What AI gets wrong about B2B tech brands.
We checked almost 50,000 answers from AI assistants about 116 B2B technology companies against verified facts. Two in three contained at least one mistake. We also traced where each mistake most likely came from, and that tells you a good deal about how accurate ChatGPT and its rivals are likely to be about your own company.
- Answers checked
- Almost 50,000
- Companies
- 116 in B2B technology
- Period
- April to September 2026
- Data
- Agentcy, our AI visibility platform
Most answers get something wrong, usually something small.
We asked ChatGPT, Gemini, Perplexity and other assistants the questions B2B buyers ask when they research suppliers, then compared every answer with a fact pack we had verified for each company. Their tone was mostly warm. On the details buyers use to compare suppliers, they were often out of date.
- 65%
- of answers had at least one error about the company
- 15%
- had an error our judge rated serious, roughly one answer in seven
- 4.5
- verified facts left out of the average answer, more than twice the number of errors
Based on Agentcy accuracy checks of almost 50,000 answers about 116 B2B technology companies between April and September 2026.
The most common mistake is confident detail nobody published.
More than half of the 80,000-plus errors we logged were specific details the company had never put in public, such as a figure, a feature name or a certification. Buyers ask precise questions. When no public page answers them, the assistant fills the gap with something plausible, built from what it half knows about the company.
The percentages add up to more than 100 because a single error can touch several topics.
One answer we reviewed credited a compliance platform with halving a customer’s audit time. Another quoted a guaranteed service level for a hardware company. Neither company had published anything of the kind, and that is typical of the errors in this group. Our examples are composites, and we don’t name any of the companies in the research.
Positioning errors run both ways. A broad automation platform was described as mainly an RPA tool, while a data provider was presented as the market leader in Europe with nothing in its fact pack to support it. Errors about funding, company size, pricing or the people who run a company were less common, although they were more likely than other errors to be rated serious.
AI leaves out more than it gets wrong.
The average answer left out four or five verified facts. Those gaps are harder to spot than errors, because a reader can’t notice what isn’t there. About a third of the missing facts concerned the range of products and capabilities a company offers. Another fifth covered security and compliance, mainly data residency and certification, which are often the first things buyers in regulated sectors check.
Most errors don’t come from any page.
Our judge only blamed a page when that page plainly supported the wrong claim, so the attribution is deliberately conservative. On that basis, about two thirds of errors had nothing behind them at all. The model seems to have known the company only vaguely and filled in the rest from memory.
About 59,000 errors with a recorded source, June to September 2026.
When a page is to blame, it’s more often your own, but third-party errors do more harm.
The company’s own pages were behind twice as many errors as third-party pages. Usually the assistant had stretched something the company said into a claim it hadn’t made, and only 7% of those errors were rated serious. Third-party pages caused fewer errors, although 15% of theirs were rated serious.
Some assistants, and some kinds of question, are worse than others.
Gemini got something wrong in more than eight answers out of ten, while ChatGPT made at least one error in about half of its answers. Across all the assistants, questions about a company’s press coverage, its executives and its pricing produced the lowest factual scores.
| Assistant | Factual score | Answers with an error | With a serious error |
|---|---|---|---|
| ChatGPT | 79 | 49% | 5% |
| Perplexity | 71 | 63% | 16% |
| Gemini | 62 | 83% | 27% |
Factual score out of 100, from more than 15,000 checked answers per assistant.
It’s getting harder to pass.
ChatGPT’s factual score slipped from 86 in April to 72 in September, and Gemini dipped over the summer, while Perplexity ended the period higher than it started. Some of that decline is down to us. Our fact packs became more detailed over those months, which made the test stricter. About one error in ten came back in later answers, and we’d chase those repeat errors first.
How to fix what AI gets wrong about your brand.
Most errors fill a gap in what the company has published, so the fix is to close those gaps in public. We’d start with your own site, because you control it and it’s the quickest source to change.
Publish the details buyers ask about
Put your pricing approach, certifications, integrations and data residency in plain text on a dated page, so an assistant has something precise to quote when a buyer asks.
Read your own pages like an assistant
Look for loose claims an assistant could stretch, old product names and figures that appear without context. Tighten them before they turn up in answers.
Correct the third-party pages with most reach
Update your profiles on review sites and directories, and ask publishers for a correction where an article carries an error that keeps appearing.
Check again every month
Ask your buyers’ questions each month and keep a record of the errors that come back. Answers change slowly, so judge progress over a quarter.
Read it in full.
The twelve-page PDF tells the same story in more depth, with the full charts and a note on how we ran the analysis. It’s free to download, and there’s no form to fill in.
Download the PDF- Pages
- 12
- Format
- PDF, 0.8 MB
- Data
- April to September 2026
Try it with AI by asking an assistant to check what it says about your company. The report shows which errors are worth looking for.
More from our AI citation research.
