What AI assistants say about your hospital, and how to check
An AI answer audit is a repeatable check of what AI assistants tell patients about your hospital. You ask a fixed set of patient-style questions, record the answers, score them for accuracy and presence, then trace every wrong fact to its likely source. Most errors come from stale listings, thin specialty pages and old doctor profiles, which the hospital can fix itself.
A patient in your catchment no longer types a short phrase into a search box and scrolls. More and more of them ask a full question to an AI assistant and read one composed answer. That answer might name your hospital, a rival, or nobody. It might get your cashless empanelment wrong, list a cardiologist who left two years ago, or describe a service you closed. And nobody on your team will know, because nothing shows up in your analytics when a patient reads a wrong answer and quietly goes elsewhere.
This is why I now treat an AI answer audit as a routine hygiene task, in the same bucket as checking listings and reviews. It is not a strategy project. It is a structured way of asking the questions your patients ask, writing down what comes back, and working out why.
I have written separately about how hospitals get cited by AI search, which covers the longer game of becoming a source worth citing. This piece is narrower and more practical. It is the check you run before you decide what to fix.
Why the answer matters more than the ranking
Search teams grew up on rankings. You could see your position, you could see the competitor above you, and you could argue about the gap. An AI assistant gives the patient a paragraph instead. Sometimes it names three hospitals, sometimes one. Sometimes it hedges and tells the patient to consult a doctor, which is sensible, but then offers two names anyway.
Being in that paragraph is the new version of being on page one. Being described correctly inside it is the new version of a good snippet. A hospital can be mentioned and still lose the patient if the answer says it does not accept a particular insurer, or that it has no paediatric emergency, when both are wrong.
The other shift is that the assistant blends sources. Your website is one input. So are old news stories, directory listings, doctor profile aggregators, review platforms and forum threads. The answer is a composite, and a composite can be wrong in ways that no single source is. That is what makes checking it by hand worth the effort.
Setting up an AI answer audit: the question set
The quality of the audit depends almost entirely on the questions. Generic prompts like “best hospital in the city” tell you very little. Patients ask much more specific things, and they ask them in their own words, often in a regional language or in mixed Hindi and English.
I build the question set from real sources: contact centre call reasons, WhatsApp enquiry logs, the search terms report and what front desk staff say people ask at the counter. From those I write between twenty and forty questions, grouped by the job each one does.
- Discovery questions: which hospital to go to for a knee replacement, a cardiac second opinion or a high-risk pregnancy in a named locality.
- Fact questions: OPD timings, whether the hospital has a particular department, whether it is empanelled with a named insurer or scheme, parking, how to reach it.
- Doctor questions: who the senior specialists are in a department, and whether a named doctor still consults there.
- Comparison questions: your hospital against the two rivals patients most often weigh you against.
- Reputation questions: whether the hospital is good, what people complain about, whether it is expensive.
Keep the wording close to how people actually speak. “Is there a good heart hospital near Whitefield that takes my company insurance” is closer to reality than any keyword. Write some questions in the regional languages your catchment uses. Answers in those languages often draw on a thinner set of sources, and the errors there can be worse.
Once the set is written, freeze it. The whole value of the audit comes from asking the same questions each cycle so that you can see change.
Running the questions without fooling yourself
Assistants personalise. They remember earlier turns in a conversation, sometimes they remember the user, and some of them take location into account. If your digital manager runs the audit from a logged-in work account that has spent months asking about your hospital, the answers will flatter you.
Use clean sessions
Start each question in a new conversation. Use a browser profile or account with no history of asking about your brand. If an assistant offers a memory setting, switch it off for the audit. Do not ask follow-ups in the same thread unless the follow-up is part of the script.
Cover more than one assistant
Patients do not all use the same tool. Run the set across the handful of AI assistants that are commonly used in your market, plus the AI summaries that now appear at the top of ordinary search results. They draw on different sources and they disagree with each other more than you would expect.
Mind location and phrasing
Where an assistant uses location, run the discovery questions as if from two or three parts of your catchment, not just from the hospital’s own postcode. And run each question once. If you rephrase until you get an answer you like, you have measured your patience, not the assistant.
What to record, and how to score it
Record the full answer text, the date, the assistant, the language and any sources it shows. Screenshots are useful for internal presentations but a spreadsheet of text is what you will actually analyse.
I score each answer on four simple dimensions, each as a plain yes, partly or no. Keep it simple enough that two people scoring the same answer reach the same result.
Presence. Is the hospital mentioned at all where it reasonably should be? For discovery questions in your strongest specialties, absence is the most important finding in the whole exercise.
