When AI gets your business wrong

The practical response is to make your public information clearer, investigate the sources behind errors and keep a record of what changes. You cannot control every generated answer. You can control the quality of the information you publish and how quickly you respond when a material error appears.

Separate an invented answer from an outdated source

An AI hallucination is a plausible but inaccurate generated statement. Not every wrong answer has the same cause. A search-enabled assistant might faithfully repeat an obsolete directory listing. Another answer might confuse two organisations or produce an unsupported claim without a source.

Start with five categories: incorrect prices, inaccurate service descriptions, mistaken ownership, confused identity and outdated contact details. Treat statements about individuals, security incidents or contractual commitments as particularly sensitive. Preserve the exact wording before trying to correct it.

A single answer is an observation, not evidence that every customer receives that answer. Responses vary with the question, previous conversation, model, search mode, language and location.

Keep a useful error log

Test the questions a buyer would actually ask: what the business does, where it operates, what a service includes and how to contact it. Use fresh conversations and record the date, product, prompt and whether web search was enabled. Save the answer and its citations.

Then compare each factual claim with a current source you control. Mark it correct, incorrect, outdated or unverifiable. Do not count a positive description as a success if the underlying facts are wrong. Our guide to measuring AI search visibility explains how to keep these checks comparable.

Prioritise mistakes by likely impact. A wrong bank detail needs a different response from a slightly imprecise description of your service area. For a serious allegation, preserve evidence and involve the person responsible for communications or legal matters.

Correct the source first

If a cited page contains the error, update it or contact its publisher. Check old PDFs, company profiles, directories and archived service pages. Make clear which price or offer is current. If a service has ended, explain that on a suitable replacement page instead of leaving conflicting offers live.

Publish the correct facts in ordinary, readable text on your own website. Your legal name, trading name, contact details and service descriptions should agree across the homepage, contact page and terms. For a European audience, distinguish the countries you serve from the country where your company is registered or its servers are located.

Use Organisation structured data where it accurately represents the visible information. It helps describe your organisation; it cannot force an assistant to accept a correction. Link only to authentic official profiles. Do not create promotional encyclopedia entries simply to obtain an identifier.

Report the answer and check again

Use the feedback route offered by the product, including the incorrect statement and a reliable correction source. Avoid putting confidential customer or employee information into a public report. Where a product offers a specific privacy or removal process, follow that process rather than treating a general dislike button as a guaranteed removal request.

For search-based answers, make updated pages crawlable and include their canonical URLs in your sitemap. A crawl request is a discovery aid, not a deadline for an answer to change. There is no universal correction period: retrieval, indexing and model updates are different processes.

Repeat the same test later and add the result to the log. A changed answer is useful evidence, but one successful retest does not prove that the underlying error has disappeared everywhere.

Prevent avoidable contradictions

Assign an owner to public business facts. Review them after a price change, relocation, acquisition or service change, with a periodic check between those events. Give old promotional content an explicit end date and link to the current offer.

An optional llms.txt file can organise links for tools that use the proposal. It is neither a correction channel nor a requirement for visibility. Google states that special AI files are unnecessary for its generative search features. Google Search guidance.

Reliable hosting supports access to your current facts. ResaHost web hosting covers the technical foundation; editorial corrections and any specialist reputation work need their own named owner and agreed scope.

Common questions