When you ask ChatGPT about a specific company, it gets at least one thing wrong in about half of its answers. We checked almost 50,000 answers from AI assistants about 116 B2B technology companies against verified facts. ChatGPT made an error in 49% of its answers and a serious error in about one in 20, which made it the most accurate of the three assistants we could compare properly. Perplexity got something wrong in 63% of its answers, and Gemini in 83%.
This research answers a narrow question. We didn’t test ChatGPT on maths, coding or general knowledge. We asked the questions B2B buyers ask about the companies they are considering, then compared each answer with a pack of verified facts about that company, which shows how well ChatGPT describes a business to the people who might buy from it.
We scored every answer out of 100 for factual accuracy and logged each wrong claim with how serious it was. Each of the three assistants had more than 15,000 answers checked.
| Assistant | Factual score | Answers with an error | Answers with a serious error |
|---|---|---|---|
| ChatGPT | 79 | 49% | 5% |
| Perplexity | 71 | 63% | 16% |
| Gemini | 62 | 83% | 27% |
Based on Resonance’s analysis of almost 50,000 answers about 116 B2B technology companies between April and September 2026. We also checked Claude, Copilot, Google’s AI Mode and Google’s AI Overviews, although the samples were too small for a fair comparison.
ChatGPT comes out best. Being wrong in half of its answers is still a lot when the answer is about your company, although most of its mistakes were small and only about 1% of its answers were graded critical.
Across everything we checked, the assistants were kinder than they were accurate, averaging 82 for tone against 71 for factual accuracy. Your buyer gets a friendly description of your company that is frequently out of date on the detail.
We sorted more than 80,000 errors from all the assistants by what they were about. One pattern accounted for more than half. In 57% of errors the assistant had added specific details the company had never published, mostly because a buyer had asked a precise question that no public page answered.
Two composite examples show what this looks like without naming anyone. One answer credited a compliance platform with halving a customer’s audit time. Another quoted a guaranteed service level on behalf of a hardware company, and neither company had published anything of the kind.
Product features and capability came up in about a third of errors, as integrations, deployment options or features that were described wrongly or made up. Compliance and certification appeared in about one error in ten, while pricing and licensing accounted for about 7% of errors, as did numbers and statistics.
Positioning errors went both ways. We saw a broad automation platform described as mainly an RPA tool, and a data provider presented as the market leader in Europe without any support in its fact pack. Errors about funding, company size, pricing or the people who run a company were less common, although they were more likely than others to be rated serious, and questions about a company’s press coverage and its executives produced the lowest factual scores of any question type.
Omissions are harder to spot than errors because nothing on the screen looks wrong. The average answer missed four or five verified facts about the company, compared with fewer than two errors. About a third of those gaps concerned the range of products and capabilities a company offers. Another fifth were about security and compliance, mainly data residency and certification, which buyers in regulated sectors tend to look for first.
One test is worth doing yourself. Ask ChatGPT what your company offers and compare the answer with your full product list, since a missing product rarely shows up any other way.
From June we also recorded where each error most likely came from. Our judge blamed a page only when that page plainly supported the wrong claim, so these figures are deliberately cautious.
About two thirds of errors across all the assistants had nothing behind them, which suggests the model knew the company only vaguely and filled in the rest from memory. For ChatGPT the figure was 61%. The company’s own pages were behind 18% of errors overall and 27% of ChatGPT’s, mostly where the assistant stretched something the company had said. Third-party pages accounted for 9% of errors overall and 6% of ChatGPT’s, and the source of the remainder was unclear.
When a page is to blame, it is more often your own. Third-party errors do more harm, though, with 15% of them rated serious compared with 7% of the errors that came from a company’s own pages. Changing what a model remembers takes longer than fixing your own site, because it depends on the rest of the web describing you consistently over many months.
If you want to know which pages the assistants lean on in the first place, we looked at that in What AI reads before it recommends.
The full research, free
What AI gets wrong about B2B tech brands has the full charts, the themes behind the errors and a comparison of the assistants, along with a note on how we did the checking. It runs to 12 pages, and there’s no form to fill in.
ChatGPT’s factual score slipped from 86 in April to 72 in September. Over the same period Gemini dipped during the summer, while Perplexity ended higher than it started, at 77.
We would treat ChatGPT’s drop with some care. Our fact packs became more detailed over those months, which made the test harder to pass, so some of the fall comes from ChatGPT’s answers and some from us checking more closely.
About one error in ten came back in later answers. Those repeat errors are the ones we would chase first.
Start by asking ChatGPT the questions your buyers ask, including questions about you by name, about your category and about how you compare with competitors on price and compliance. Then put the same questions to the other major assistants, because they get different things wrong.
For each answer, note the errors and gaps along with the sources it cites, and compare it with a short fact sheet of the things you would want any buyer to know. Do this every month. The answers change as the models and their sources change.
Once you have more than a handful of questions, the process gets slow. Agentcy, the platform behind this research, asks the questions on a schedule and checks each answer against your facts, and if you’d like to see what it finds for your company, we’ll run a free AI accuracy snapshot.
Most errors fill a gap in what a company has published, so the work is to close those gaps in public. We would start with your own site because it is the quickest source to change.
Our B2B tech PR team does this work with clients, alongside the media relations that shape what the assistants read.
Mostly on tone, much less on detail. In our checks of B2B technology companies it got at least one thing wrong in about half of its answers and made a serious error in about one in 20, which still made it the most accurate of the three assistants we compared.
Sometimes, and invented detail was the most common error we found. The assistant added a figure, feature, integration or certification the company had never published, most often when a buyer asked a precise question that the company’s public pages didn’t answer.
ChatGPT draws on what it learned in training and, when it searches, on the pages it cites. In our data about two thirds of its errors had no supporting source, which means they came from memory. When a page was behind an error, it was more often the company’s own site than a third-party one.
You can’t edit its answers directly, although you can change what it reads. Publish the specifics on your own site in plain, dated text and fix any claims that could be stretched, then correct the review sites and directories it relies on. Answers drawn from memory change more slowly, so keep your story consistent and check again every month.
Of the three assistants we could compare fairly, ChatGPT was the most accurate, with a factual score of 79 out of 100 compared with 71 for Perplexity and 62 for Gemini.
Ask the main assistants the five questions your buyers ask most, and write down every mistake and every gap you find. Fix the pages you control. A month later, check what has changed.
If you would rather we did it, ask for a free snapshot and we’ll send you what the assistants say about you, along with where their mistakes come from.
This post draws on What AI gets wrong about B2B tech brands (Resonance, 2026), which used Agentcy to check almost 50,000 answers from ChatGPT, Gemini, Perplexity and other assistants about 116 B2B technology companies against verified facts between April and September 2026. Error sources were recorded from June 2026.