How is brand hallucination rate measured?

Run a fixed prompt set about your brand — pricing, features, integrations, leadership, history — across engines on a schedule, then grade each answer against a fact sheet. Hallucination rate is the share of answers containing at least one material false claim, tracked per engine and per topic.

What is an acceptable hallucination rate?

There is no universal threshold, but practitioners treat errors on commercially sensitive facts — price, security posture, availability — as zero-tolerance items requiring a correction loop, while tolerating occasional slips on trivia. The severity weighting matters more than the raw percentage.

What Is Hallucination Rate?

Last updated: July 5, 2026 · Menra Team

Hallucination rate is the measured frequency with which AI engines state false facts about your brand — wrong prices, invented features, fabricated URLs, misattributed products. It converts anecdotal "ChatGPT said something weird about us" moments into a tracked metric with per-engine, per-topic baselines and thresholds that trigger a correction loop.

How is it measured in practice?

The method is an evaluation harness pointed at your own brand. Teams maintain a fact sheet — the canonical answers on pricing, plans, integrations, compliance, founding facts — and a prompt set designed to elicit those facts ("what does X cost," "does X integrate with Salesforce," "who owns X"). Each scheduled run grades answers as correct, incomplete, or false. Two design choices matter: grade material errors, not stylistic ones, and log the cited sources, because the citation usually reveals where the false fact lives on the open web.

What drives the rate up or down?

Coverage gaps. Models fill missing facts with plausible ones; a brand without a clear pricing page invites invented prices.
Stale third-party sources. Old review-site listings and outdated articles outrank a thin owned page in retrieval.
Entity confusion. Similar brand names produce misattribution, the hardest class to fix.
Retrieval vs. parametric answers. Grounded answers with citations hallucinate less than answers from parametric memory alone; engines and modes differ accordingly.

Example

A fintech finds 4 of 30 fact prompts produce errors in one engine — all traceable to a 2024 pricing snapshot on a comparison site. After the site updates and the source page is corrected, the next month's run shows the errors gone: rate as a before/after instrument, not just a scoreboard.

Related terms

See brand hallucination, hallucinated URL, and misattribution. The measurement workflow is part of tracking AI mentions.

Frequently asked questions

How is brand hallucination rate measured?: Run a fixed prompt set about your brand — pricing, features, integrations, leadership, history — across engines on a schedule, then grade each answer against a fact sheet. Hallucination rate is the share of answers containing at least one material false claim, tracked per engine and per topic.
What is an acceptable hallucination rate?: There is no universal threshold, but practitioners treat errors on commercially sensitive facts — price, security posture, availability — as zero-tolerance items requiring a correction loop, while tolerating occasional slips on trivia. The severity weighting matters more than the raw percentage.

Keep exploring

See how AI engines talk about your brand — track mentions across ChatGPT, Perplexity, Claude, Gemini and 5 more. Start with Menra