup uses both: the bot answers what your content covers, says so when it does not know, and passes the rest to your team with the conversation attached.
Is a business responsible for what its chatbot says?
In Moffatt v. Air Canada (February 2024), British Columbia's Civil Resolution Tribunal held the airline responsible for wrong fare advice from its website chatbot, noting that it "makes no difference whether the information comes from a static page or a chatbot." It is a small claims decision in one Canadian province and this is not legal advice, but it is a sensible assumption to work from.
Which AI model is the most accurate?
It depends on the task, and the rankings change every few months. On Vectara's summarising test, the best and worst hallucination rates were more than ten times apart. For a support chatbot the model is one factor; the content it reads and what it does when unsure are the others. If you can switch models, run the same 50 questions on each and compare.