Accuracy. Are the facts stated about you correct? Timings, departments, empanelment, doctors, address, contact routes. One wrong fact makes the answer a “no” on accuracy.
Framing. Is the description fair? Assistants sometimes pick up an old complaint thread or a years-old news story and present it as the defining fact about you.
Sourcing. Where the assistant shows sources, are any of them yours? If every source is a third party, you have learnt that your own pages are not doing the work.
The AI search visibility audit tool on this site gives a starting framework for this scoring if you would rather not build the sheet from scratch.
The errors that show up again and again
The specific errors vary, but the families are consistent. In my experience most wrong answers about a hospital fall into a small number of types.
The first is the departed doctor. Aggregator profiles and old press coverage outlive a doctor’s tenure by years, and assistants happily recommend a specialist who now consults across town. This is the error clinicians notice first and complain about loudest.
The second is the missing service. The hospital offers something, but it lives in a PDF brochure, a banner image or a single line on a general page. Language models read text, and they need it stated plainly and somewhere they can find it.
The third is stale commercial information. Insurer and TPA empanelment changes, scheme participation changes, visiting hours change. The assistant repeats whatever it last found, which may be a directory listing nobody has touched since the hospital opened.
The fourth is identity confusion. Groups with similar names across cities, or a hospital that changed its name after an acquisition, get blended. Facts from one unit get attached to another. This is especially common for multi-unit groups whose unit pages look nearly identical.
The fifth is reputation drift. A handful of angry reviews or one old news item becomes the summary. You cannot delete history, but you can make sure it is not the only substantial thing written about you.
Tracing a wrong answer back to its source
Fixing the answer directly is not an option. There is no form to submit to an assistant. What you can do is fix the sources it draws on, and that means working out which ones those are.
When an assistant shows citations, start there. When it does not, search for the wrong fact in quotes on an ordinary search engine. The stale doctor name or the outdated timing will usually turn up on a handful of pages: an old listing, a doctor directory, a local news site, sometimes a forgotten microsite your own team built for a campaign years ago.
Your own properties come first. I have lost count of the times the source of a wrong answer turned out to be the hospital’s own old page, still live, still indexed and contradicting the new one. Next come the listings you control, especially your business profile on Google Maps and the major health directories. Then come the pages you can influence but not edit, such as media coverage and third-party profiles, where a polite correction request often works.
Write the source next to each error in the audit sheet. Over two or three cycles, you will see that a small number of sources explain most of the problems.
Fixing what you control, and accepting what you cannot
The fixes are mostly unglamorous. Retire or redirect old pages. Update every listing with current timings, departments and contact routes. Make sure each specialty page states, in plain text, what the department treats, which doctors consult there and which insurers and schemes apply. Put the facts people ask about into sentences on the page, not only into images or downloads.
Doctor pages deserve their own sweep. Every consulting doctor should have a current profile on your site with their department and qualifications, and departed doctors should come off promptly, with the page redirected to the department. Coordinate with HR so that exits trigger the change, rather than relying on marketing to notice.
Reviews and reputation are slower. The approach I describe in Google reviews are the hospital’s real front desk applies here too: respond properly, ask satisfied patients to review, and let the volume of current, balanced experience outweigh the old. Treat the AI summary as a lagging signal of what you do in the real world, much as reputation itself lags.
And some things you will not fix quickly. Assistants refresh on their own schedules. A corrected listing may take weeks or months to show up in answers. That is why the audit has to be periodic. A single check tells you the problem. Only the repeat tells you whether your fixes are working.
Who owns the audit, and how often to run it
This sits naturally with whoever owns search and local listings, usually inside the digital team. But the findings belong to several people. Doctor errors go to the medical administration and HR loop. Empanelment errors go to the TPA desk. Service errors go to the unit head. The digital team runs the check and routes the fixes; it should not be the only team that fixes things.
For most hospitals, a full run each quarter is enough, with a smaller spot check of the top discovery questions every month. Run an extra check after any big change: a new unit, a rebrand, the exit of a well-known consultant, or a news story. For groups, run each unit separately. A group-level answer can look healthy while a smaller unit is invisible or misdescribed.
Report it simply. Leadership does not need the full sheet. They need to know which discovery questions you are absent from, which wrong facts are being repeated about you, and what has changed since the last check.
Your first fortnight of checking
In the first week, build the question set from real enquiry data and get the unit head and the contact centre lead to review it. Then set up clean sessions and run the full set across the assistants your patients use, in each relevant language. Record everything in one sheet.
