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

Designing a Prompt Set to Measure Share of Voice in AI Answers for a Hospital

A method for building a weighted prompt set across specialty, city, language and intent, running it on several AI surfaces and turning the answers into a share-of-voice score a hospital board can track.

Designing a Prompt Set to Measure Share of Voice in AI Answers for a Hospital

Every hospital marketing team that has started looking at AI answers asks the same thing: how often are we mentioned, compared with the hospitals next door? The usual response is to type a handful of questions into a chatbot and screenshot the result. That produces an anecdote. A share-of-voice number needs a designed set of prompts, a fixed way of running them and a scoring rule agreed before the first answer comes back.

This article is for marketing leads who need to produce that number for a hospital or hospital group, and for the analysts who will run it. It describes the method layer: how to build the prompt set, weight it, run it and score it. It does not report results, because the results depend entirely on your hospital, your city and the day you run the test.

What share of voice means here

In AI answers, share of voice is the proportion of weighted opportunities in which your hospital appears, relative to all hospitals that appear. An opportunity is one prompt run on one surface. You can measure three levels of appearance, and it helps to keep them separate:

  • Mention. The hospital name appears in the answer.
  • Recommendation. The hospital is offered as an option for the user's stated need.
  • Citation. One of your pages is listed as a source.

A hospital can be cited and not mentioned, or mentioned without a link. Reporting only one conflates different things. We track all three and report mention rate as the headline, with recommendation and citation as supporting columns.

Build the prompt taxonomy

A flat list of fifty questions will over-represent whatever the team found easy to think of. Build the set from dimensions instead.

DimensionValues to includeWhy it matters
SpecialtyYour priority service lines, for example cardiology, orthopaedics, oncology, maternityVisibility differs sharply by service line
GeographyYour city, nearby cities, "near me" style, and a medical-tourism origin market if relevantLocal and tourist queries behave differently
LanguageEnglish, Hindi and the regional language your patients useAnswers and sources change with the language
IntentSymptom, treatment, cost, doctor choice, hospital comparison, second opinion, logisticsMaps to stages of the patient journey
Brand framingUnbranded, your-brand, competitor-brand, doctor-nameSeparates discovery from reputation
PersonaPatient, family member, referring doctor, international patientWording changes with the asker

You do not need every combination. Pick the cells that matter commercially and fill each with two or three phrasings.

Sizing the set

Fifty prompts is a workable starting size for one hospital. A sketch, with hypothetical allocations:

  • Twenty unbranded discovery prompts across your four highest-value service lines, five per line.
  • Ten doctor-choice and hospital-comparison prompts.
  • Eight cost and logistics prompts.
  • Six prompts in Hindi or a regional language.
  • Six branded prompts (hospital name, key doctors, a few reputation questions).

Write the prompts as a real patient would. "Best hospital for knee replacement in Pune" is more representative than "orthopaedic hospital Pune top rated". Pull phrasing from Search Console queries, call-centre notes, WhatsApp enquiries and Google's People Also Ask boxes. Avoid including any patient-specific details; use generic situations only.

Clinical content in the answers is not yours to judge. If the audit uncovers an answer with dubious clinical content about a procedure you offer, pass it to the clinical lead for review rather than rewriting it.

Weights

Not all prompts are equally valuable. Assign each a weight from one to five before running anything. Suggested criteria, which you should adjust:

  • Commercial value of the service line (margin, capacity, strategic priority).
  • Likely volume of the underlying question.
  • Proximity to booking, so "which hospital for bypass surgery" scores above "what is angina".

Store the weights in the same sheet as the prompts and lock them for the quarter. Changing weights mid-quarter makes trends meaningless. If you must change them, rebase and note the discontinuity.

Choose the surfaces

Run the same prompts on the surfaces your patients use. For most Indian hospitals that means Google, including AI Overviews and AI Mode where they appear, ChatGPT, Gemini and Perplexity. Add Copilot or Claude if your audience uses them. Google states that AI Overviews and AI Mode traffic is counted in the Web search type of the Search Console Performance report, with no separate breakout, so the prompt set is your main way of seeing appearance rather than clicks.

