A patient asks an AI assistant which hospital in Nagpur does robotic knee replacement, and your hospital is named. The trouble is that you do not offer it. Or the answer lists a paediatric cardiac unit that closed two years ago, or a consultant who moved to another city. The error is rarely invented from nothing. In most cases the model is repeating something that is written down somewhere, and that somewhere is often your own site or a directory you forgot.
This post is for marketing teams at hospitals and multi-speciality centres who have found, or suspect, a wrong service attribution. It answers one question: how do you find the source of the false claim, and what do you change so it stops? It is a method, not a set of results. We do not report test outcomes here; you run the test on your own brand.
Wrong service claims are a patient-safety issue as well as a reputation issue. A person may travel for a procedure you cannot perform. Bring your clinical lead and compliance owner into the review before you publish any correction, and do not describe what a procedure involves without clinical sign-off.
Step 1: Build a 50-query test set
Fifty queries is enough to see patterns and small enough to repeat every month. Build it from real patient language, not from your service list. Split it roughly like this:
- Service existence (15): "does [hospital] do [procedure]", "[procedure] hospital in [city]".
- Specialty and department (10): "best [specialty] hospital in [city]", "[hospital] cardiology department".
- People (10): "[doctor name] specialty", "who is the head of [department] at [hospital]".
- Facilities and access (8): emergency, ICU, day care, international patient desk, cashless insurance.
- Comparison (7): "[hospital] vs [competitor] for [procedure]".
Include at least ten procedures you do not offer but a neighbouring hospital does. False positives hide there. False positives tend to hide there, because hospital pages share much of the same vocabulary.
Run every query on each surface your patients plausibly use: ChatGPT with search on, Google AI Overviews and AI Mode, Gemini, Perplexity and Microsoft Copilot. Use a clean browser profile and record the date, surface, location setting and whether the answer cited sources. Run each query at least twice. Answers vary between runs, so a claim that appears once is a signal; a claim that appears in most runs is a finding.
Step 2: Score each answer
Use a simple sheet with one row per query and surface. Columns that matter:
| Column | What you record |
|---|---|
| Hospital named | Yes or no |
| Claim type | Service, doctor, location, facility, accreditation, price |
| Claim correct | Correct, wrong, outdated, unverifiable |
| Sources shown | URLs the surface displayed |
| Quoted wording | The exact sentence that is wrong |
| Run count | How many of the repeat runs showed it |
Have someone who knows the hospital mark each claim, with the department head confirming anything clinical. Do not let the person who wrote the website be the only reviewer; they will read the claim through the intent of the page, not through what a machine would extract from it.
Sort the wrong claims by risk. A wrong visiting-hours answer is an annoyance. A wrong claim that you perform a transplant, a cancer therapy or fertility treatment is the first thing to escalate.
Step 3: Trace each wrong claim to a source
Models do not store a labelled list of facts about your hospital. They assemble answers from training data and, where search is on, from pages retrieved at the time of the question. For the retrieval surfaces you can often see the pages used. For the rest, you search for the wording yourself.
Where citations are shown. Perplexity and ChatGPT search display source links; Google's AI Overviews show supporting links. Open each one and search the page for the wrong term. Google's AI features documentation explains that these features may issue multiple related searches across subtopics and data sources, so the supporting page may not be the one you would have predicted. Check every link, not just the first.
Where nothing is cited. Take the exact wrong phrase and search for it in quotes on Google and Bing. Then search for the procedure name together with your hospital name, and for the procedure with each former doctor name. If the claim appears in a model answer without any retrievable page, it may come from older training data. Then the fix is to make current, correct pages abundant and consistent, because you cannot edit the model.
Classify what you find into these source types. Each has a different owner and a different remedy.
Your own site
This is the most common origin and the easiest to fix. Look for:
- Retired service pages still live, or redirected to a generic department page that still lists the service.
- Template boilerplate: a footer, mega-menu or sidebar listing every service across all branches, so a page about one branch appears to mention services offered at another.
- Branch pages copied from the flagship hospital with the service list intact.
- Old press releases and news items announcing a planned service that never launched or has since stopped.
- Doctor profiles that list a procedure the doctor performed at a previous hospital.
- PDF brochures and camp flyers, which are rarely updated and often outrank the web page.
Crawl the site and extract every page that contains the procedure name. A technical SEO audit can include this inventory. Match the list against the current service register that the clinical team signs.
Directories and aggregators
Listing sites, doctor-booking platforms, insurer network pages and medical tourism facilitators often hold their own service lists, written years ago and never revised. Search the procedure name on each directory that lists your hospital. Record the profile owner, the login and the last update date. Our local citations list covers the common Indian directories; use it as the starting inventory for this audit.
Structured data and profiles
Check your own markup. A medicalSpecialty value or a list of procedures in JSON-LD that is broader than reality will be read by search systems as a declaration. Also check Google Business Profile services and categories, social profiles and any knowledge-base entries your hospital has on Wikipedia or Wikidata. The schema markup guide shows which properties to keep narrow.
Third-party content
Local news, listicles, "top hospitals" pages, forum threads and reviews may state that you offer a service. You cannot edit these, but you can ask the publisher for a correction and publish an accurate source that outranks them.




