AI crisis management: the new communication problems large companies face

For years, a communication crisis had an identifiable origin: a story in the press, a viral complaint, a spokesperson who got it wrong. Today there is a new and much quieter kind. A customer asks Google or ChatGPT for your company's phone number, the AI answers with complete confidence and the number it gives is out of date. Nobody published anything against you, there is nobody to call for a correction, and the error repeats every time someone asks the same question. That is AI crisis management, and it is a problem that communications teams at large companies are only starting to see.

The real case: a client, a wrong phone number and 12 URLs

We saw it up close with a client that operates in several countries (we cannot name it for confidentiality reasons). When people searched Google Colombia for the company's phone number, the AI Overview showed as the main number one that was not its contact center. The AI had not made it up: it came from a network of third-party sites not affiliated with the company, the kind that publish "customer service numbers" for dozens of companies using the same template.

We then repeated the analysis in Chile, this time asking ChatGPT, Perplexity, Gemini and Claude directly, all four with web search enabled. The result was uneven: two assistants always gave the correct number, one failed intermittently and one failed in every test query. When we traced the origin, we found 12 URLs across 8 different domains publishing the outdated number.

The most uncomfortable finding was something else: the old number was still published on the company's own official site, on an FAQ page for another country and in a 2019 terms and conditions PDF. The AI was not hallucinating. It was correctly reading information the brand itself had never taken down.

What the data says: the phone number is what fails most

The case is not an oddity. US agency Seer Interactive asked seven AI models 178 questions about brand phone numbers and found that in 36% of cases the number given did not match the one on the brand's official customer service page. A key detail for communications teams: only 41% of the sources the AI cited were owned by the brand. The other 59% were third-party sites.

A second study, by the platform Insites, analyzed 10,000 local businesses in the United States and compared what ChatGPT and Perplexity answered against each company's Google Business Profile. The AI knew almost all of the businesses, but got every detail right only about half the time.

Indicator (Insites, July 2026)	ChatGPT	Perplexity
Businesses the AI identifies	93.59%	93.15%
Answers that fully match the Business Profile	55.58%	48.05%
Wrong phone number	30.09%	38.81%
Wrong website	15.59%	16.9%

On both platforms, the wrong phone number was the most frequent source of error. It matches exactly what we found in the field.

Why AI repeats outdated information

An AI assistant does not "know" your company's phone number: it builds the answer from what it finds published. And on the internet, old information does not disappear by itself. In the cases we have audited, the wrong fact reaches the answer through four routes:

• Third-party sites optimized for that exact search: unaffiliated pages that rank for "[brand] phone number" better than the official help page.
• Directories and company registries: a single data provider with the old number can feed several listings at once.
• The brand's own site: FAQ pages for another country, old PDFs, pages nobody remembers exist.
• No source of truth: without a verified Google Business Profile or structured data marking which fact is the official one, the AI has nothing to anchor to and averages what is out there.

That is why in many cases "hallucination" is the wrong word. If you are interested in that other scenario, when the model simply invents things, we cover it in the guide to fixing what AI gets wrong about your brand.

What comes next: wrong addresses, stores that no longer exist and old opening hours

The phone number is only the first symptom. Any operational fact that changes over time is a candidate for the same problem, and companies with many branches are the most exposed. The platform Searchable asked ChatGPT, Gemini and Perplexity more than 72,000 questions about UK retail chains and, as reported by Search Engine Journal, 64% of the companies received at least one false fact and one in ten location answers placed the store in the wrong postcode. In an earlier test with 165 London businesses, 93% had at least one basic fact wrong or missing.

The same article cites BrightLocal's annual survey: 45% of consumers already use ChatGPT or similar tools to look for local business recommendations, up from 6% the year before. More people asking means more people receiving the bad fact.

Fact that can fail	Where the error usually comes from	What happens to the customer
Customer service phone number	Third-party sites, directories, old pages on the brand's own site	Calls a number that does not answer or does not belong to the company
Store or branch address	Listings not updated after a closure or a move	Arrives at a location that no longer exists
Opening hours	Press releases and listings from previous years	Finds the branch closed
Website	Old or lookalike domains	Lands on a page that is not the official one
Policies and conditions	Outdated PDFs and legal documents	Complains about a condition that no longer applies

Each row is a social media complaint waiting to happen, and it is worth adding it now to the company's reputational risk matrix.

When the error stops being an accident: numbers planted by scammers

There is a worse scenario than outdated information: false information placed on purpose. In August 2025, The Washington Post reported the case of a Las Vegas businessman who searched Google for a cruise line's customer service number, called the number highlighted in the AI Overview and ended up giving his card details to an impostor. According to the AI Incident Database record, the fraudulent charges added up to USD 768.

In September 2026, the firm Vigilance Security documented the industrial version of the same mechanism: a campaign that floods the web with posts, PDFs and fake support pages so that ChatGPT, Gemini and Google's AI Overview hand out fraudulent phone numbers. According to Dark Reading, the researchers detected at least 374 affected companies, including airlines, banks and travel platforms. One caveat: the firm has not published the full list of pages and queries, so the figure cannot be verified independently.

For a large company, the consequence is direct. The scammed customer does not distinguish between "the AI got it wrong" and "the company failed me", and the complaint reaches your brand either way.

An AI crisis management protocol in five steps

1. Audit the critical questions. Phone number, address, opening hours, official site and policies, asked in Google's AI Overview, ChatGPT, Perplexity, Gemini and Claude, country by country. Save the exact answer and the date.
2. Trace every wrong fact back to its source. Search for the wrong number or address in quotes and list every URL that publishes it, including those on your own site.
3. Clean your own house first. Remove or correct old FAQ pages, PDFs and pages on the official domain. It is the only part that depends entirely on the company.
4. Reinforce the source of truth. A verified Google Business Profile, structured contact data on the help page and corrections in the directories that feed other sites. Publishing the correct fact in authoritative media, with a link building and press strategy, gives the AI more support for the current fact than for the old one.
5. Measure again. Repeat the same queries and keep a history. Models change, and a fact that comes out right today can fail again after an update.

If you want to start with the first step, the free AI audit shows in minutes what the assistants are saying about your brand.

AI is a spokesperson nobody trained

Large companies invest in preparing their spokespeople, writing key messages and monitoring the media. Today there is one more spokesperson, one that answers thousands of questions a day about the brand, that nobody hired and that repeats whatever it finds. Managing it does not require a new discipline, but applying the usual one to a new channel: knowing what is being said, correcting the source and measuring whether the correction worked.

At ESBUENISIMO LINKS we work that full cycle: auditing AI answers, tracing sources, publishing in the media and following up. Start with the free AI audit to see where your brand stands.