Decide whether you test with browsing or search turned on. Answers with live retrieval cite pages; answers from model memory often do not. Record the setting.

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

Book my free audit

Run protocol

Consistency matters more than sophistication. Write these rules down.

  1. Fresh sessions. New chat or logged-out browser for each prompt, no history.
  2. Location. Run from a consistent location for local prompts, and state it. Use a VPN or a team member in the target city for out-of-city tests.
  3. Repeats. Run each prompt three times per surface, because answers vary between runs. Record each.
  4. Window. Run all surfaces within the same two-day window, on a fixed day of the month.
  5. Raw capture. Save the full text of each answer and the cited URLs, not only your interpretation.
  6. No patient data. Never type real patient details into any tool.

For fifty prompts across four surfaces at three repeats, that is six hundred runs. That is manageable only if you split it across people or semi-automate capture. Check each surface's terms before automating, since some prohibit scraping.

What to record per run

FieldExample values
promptid, rundate, surface, mode (search on/off)P-014, 2026-11-02, Perplexity, search on
language, locationHindi, Pune
answer_textFull text
hospitals_namedList, in order of appearance
your_position0 if absent, else rank among named hospitals
recommendedYes / No
cited_urlsList, each tagged as own / competitor / directory / media / other
accuracy_flagCorrect / wrong detail / unclear

Accuracy is worth capturing from the start. A mention with wrong information is a risk, not a win.

Scoring

Per run, assign:

  • Mention: 1 if named, otherwise 0.
  • Recommendation: 1 if named in a way that suggests choosing it, otherwise 0.
  • Citation: 1 if one of your pages appears in the sources.

Take the average across the three repeats of each prompt. Multiply by the prompt weight. Then compute:

Share of voice = weighted mentions of your hospital divided by the weighted sum of mentions for all hospitals named in the set.

Report the denominator hospitals as well, because your share depends on who else is in the answers. A worked example, with entirely hypothetical numbers: suppose a 200-bed hospital in Pune is mentioned with a weighted score of 36 out of a total of 240 weighted mentions across all named hospitals. Its share of voice is 15 per cent. Next quarter, 48 out of 255 gives about 19 per cent. The figures are illustrative only and say nothing about how any real hospital performs.

Report by cell, not just overall: share by service line, by language and by surface will show where to act. A flat overall figure can hide a service line at zero.

Failure modes

  • Prompts that favour you. If the set was written by people who know your strengths, unbranded prompts drift towards them. Have someone outside the team review.
  • Single runs. One run is a coin flip. Use repeats.
  • Mixing modes. Search-on and search-off runs measure different things.
  • Ignoring wrong answers. Share of voice rises while the answers carry a wrong phone number. Track accuracy alongside.
  • Moving the goalposts. New prompts added without versioning make trend lines unreliable. Version the set, and keep a stable core of at least thirty prompts across quarters.

Reporting

Give the board three things: the headline share, the trend by cell, and the five source pages or domains that appeared most often in answers where you were absent. The last item turns measurement into action. It tells you which directory, media outlet or competitor page the AI is relying on.

AiLLMz organises citations and publication workflows that support visibility across AI search and LLMs, and can sit alongside the tracking sheet described here. Our guide to AI search optimisation and the AI search optimisation service cover the work that follows measurement, while LLM citation building addresses the source gaps this audit exposes. For ongoing reporting set-up, see SEO reporting and analytics.

Where to start

  • Pick four service lines and the languages your patients use.
  • Draft forty to fifty prompts across the taxonomy and weight them.
  • Have a clinician and a call-centre lead review the wording.
  • Run the first baseline on one fixed week, with three repeats per surface.
  • Publish the method to the team before the numbers, so nobody disputes the scoring afterwards.
  • Repeat monthly, and re-weight only once per quarter.
Filed underAI share of voice measurementAI visibility prompt setmeasure hospital visibility in ChatGPTAI answer tracking healthcareLLM brand mention tracking

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 designing a prompt set to measure share of voice in ai answers.

Explore client work →