What AI reads before it recommends
The content types and page features that AI assistants cite most, from over 1.3 million citations.
Earned media in the age of AI answers
Which media titles the assistants read for B2B tech, and why trade and specialist press come out on top.
The UK B2B Tech PR AI Visibility Index
Which UK B2B tech PR agencies the AI assistants recommend, measured every week on the same set of questions.
AI accuracy statistics for B2B brands.
We’ve rounded the report’s headline numbers so they’re easier to quote. All of them come from almost 50,000 AI answers about 116 B2B technology companies, checked against verified facts between April and September 2026.
- 65%of AI answers about a company contained at least one error.
- 1 in 7answers had an error rated serious, and 4% of all answers were graded critical.
- 4.5verified facts were left out of the average answer, compared with 1.7 errors, so omissions outnumber mistakes.
- 57%of errors were invented specifics, such as figures, features or claims the company never published.
- 65%of errors had no source behind them, which means the model was working from memory.
- 18% vs 9%of errors traced to the company's own pages, against third-party pages.
- 15% vs 7%of third-party errors were serious, against errors from the company's own pages.
- 79 / 71 / 62factual scores out of 100 for ChatGPT, Perplexity and Gemini.
If you’re short on time, our blog post How accurate is ChatGPT? covers what we found about ChatGPT in particular, and takes about five minutes to read.
If you’re citing this research, you’re welcome to use these figures, and we’d ask you to credit and link to the source: Resonance (2026), What AI gets wrong about B2B tech brands. https://www.resonancecrowd.com/research/ai-accuracy-b2b-tech-brands
Questions we get asked about AI accuracy.
How accurate are AI assistants about B2B technology companies?
In our research the assistants were mostly right on tone and loose on detail. When we checked almost 50,000 answers about 116 B2B tech companies, 65% contained at least one error and about one in seven contained a serious one. The answers also left out four or five verified facts each, and they were often out of date.
What kind of mistakes does AI make about companies?
The most common mistake is invented detail, which appeared in more than half of the errors we found. In those cases the assistant added a figure, feature, integration or certification that the company had never published. Errors about funding, pricing and people were rarer, although they were the most likely to be rated serious.
Do AI errors come from a company's own website or from other sites?
In most cases the error doesn’t come from a website at all. About two thirds of errors had nothing cited behind them, so the model was working from memory. When a page was to blame, it was the company’s own site about twice as often as a third-party one, although third-party errors were twice as likely to be serious (15% against 7%).
Which AI assistant is most accurate about B2B brands?
Of the three assistants we could compare fairly, ChatGPT was the most accurate. It got something wrong in about half its answers and scored 79 out of 100 for factual accuracy, compared with 71 for Perplexity and 62 for Gemini, which made an error in more than eight of every ten answers.
What does AI leave out about companies?
Assistants most often leave out the breadth of what a company sells. About a third of missed facts concerned products and capabilities the answer didn’t mention, and another fifth concerned security and compliance, mainly data residency and certification.
How can a company fix what AI says about it?
Start with your own site, because it’s the quickest source to change. Publish the details buyers ask about, such as your pricing approach and your certifications, in plain dated text, and tidy up loose claims an assistant could stretch. After that, correct the review sites and directories that carry the worst errors, and check the answers again every month.
Why does ChatGPT get facts wrong about my company?
Mostly because it’s filling gaps in what it knows about you. In our research, about two thirds of errors had no source behind them, which suggests the assistant knew the company vaguely and supplied plausible detail such as a feature, a figure or a certification. When buyers ask precise questions that your public pages don’t answer, it will often write an answer anyway.
How do I monitor what AI says about my brand?
Put your buyers’ questions to the assistants on a regular schedule. Record each answer and the sources it cites, and note any errors. Check at least the three largest assistants, because their answers about the same company can differ considerably. AI brand monitoring tools, including Agentcy, the platform behind this research, can automate the questions and check each answer against verified facts about your company.
Can you correct what ChatGPT says about your company?
You can’t edit ChatGPT’s answers directly, although you can change what it reads. Publish the details buyers ask about on your own site in plain, dated text, and fix any loose claims an assistant could stretch. It also helps to correct the review sites and directories it relies on, along with any articles that repeat an error. Answers that come from model memory change more slowly, which is why a consistent story across the web and a monthly check both help.
How was the research done?
Agentcy, our AI visibility platform, collected the answers between April and September 2026. An AI judge compared each one with a verified fact pack for the company and logged every wrong claim, how serious it was and, from June, where it most likely came from. We don’t name any company in the report.
Find out what AI gets wrong about you.
Tell us who you are and we'll ask the leading AI assistants the questions your buyers ask about you. We'll send you a short snapshot of the errors and gaps in their answers, along with where each mistake most likely came from.
The snapshot is free, with no obligation. We run it on Agentcy, the platform behind this research.
By submitting this form, you agree to our privacy policy.