Skip to main content
NewNMC’s advertising guidelines for doctors are in force from 6 Oct 2026. See what to change, and get a free compliance auditNew NMC advertising rules for doctors: what to change
AI Search

Hallucination Correction Playbook: When AI Gives the Wrong Phone, Address or Doctor

A step-by-step process for detecting, documenting and correcting AI answers that give a hospital the wrong phone number, address, timings or doctor, with source fixes, platform feedback and a repeat-check cadence.

Hallucination Correction Playbook: When AI Gives the Wrong Phone, Address or Doctor

A patient asks an AI assistant for the casualty number of a hospital and receives a number that belongs to a different department, or to a hospital that closed. Another is told that a surgeon works at a branch where they no longer practise. These errors are not rare curiosities. For a hospital, a wrong phone number or address is a service failure and sometimes a patient-safety one.

This article is for hospital marketing and digital teams, and for the operations leaders who own contact information. It sets out a process: detect the error, document it, classify its cause, fix the source, report it to the platform and check that it stays fixed. We describe a method, not a guarantee. No hospital can force an AI system to change an answer, and the platforms say so in their own feedback tools. What you can do is remove the wrong evidence and add clear right evidence.

Why the errors happen

Assistants assemble answers from retrieved pages, stored knowledge and patterns learned in training. Wrong contact details typically come from one of five causes:

  1. Stale sources. An old directory listing, press release or PDF with a former number.
  2. Conflicting sources. Your site, your Google Business Profile and a directory give three different numbers, and the system picks one or blends them.
  3. Entity confusion. Two hospitals with similar names, or two branches in one city, are merged.
  4. Missing sources. There is no accessible page for the fact, so the system guesses from nearby text.
  5. Model invention. The system fabricates a plausible-looking value. This is the hardest to fix, because there is no wrong source to correct.

Knowing which cause you face decides the fix.

Step 1: Detect

Make detection routine, not reactive. Build a small "facts audit" prompt set of twenty to forty prompts covering every fact a patient depends on:

  • Main number, casualty or emergency number, appointment number and WhatsApp number.
  • Address, landmarks and parking for each branch.
  • Visiting hours, OPD timings and 24-hour services.
  • Doctors by name and specialty, and the branch they practise at.
  • Services offered and not offered.
  • Insurance and cashless panels.
  • Website and appointment link.

Run them monthly on the assistants your patients use, three times each, in a fresh session. Also add detection sources that cost nothing: front-desk and call-centre staff who hear "the internet told me" complaints, social media mentions and Google Business Profile questions. Give staff a simple internal form for reporting a wrong detail with a screenshot.

Step 2: Document

For every error, create a record in a shared log. Capture the following.

FieldWhy it matters
Date, time and surfaceAnswers change; you need a timestamp
Exact promptReproducibility
Full answer text and screenshotEvidence for platform reports
Wrong fact and correct factClarity
Cited sources, if anyPoints to the source of the error
SeverityDrives urgency
Owner and statusAccountability

Set severity with clinical and operations leaders. A wrong emergency or casualty number, or a wrong location for a time-critical service, is the highest level and should be escalated the same day. A wrong consultation fee or an outdated doctor is medium. A slightly off description is low.

Never copy patient details into the log. If a user-reported error includes patient information, remove it before logging.

Step 3: Classify the cause

Use the cited sources, if any, to decide.

  • A cited page has the wrong fact. Fix that page or ask its owner to.
  • Cited pages disagree. Reconcile your own sources first.
  • No sources are cited and the fact is wrong. Likely invention or a stale memory. Add correct sources and report it to the platform.
  • The answer confuses you with another entity. Disambiguate your own entity pages, and link to official profiles.

Where an assistant shows no citations, run the same prompt with search enabled, which often reveals the page it is leaning on.

Step 4: Fix the source of truth

Treat one internal record as the master for each fact, and push it everywhere else.

  1. One owner, one file. Operations holds a verified record of numbers, addresses, timings, doctor rosters and services. Marketing publishes from it.
  2. Your website first. Put the facts in HTML on a contact or locations page, in text, not images. Give each branch its own page with its own number and address. Use telephone, address and openingHoursSpecification in structured data that matches the visible text. Our schema markup service and healthcare schema guide explain the types to use, and the glossary entry on NAP consistency covers why identical name, address and phone details matter.
  3. Google Business Profile. Update each profile and check for suggested edits that others have submitted. Our Google Business Profile guide covers verification and management.
  4. Directories and aggregators. List every place the hospital appears: health directories, insurer panels, map and navigation apps, government listings and old press pages. Submit corrections, keep a tracker of requests and follow up. Our reputation SEO service handles this at scale.
  5. Doctor profiles. Keep each doctor's page, branch and specialty current. When a doctor leaves, redirect or update the page and the directory entries promptly, and publish an accurate note instead of leaving the old one live.
  6. Old documents. Remove or update stale PDFs, brochures and press releases that carry old numbers. If you cannot delete them, replace them with updated versions at the same URL.

