Hallucination

Large language models generate text token-by-token, optimizing for plausibility rather than truth. When an LLM hasn't seen a specific business fact in its training data, or when multiple conflicting facts exist in the web corpus, the model fills gaps by predicting the most statistically likely next word — which often sounds right but is wrong.

For local businesses, hallucinations cluster around: phone numbers (model invents a digit), addresses (confuses two nearby locations), business status (recommends a permanently closed location as open), and brand attribution (assigns your business to a competitor or conflates two similar names). The risk amplifies when source-level training data is thin — small cities, niche services, recently opened businesses.

When Google AI Overviews, ChatGPT, Perplexity, or Claude make a local business recommendation, hallucinations directly impact which business wins the user's click. If an AI recommends a closed competitor (hallucination) instead of your open business, you lose that conversion. If the AI invents a phone number for your business, users call a wrong number or assume you're unreachable.

Hallucinations are harder to fight than ranking losses because they're not about your on-page SEO or backlink profile — they're about whether the LLM's training data includes your correct NAP, recent reviews, and accurate business status. Monitoring AI visibility catches when hallucinations assign your brand to competitors or surface closed locations.

A hallucination differs from bad source attribution. Attribution errors point users to a real source but misinterpret it. Hallucinations invent facts with no source. An AI might cite a correct source but misread it (attribution error) or cite a non-existent source for a false fact (hallucination). For local SEO, the end result is similar: users get wrong information. But hallucinations are harder to correct because they're not tied to a specific webpage you can update.

Both expose your brand to AI-search risk. AI mentions monitoring detects when hallucinations or misattributions surface in LLM responses about your business.

You cannot eliminate hallucinations — they're a fundamental property of LLMs — but you can reduce the chance your business gets hallucinated:

• Consistent NAP across all directories ensures the LLM's training data has high-confidence identity signals
• Recent reviews on Google and other platforms confirm your business is currently open and active
• Schema markup (LocalBusiness, PostalAddress) makes your canonical data machine-readable
• Public business records (state registration, chamber membership) create additional sources LLMs can cross-reference

Monitoring tools like the AI Mentions API alert you when hallucinations surface, so you can respond with content fixes, schema updates, or directory corrections that increase the LLM's confidence in the truth.