Lancashire Digital • Kevin Lancashire, October 2026

AI & Brand Strategy

What AI assistants say about your brand is often wrong – and it misleads your customers

The bottom line first: when a customer asks an AI assistant about your product today, they get an answer that sounds competent – but no one at your company ever approved it, and often it isn’t true. This isn’t a fringe image problem. It’s misinformation delivered to your customers, phrased in your name.

The Core: You don’t eliminate wrong answers by hoping; you do it by controlling the sources. Only when you know what the models draw on can you steer what they say.

The Problem

Unapproved Billboards

Picture a billboard that gets your message wrong – wrong price, wrong feature, a promise you never made. You’d take it down immediately. That is exactly what happens inside AI assistants, constantly, with two differences: you never approved this billboard, and it rewrites itself with every query. The person accountable carries the consequence without ever having seen the statement.

Budgets

Rising AI Investments

Enterprises are increasing their AI investments sharply, while average IT budgets grow only modestly – by around 3.7%¹. So the money flows precisely to where the output is least controlled.

Impact

Cost of Errors

The more customers rely on AI answers, the more every wrong one costs. The pressure makes this more urgent, not less.

The Evidence

The Hallucination Reality

There is no single “error rate” – hallucination rates vary widely by model, task and domain. But where you look closely, they are substantial.

01

Scientific References

Models fabricated between 28.6% and 91.4% of citations depending on the system² – references that look real and aren’t.

02

Legal Facts

On specific, verifiable legal questions, models hallucinated facts between 58% and 88% of the time³.

03

Prompt engineering isn’t enough

Across 300 clinical case vignettes, each seeded with one fabricated medical detail, six models reproduced that fabrication in 66% of cases on average. A prompt written specifically to induce caution brought this down to 44% – it did not remove the problem⁴.

The Brand Impact

The same failure pattern hits your brand: invented features, wrong terms, outdated details – delivered with confidence. Automated scores alone don’t catch this. A visibility dashboard tells you that you appear – not what is being said wrong about you. That requires a manual audit: someone who reads what the models actually answer and recognises where the statement tips over. At a global pharmaceutical group I was accountable for three internal audits in digital governance and closed the findings they raised; the result was a group-wide directive for digital compliance. I know what to look for.

Solution Approach

Take Control of the Sources

Establish a baseline – manually verified. Get your sources under control: clean, discoverable, consistent references. Monitor continuously across the LLMs you choose, independent of any single provider.

Let's Connect

The difference from a dashboard subscription: you don’t get a score, you get a verifiable list of what needs correcting – and a way to do it. I’m currently looking to connect with people who are working on this. If this is on your radar, I’d be glad to grab a coffee and compare notes. DM me.

Methodology More on the approach: the AI Answer Audit. Technology The tool behind it (Beta): Seravel.

Drafted with AI assistance; every figure and source was manually verified against the primary paper – which is rather the point.

Sources

• [1] Gartner, 2027 CIO Agenda – AI funding rising sharply against average IT budget growth of ~3.7%. gartner.com/en/chief-information-officer/insights/cio-agenda
• [2] Chelli, M. et al. (2024): Hallucination Rates and Reference Accuracy of ChatGPT and Bard for Systematic Reviews. Journal of Medical Internet Research. (Fabricated references: 28.6% GPT-4, 39.6% GPT-3.5, 91.4% Bard)
• [3] Dahl, M. et al. (2024): Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models. arXiv. (Legal hallucinations: 58% ChatGPT-4 to 88% Llama 2 on verifiable case questions)
• [4] Omar, M. et al. (2025): Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Communications Medicine 5, 330. doi.org/10.1038/s43856-025-01021-3 (66% → 44% is the mean across six models with deliberately seeded false details; the body text reports 65.9% and 44.2%.)