When enterprise brands voice concerns about AI inaccuracies, the assumption tends to be that AI is simply hallucinating responses to prompts it does not know the answers to. However, recent data from Bluefish Research suggests AI hallucinations are only part of the story. In nearly a third of cases, AI didn't invent the wrong answer, it actually read the inaccuracy somewhere in its training or RAG data. 

Bluefish used its new AI Accuracy technology to track millions of brand claims — verified product details, finance rates, medical dosages— to measure how accurately they appeared in AI responses. Across inaccurate claims found with an associated verified source, our research showed that errors fell into one of three buckets: 

1. 30% of inaccurate responses can be traced directly to a cited source carrying an incorrect claim. Essentially, AI found wrong information somewhere on the internet and repeated it. This bucket is the most straightforward for brands to fix — identify where the error is being seeded and address it at the source. 

   
   

2. Another 32% of inaccuracies tied to citations that didn't directly address the inaccurate claim. The model simply filled in the gaps where its cited sources didn’t address topics directly. For brands, this bucket can be mitigated with authoritative content strategies. 

   
   

3. Egregiously, the last 38% of inaccuracies were claims made by AI that directly contradicted one or more of its own cited sources. Meaning the right information exists, but AI ignored it. 

Brands that can see which bucket their errors fall into can prioritize action accordingly. For inaccurate AI claims that have a traceable origin, they are diagnosable and fixable. For more ambiguous claims, they actually light up a clear path to content gaps to be filled. And when AI contradicts a source that already had the right answer, that's an authority problem — the brand's messaging exists, but AI didn’t trust it enough to base an answer on it. With transparency, all three buckets of inaccuracies have a path forward, whether it’s earning presence, building authority, or driving repetition with PR and other partners. 

A lot of AI Inaccuracies Involve Numbers 

Bluefish Research also found that across major commercial brands and AI channels, many AI errors included incorrect numbers — prices, rates, counts, dosages — accounting for 64% of inaccuracies. In particular, Finance and Pharma have the highest-stakes outliers: In Finance, errors spread across rates, prices/fees, loan terms, and credit-score minimums. In Pharma, 21% of its inaccuracies reflected incorrect efficacy percentages and 27% reflected incorrect dosage recommendations. 

The Internet (3rd Parties) is Getting Your Brand Wrong 

Unsurprisingly, when looking at traceable inaccuracies, 60% of citations propagating incorrect information tracked back to blogs/topic-driven independent media sites, while brand-owned sites only account for about 20% of citation-backed errors. Examples include personal finance bloggers citing a credit card APR, travel points sites quoting a baggage fee, or wellness blogs sharing dosage information. These third-party sites carry significant weight in AI responses, often much more than the brand's own page on the same topic. 

Implications for Brands 

The findings below reframe what AI accuracy means for enterprise brands. Brands are increasingly investing in AI visibility, topic authority, and source optimization to improve how they appear in AI responses. That investment is well-placed, but it builds on a flawed foundation if the way you show up to consumers is fabricated or misleading.

Optimizing for AI without fixing accuracy is like running a paid campaign to a broken landing page — you're driving traffic to the wrong answer. And unfortunately, high consumer confidence in AI coupled with condensed consumer journeys mean brands won’t have another opportunity to remedy their brand image before a purchase decision is made. 

As brands build more robust AI Marketing practices, it’s important to note that Accuracy is not a one-time audit. AI responses change as models update, sources shift, and brand information evolves. A price that was correct last quarter may be wrong today. A dosage that appeared accurately in one model may be misrepresented in another. Brands that treat accuracy as a continuous signal will build the most durable AI performance over time. 

AI Accuracy is Specific, Diagnosable, and Necessarily Addressable 

Inaccuracy in AI is a specific, diagnosable, and in many cases addressable problem. Brands that understand where errors originate — which sources feed them, whether the model is echoing bad data or ignoring good data — will drive maximum actionability and ROI in their AI Marketing practice. Accuracy evaluation and error transparency help brands clearly prioritize organic and paid media investments, transforming Answer Engine Optimization into a growth driver for their business. 

Bluefish's AI Accuracy Platform provides brands with claim-level accuracy data, citation root-cause analysis, and channel-by-channel breakdowns across Amazon/Alexa, Google AI Mode, ChatGPT, Claude, Perplexity, and more. Reach out here to learn more.