In the second week, score the answers, group the errors by type and trace each one to a source. Fix your own pages and listings immediately, because those are within your control and often explain the largest share of the damage. Send correction requests to the third-party sites that matter. Share a one-page summary with the unit head, with the three most important findings at the top.
Then put the next run in the calendar. The first audit is almost always uncomfortable. That is the point of doing it. The second one, a quarter later, is where you find out whether the hospital is becoming easier for an AI assistant to describe correctly, and that is the only result worth reporting.
Questions people ask
An AI answer audit is a structured, repeatable check of what AI assistants tell people about your hospital. You ask a fixed set of patient-style questions in clean sessions, record the answers, and score them for presence, accuracy, framing and sourcing. The output is a list of wrong or missing facts, each traced to a likely source, and a set of fixes owned by named teams. It is a hygiene task, not a campaign.
A normal SEO audit looks at your pages, rankings and technical health. An AI answer audit looks at the composed answer a patient reads, which blends your website with listings, directories, news and reviews. You can rank well and still be described wrongly. The two are connected, because your pages are an input, but the audit starts from the answer and works backwards to the sources rather than starting from the site.
Because patients increasingly decide before they ever call. If an assistant says your unit lacks a service, names a doctor who has left, or gets your insurer list wrong, the patient goes elsewhere and you never see the lost enquiry. The audit makes that invisible leak visible, and most of the fixes it points to sit with your own unit: doctor lists, department pages, timings and empanelment details.
For a single hospital, the first full audit usually takes a couple of weeks of part-time effort from one digital team member, with input from the contact centre and the unit. Building the question set properly is the slowest part. Later runs are much quicker because the questions, the sheet and the scoring rules already exist, and you are only looking for what changed.
No legitimate route exists to edit an assistant’s answer directly, and I would be wary of anyone who promises one. What you can do is correct and strengthen the sources those assistants draw on: your own pages, your listings, doctor profiles and third-party coverage. That work is slower than a paid fix would be, but it is real and it compounds over time.
The audit itself costs mainly staff time, and most of the fixes are content and listing updates the digital team should already be doing. The bigger cost is the enquiries you lose while wrong answers circulate, which is hard to size but real. I present it as a low-cost control that protects the return on everything else spent on patient acquisition, not as a new line of investment.
Discovery questions in your strongest specialties matter most, because absence there means you are not in the consideration set at all. Next come fact questions about empanelment, timings and doctors, because wrong facts turn interested patients away. Build all of them from real enquiry data, such as contact centre call reasons and WhatsApp logs, so that you are testing what patients actually ask rather than what marketing assumes they ask.
Include them from the first run. Write a subset of the question set in each language your catchment uses, including the mixed Hindi and English many people type. Answers in regional languages often rely on fewer sources, so errors can be more frequent and harder to spot. You will need a team member fluent in each language to score those answers properly rather than relying on translation.
Watch for wrong doctor attributions, services described as available when they are not, and any answer that appears to make clinical claims on the hospital’s behalf. The audit is not a clinical review, but it will surface statements about your clinical services that patients are reading. Those findings should come to you so that the correct information on the hospital’s own pages is clinically accurate and approved.
Usually very little beyond supporting the website fixes. IT may need to help retire old microsites, set up redirects, clean up duplicate pages across units and make sure important pages are not blocked from being read. They may also need to provide clean browser profiles for the audit. The audit itself is run by the digital team, not IT.
HR holds the most reliable record of which doctors have joined and left. Wrong doctor information is one of the most common and most damaging errors, so the fix is a simple process: every consultant exit or joining triggers an update to the website, listings and directory profiles. Without that link to HR, marketing finds out late and the old information lingers for months.
An agency can run the mechanics, but the question set and the interpretation should stay in-house. The questions have to come from your real enquiries, and the fixes depend on people inside the hospital: HR, the TPA desk, unit heads and clinicians. If you outsource it, insist on receiving the raw answers and the sources, not only a summary score, so you can check the work.
A short summary each quarter is enough. Show which important discovery questions the hospital is absent from, which wrong facts keep appearing, and what has improved since the last check. Frame it as reputation and demand protection. Boards understand that patients are increasingly influenced by what they read before they call, and a simple trend is more useful to them than detailed scoring.
A full run each quarter suits most hospitals, with a monthly spot check of the most important discovery questions. Run an extra check after a rebrand, a new unit opening, the exit of a well-known consultant or significant press coverage. Keep the question set frozen between runs so that changes in the answers reflect changes in the world, not changes in how you asked.
