When AI gets your business wrong: finding and fixing hallucinations.

Engines rarely invent from nothing. Almost every hallucination we've traced has a documentary source — which is good news, because documents can be fixed. The usual suspects: stale directory pages autogenerated from old data feeds, a previous tenant's record still attached to your address, outdated press coverage, conflicting name-address-phone records across platforms, and abandoned profiles nobody remembers creating.

You have to go look. Run your name and your key people's names through all five engines in fresh sessions, then go past identity into judgment: "what does [business] specialize in," "is [name] good," "[business] reviews." Record what each engine asserts, dated. (The fuller protocol is in our DIY audit guide.) Anything wrong goes on the list with two attributes: what the error is, and how much it costs. Wrong founding year is cosmetic. Wrong specialty list bleeds leads.

There is no customer-service line for a language model. Correction is source work. Browsing engines cite their inputs — read every citation and you'll usually find the stale or conflicting page feeding the error. Then three moves, in order: claim and correct the source where the platform allows it; displace it with stronger, correct pages where it doesn't; and reinforce your own entity data — schema, knowledge panel, fresh reviews — so engines have something better to lean on.

Then re-test weekly. Browsing engines typically flip within weeks of the source landscape changing. Base-model hallucinations are stubborn — they persist until a retraining cycle, which is a reason to start now, not a reason to skip the work. The fix you ship today is what the next snapshot learns.

Accuracy is a tenth of our visibility score, and it's the only component that can go negative in effect: being described wrongly can cost more than not being named at all. Audit what engines say, not just whether they say your name.

A composite from real engagements, details blended: a B2B SaaS company keeps being described by ChatGPT and Copilot as an e-commerce brand. Inbound leads skew wrong, buyers arrive asking for services the company does not offer, and the team assumes the AI is simply broken.

Tracing took an afternoon. Perplexity's citations on the "what does [company] specialize in" query included a stale business directory page — autogenerated years earlier from a syndicated data feed — listing the company under categories inherited from the previous tenant of its office address, an e-commerce operation that shut down in 2021. Two other directories had syndicated the same bad feed. The engines weren't hallucinating; they were faithfully reporting a corrupted record.

The fix ran the standard three moves: claim-and-correct on the two directories that allowed it; a takedown request on the third, displaced in the meantime by building out correct, service-specific profiles on the directories engines actually cited; and a reinforcement pass — Service schema on every product page, the knowledge panel's categories tightened, a fresh review push that named the actual product in buyers' own words.

Result: Copilot corrected in 3 weeks, Perplexity in 5, ChatGPT-with-browsing in 7. The base model held the old story until the following retraining cycle — then flipped without further work, because the record it re-read had been repaired months earlier.

The lesson generalizes: every week the source landscape stays wrong is a week the error compounds across engines. The work is rarely glamorous — it's directory archaeology — but it's deterministic in a way most marketing isn't. Fix the record and the answers follow.

Hallucinations recur. Indexes refresh, models retrain, a directory re-syncs from a bad feed — and an error you killed in March is back in August. This is a monitoring problem, not a one-time fix, which is why our ongoing tier re-runs the full query map weekly with dated screenshots. Catch the error the week it appears and it's an annoyance. Catch it a quarter later and it's a quarter of misinformed buyers.