What ChatGPT Actually Searches When a Patient Asks "Best IVF Clinic in Delhi"

A patient types "best IVF clinic in Delhi" into ChatGPT. The reply names a few clinics, adds a short comparison and cites a handful of pages. What the patient never sees is the work behind it: the question is rewritten into several narrower searches, and each of those searches is answered by different pages.

This article is for hospital and clinic marketing leads who already know that AI search matters and want a working method. It answers one question. If an AI assistant decomposes a recommendation query, which sub-queries does it run, and does your site have a page that can answer each one? Our earlier guide to ranking in ChatGPT search covers the basics of being visible. This piece goes narrower: how to find and cover the hidden queries.

Nothing here promises a citation. It is a gap-analysis method, and the clinical content in any answer must be reviewed by the provider's own doctors.

Why the visible query is the wrong unit

Scale explains why this is worth the effort. ChatGPT reached 900 million weekly users in early 2026, and OpenAI has said that more than 230 million people a week ask it health and wellness questions. Similarweb's tracking suggests ChatGPT's share of generative-AI web traffic fell from about 76% to about 53% between June 2025 and May 2026 while Gemini rose to roughly 27 to 28%, so a single-platform strategy is already out of date (Similarweb, July 2026).

Those users do not type keyword strings. They ask conversational, multi-part questions. OpenAI's help documentation says that when ChatGPT search works with partner search providers it generally turns the question into one or more targeted queries. Google describes a similar behaviour in its AI features documentation: AI Overviews and AI Mode may use a "query fan-out" technique, issuing related searches across subtopics to find supporting pages.

The practical consequence is that "best IVF clinic in Delhi" is not one ranking contest. It is a bundle of smaller contests, and a clinic can lose the recommendation by having no page for one sub-question even if its homepage is strong.

What the fan-out for an IVF recommendation probably contains

Neither OpenAI nor Google publishes the exact sub-queries, so treat any list as a hypothesis to test. A reasonable decomposition of the Delhi IVF question has six facets:

• Entity and location: IVF clinics in Delhi and the National Capital Region, by area.
• Credentials: the doctors' qualifications, the clinic's registration and accreditation status.
• Process: what an IVF cycle involves, how many visits, what the timeline looks like.
• Cost: price ranges, what is included, how payment works.
• Evidence and reputation: reviews, patient discussions, and how others describe their experience.
• Alternatives: IUI versus IVF, and when a doctor suggests which.

Each facet maps to a sub-query that the assistant can run separately. A sentence such as "IVF cost Delhi what is included" is answered by a page that states what is included. A sentence such as "fertility specialist Delhi registered clinic" is answered by pages that show verifiable credentials.

Success-rate claims need particular care. They invite comparison, they are easy to present misleadingly, and they sit in a regulated, sensitive field. If the clinic publishes any figure, it needs a stated denominator, a time period, clinical sign-off and a check against the advertising rules that apply to the doctors involved. Our IVF marketing guide and the post on marketing fertility clinics sensitively cover that side.

Step 1: Capture the real sub-queries instead of guessing

You have three ways to get closer to the truth, in rising order of effort.

Read the interface. Where ChatGPT shows its search activity and a Sources panel, note which pages were cited and what the reply addressed. The sources show you which sub-questions had a good answer somewhere.

Use the API. OpenAI's web search tool documentation says the search action will usually, though not always, include the queries that were searched. Running your prompt set through the API with search enabled gives you a log of rewritten queries you can count and cluster. Store the prompt, the date, the model name, the queries and the cited URLs. Model behaviour changes, so dated logs matter more than any single run.

Ask the model to decompose. Prompt an assistant: "A patient asks X. List the separate questions you would need answered to respond well." This does not reveal the real fan-out, but it produces a sensible checklist that you can compare against the logged queries.

Run each prompt several times. Output varies between runs, so one answer proves nothing. What you want is the sub-queries that recur.

Step 2: Build a coverage map

Put the recurring sub-queries in a table and add three columns: the page on your site that should answer it, whether that page exists, and whether it answers the question in the first screen.

The point of the table is to find holes. Clinics commonly have a strong IVF service page and nothing on cost, nothing on the comparison with IUI, and doctor profiles that list titles without registration details. Those holes are where an assistant will reach for another site.

When writing the missing pages, give each one a single job and answer the question in the opening lines. An assistant's retrieval step works on passages, so a clear, self-contained answer near the top of a page is easier to use than one buried under a brand story. Our AI search optimisation guide describes the content patterns in more detail.

Step 3: Make the entity easy to verify

Fan-out queries often ask the same question about a clinic from different angles: where it is, who works there, what it is registered for. Inconsistent answers across your website, Google Business Profile and directory listings weaken confidence. Reconcile the name, address, phone number, doctor names and services across those sources before writing new content.

Check the basics a crawler needs. The pages must be indexable, served as HTML that contains the answer text, and not hidden behind a form. Google's documentation states that a page must be indexed and eligible to appear in Search with a snippet to be a supporting link in its AI features, and that there are no additional technical requirements. Treat that as a floor. Schema markup helps machines read the entity, and our healthcare schema guide covers the types to use.

Step 4: Measure with a fixed prompt set

Build a set of 25 to 40 prompts that reflect how Delhi patients actually ask. Include formal English, short phrases, and mixed Hindi and English. Vary the intent: choose a clinic, compare two treatments, ask about cost, ask about a procedure step.

For every run, record:

• whether your clinic is named;
• which pages were cited, and whether any of them are yours;
• which competitors and third-party sites are named;
• the sub-queries, where the tool exposes them.

Repeat monthly on the same prompts. Citations can vary between runs, so look for movement across several runs and several prompts. A change on one prompt is noise. A pattern across a facet, such as cost, is signal.

AiLLMz, one of our products, automates this kind of repeated prompt logging, but a spreadsheet and a disciplined weekly routine do the same job at small scale.

Failure modes to watch for

Optimising for the head term only. If your whole effort goes into the phrase "best IVF clinic in Delhi", you are competing for the one query the assistant is least likely to run literally.

Copying competitor language. An assistant that has several near-identical pages gains nothing from yours. A cost page that explains your own package components is more useful than a generic one.

Treating citations as the goal. A citation is not an enquiry. Check that the cited page leads to a clear next step, and that the call and form paths work on a phone. The conversion fixes post is a useful companion here.

Ignoring non-web evidence. Reviews and discussion forums feed some answers. If your Google reviews are old or unanswered, the reputation facet will be weak whatever your website says.

Where to start this week

1. Write ten patient prompts for your highest-value service line and run each one five times.
2. Log the sub-queries and cited URLs in one sheet.
3. Cluster the sub-queries into five or six facets.
4. Mark which facets have no page, or a page that does not answer in its first screen.
5. Write or fix the two weakest pages, with clinical review before publishing.
6. Re-run the same prompts in thirty days and compare facet by facet.

If you want the work done to a defined method, our ChatGPT ranking service page describes how we run it.