Disambiguation matters if you share a name with another hospital: include city, branch and legal name consistently in titles, headers and structured data, and link to official profiles from your pages.

Want this applied to your practice? Get a free ai search audit.

Book my free audit

Step 5: Give platforms the feedback

Feedback is useful and slow, so do it alongside the source fixes and not instead of them.

  • Google. Google documents that on an AI Overview you can use the thumbs-down control and choose "Report a problem", then add details, as described on its AI Overview feedback help page. Your most recent search and its results are included with the feedback. Keep your report factual: the prompt, the wrong fact, the correct fact and the URL of the authoritative page.
  • Chat assistants. Most have a thumbs-down or report control on each answer. Use it with a short factual description. Check each vendor's help centre for current options, including routes for correcting information about an organisation.
  • Local listings. For map listings, use the editing and reporting tools on the profile itself.

Attach your evidence, keep language neutral and do not claim legal wrongdoing in a routine factual report. If the error is about a named person and could be defamatory or privacy-related, involve your legal team before choosing a route.

Set expectations internally: platforms may not respond, and changes can take weeks or never appear.

Step 6: Verify and monitor

After fixes, re-run the failing prompt in a fresh session at one week, one month and three months. Record whether the error persists, changes or goes. If the old value persists after your sources are clean, repeat the platform report with the new evidence, and consider whether an uncorrected third-party source still exists. Search for the wrong number itself in a search engine in quotation marks. Pages that contain it are your remaining targets.

A hypothetical illustration: a 200-bed hospital in Pune finds an assistant giving a former switchboard number. The log shows it cited a regional directory. The team confirms the correct number, updates its site and Business Profile, files a correction with the directory and sends a platform report. At one month the assistant gives the right number on two of three runs; the team then finds a scanned brochure on a partner site with the old number and asks for its removal.

Safety rules for emergencies

For emergency numbers, do not wait for the platform. Publish the correct emergency and casualty numbers prominently on the homepage and each branch page, repeat them in Business Profile, train front-desk staff on what to say when callers mention wrong numbers and make sure the old number, if still active, forwards or plays a message with the right one. Telecom changes should include a step to update online listings.

Failure modes

  • Fixing only the website. The wrong value persists in directories and the AI keeps finding it.
  • Not having a master record. Every channel has a different version.
  • No owner for listings. Marketing assumes operations updated them, and operations assumes the reverse.
  • Over-reporting. Dozens of vague reports to platforms get less attention than a few precise ones.
  • Treating one run as proof. Answers vary. Test several times before declaring a fix a success.
  • Ignoring doctors who left. Outdated rosters are among the most common errors.

Make it routine

Add a facts-audit step to every phone-number, branch, doctor-roster and timings change, so listings are updated with the change itself. Review the error log quarterly for patterns: which directories repeat, which branches are most affected and which facts drift most. Our AI reputation monitoring service is one way to run the checks.

AiLLMz organises citations and publication workflows that support visibility across AI search and LLMs, which can help keep corrected sources consistent.

Where to start

  • Appoint an owner for the master facts record.
  • Run a first facts audit of thirty prompts and log every error.
  • Fix the emergency and main numbers everywhere this week.
  • Reconcile website, Business Profile and top ten directories.
  • File platform feedback with evidence for the persistent errors.
  • Re-test at one week, one month and three months, and schedule the monthly audit.
Filed underAI wrong information about hospitalChatGPT wrong hospital phone numberfix AI hallucination hospitalAI answer correction playbookhospital AI misinformation

About the author

Co-Founder & CTO · Gurugram, India

Arush builds the systems the rest of the company runs on — the patient-acquisition pipelines, the AI chatbots, the analytics that tie a single click to a booked appointment. If a campaign reports a number, it's because something his team built is tracking it.

Free guide · 15 pages

The Patient Acquisition Blueprint

A 90-day plan from first search to booked appointment.

The Pioneers Brief — one essay, three benchmarks, one teardown, every Thursday. Subscribe free →

Free ai search audit

Your next chapter starts with a conversation.

Three quick questions. We come prepared with a plan for your specialty and city.

  • A 30-minute call with a healthcare specialist
  • Your top three growth opportunities, in writing
  • Honest advice — even if that means not hiring us
  • No obligation and no hard sell
You’ll speak with a senior strategist
Consultations are hosted by Arush Thapar’s team.
Step 1 of 3Takes 30 seconds

What do you want to improve first?

Details stay privateNo obligationReply within 1 working day

Client portfolio · Healthcare websites

Work relevant to hallucination correction playbook when ai gives the wrong phone.

Explore client work →