When AI lies about your company. A brand hallucination field guide.

The hallucination rate baseline

AI hallucination is not a fringe edge case. It is a measured, structural feature of probabilistic systems applied to deterministic domains. NeuralTrust’s 2025 analysis frames it bluntly: hallucinations are “an inherent risk of using probabilistic models in deterministic domains.” They are a feature, not a bug.

The published rates by domain:

• Legal AI tools: Stanford research found 17–33% hallucination rates across major legal-research AI products.
• Medical AI: documented hallucination rates range 28–90% depending on the system and query type.
• General AI search: Columbia Journalism Review’s test of 8 engines: 60% inaccurate or misleading.
• User-side: MarTech data shows 42.1% of users encounter inaccurate AI content; 16.78% have encountered unsafe or harmful AI advice; over 40% rarely or never click through to verify a source.

Translate that to brand math: every day, AI engines make millions of confidently incorrect statements about real businesses, and most users never cross-check.

“There are three kinds of lies: lies, damned lies, and hallucinations. — UC Berkeley SCET”

The Air Canada precedent

In 2024, Air Canada’s customer-service chatbot invented a bereavement-fare policy that did not exist. A grieving passenger acted on the AI’s answer, booked the trip at full price expecting a refund, and was denied. He sued. The Canadian tribunal ruled the airline liable for the chatbot’s statements: if the AI is the customer-facing voice of the brand, the brand owns what the AI says.

The legal exposure is real and replicable. The National Law Review’s 2025 framing was unambiguous: “If your AI acts as an agent of your business, you likely bear responsibility for what it tells people.” The legal framework is still settling, but the early-precedent direction favors users harmed by hallucinations and not the brands deploying the AI.

The Soundslice case: built what the AI invented

Soundslice, a music-software company, discovered ChatGPT was telling users it had an ASCII tab import feature. The product had no such feature. Users were signing up to use it, finding nothing, and churning. ChatGPT had hallucinated the capability into existence.

The company eventually built the feature — not because they had planned to, but because the AI had created enough demand for it that shipping was the cheaper response than continued explaining. iPullRank documents this as the canonical example of AI-driven product roadmap pollution: your AI-generated brand reality starts to shape your real product strategy.

Hoka faced a similar issue from a different direction: ChatGPT was showing prospective customers incorrect pricing pulled from outdated third-party sources. Even after Hoka updated their official pricing, the AI continued surfacing the wrong number. The lag between brand correction and AI retraining is months, sometimes longer.

Why users don’t catch it

The user-behavior side of the hallucination problem is what makes it consequential at scale. Pew Research found only 1% of users click into the cited source when an AI Overview is shown. Bain’s consumer survey found ~80% of consumers now rely on zero-click results.

The trust paradox compounds: an Exploding Topics analysis of 2025 consumer trust data found 82% of users are skeptical of AI results, yet only 8% always check sources. Skepticism without verification creates the worst-case outcome — users who don’t fully trust AI answers but act on them anyway, then carry the misinformation into their decision-making.

For brands, this means: a hallucinated claim about your company in an AI Overview reaches users who are already skeptical of the medium, but who will act on the claim anyway, and will not check whether your website tells a different story.

The four classes of brand-level hallucination

From the case literature, hallucinations affecting brands cluster into four reproducible patterns:

1. Outdated training data. AI surfaces old pricing, old leadership, old policies, old product capabilities. The brand updates the website; the AI keeps quoting the old version for months.
2. Brand confusion. RankScience documented a US consulting firm whose AI responses blended its history with a UK firm of similar name, making the US firm invisible on competitive queries. Fictional-character namesake collisions are the worst case.
3. Fabricated specifics. Invented features, made-up partnerships, hallucinated awards, fake founder biographies. Soundslice is the textbook example.
4. Negative narrative drift. Third-party criticism, outdated controversies, or competitor-comparison content getting surfaced as the dominant brand frame even when the underlying issue is resolved.

The defensive GEO playbook

You cannot directly correct an AI model’s output. The only effective response is to flood the information ecosystem with accurate, structured, authoritative content that gives AI systems high-confidence material to draw from. Edelman’s 2025 framing: “Earned media is the single most important driver of brand visibility in AI-generated responses.”

The MarTech-validated defensive sequence:

1. Audit systematically. Query all major AI platforms with brand-specific prompts (“What is [Brand]?” / “Who founded [Brand]?” / “What does [Brand] cost?”). Document every false claim. Track over time.
2. Diagnose the root cause. Is the wrong claim coming from outdated training data, or from a third-party source the model is retrieving live? Different fixes apply.
3. Update owned content first. Make sure your website, knowledge graphs (Wikidata), and schema.org markup state the canonical truth clearly.
4. Earn third-party corrections. Get accurate information published on the authoritative sites the AI engines crawl. The mention is the signal.
5. Use platform feedback mechanisms. Where available (ChatGPT thumbs-down, Perplexity feedback), report incorrect statements about your brand. Slow but cumulative.
6. Monitor over time. Model retraining cycles are months. Check quarterly that corrections have propagated.

The insurance market emerges

A measure of how seriously the brand-risk side is taken: insurance products specifically covering AI-related brand damage are now in market. Munich Re launched aiSure covering AI performance failures. Willis Towers Watson partnered with Liberty Specialty Markets on similar coverage. Lloyd’s syndicates are underwriting AI hallucination liability policies.

The premiums and exclusions are still being calibrated, but the existence of the market is itself a signal. When reinsurers are willing to write coverage, they have priced the risk. The hallucination-driven brand-damage exposure is now a balance-sheet item, not a marketing concern.

For CMOs and CISOs: budget for AI brand monitoring is no longer optional. The cost of catching a hallucination early is measured in monitoring tool licenses. The cost of catching it late is measured in tribunal rulings, lost customer trust, and insurance premiums.

Every statistic in this piece follows Visigent's research standards: named sources, dated figures, corrections on request to help@visigent.com.