What Is an AI Source Map?

An AI source map is a log of the outside websites that AI tools cite again and again when they answer your patients' questions. For a health practice, it turns the vague advice to get more mentions into a short list of named places where you can fix or earn a listing. It takes an afternoon to build and needs no paid tool.

How an AI source map works

When an AI tool answers a question such as who treats thyroid conditions near me, it often shows the pages it used. Those pages are its sources. They are usually not your website. They are other places that describe practices like yours.

An MGMA-published study shows what that mix can look like. The study analyzed 33,340 ChatGPT citation events across 200 practices, 20 metro areas and 10 specialties, using GPT-5.5 through OpenAI's API in May 2026. Hospital-system staff rosters were 27.5% of the cited sources. Specialty association credentialing pages were 22.1%. Practice-owned websites were about 23%. National medical directories were 8.3%, state medical board profiles 5.1% and review aggregators 3.8%. Google Business Profile produced zero direct citations, which the author attributes to Google not exposing reviews as crawlable web addresses. The author runs a commercial service for healthcare, so treat the study as one data point and not a rule for your specialty.

The point of the study is the spread. About three quarters of the cited sources were off-site. If you only work on your own website, you are working on one source in four. A map shows you the other three.

Here is how to read what you collect. Sort each site into one of four groups.

• Shown in most runs. These are your primary targets.
• Shown in some runs. These are secondary.
• Shown once. These are the long tail, and you can skip them for now.
• Not shown. Ignore them.

I give no percentage cutoffs between the groups, because I know of no published standard. Use your own judgment and keep the same rule each time.

Why do runs matter? Answers vary. The same question asked twice can give different wording and different links, which is the subject of AI answer variability. So you run each question several times, and look for sites that keep returning.

Each site then gets one of four moves.

• Claim or correct your listing there.
• Add a profile that is missing.
• Improve the page of yours that is already cited.
• Skip it, because it is closed editorial content that you cannot reasonably join.

Do the cheapest accurate move first. A wrong phone number on a directory that appears in half your runs is a better first fix than a long outreach campaign.

There is a line I will not cross, and I would not cross it for a client either. Paying a publisher to quietly rewrite an existing article, swapping a competitor out of a "best of" list, or placing a "best clinic" piece without disclosure are not honest moves. Google's spam policies define link spam as links created primarily to manipulate rankings, and list paying money for links or for posts that contain them as an example. See backlinks for how links, mentions and citations differ. The map is for finding where patients and tools already look, so you can be accurate and present there. It is not a way to buy the answer.

Example: an AI source map at a health practice

Take a chiropractic and rehab clinic with two locations in a mid-size city. This example is illustrative, not a real clinic or a transcript. The owner writes ten questions in patient language, such as "chiropractor for sciatica near me" and "who treats whiplash after a car accident in [city]". She runs each one in ChatGPT with search on and in Google AI Mode, and again a week later. For every answer she copies the cited sites into a sheet with the question, the tool and the date.

After two rounds, the sheet shows a pattern. A regional health directory shows up in most runs. A state chiropractic association's find-a-provider page shows up in some. A local magazine's "best of" list appears once. A hospital system's physician directory appears often, but it lists only physicians on its staff, so the clinic cannot join it.

Then she checks each repeated site for her clinic.

• On the regional directory, the clinic is listed under an old name and the second location is missing. She claims the profile and corrects both.
• On the association page, her two chiropractors are listed, but one has an outdated specialty. She submits a correction through the association's member portal.
• On the magazine list, she is not included. She notes it for later and does not pay for placement.
• On the hospital directory, she cannot be listed. She skips it.

She then changes her own site as well. One of her pages is cited in some runs, so she reads it as a stranger would. The answer is buried in the fourth paragraph, so she moves it to the top.

She repeats the questions in a month. She cannot promise a change in the answers, because tools vary and she changed only a few things. What she has now is a short, dated list of places where her details were wrong and are now right. That is a measurable result, which she can check again each quarter.

How to check yours in 60 seconds

1. Open ChatGPT in a temporary chat with search on, and open Google AI Mode.
2. Ask the same patient question in each, such as "best [specialty] in [city]".
3. Copy every link or source shown into a sheet, with the question, the tool and the date.
4. Circle any site that appears more than once.
5. Search your practice name on each circled site and note whether you are listed correctly.

Repeat with two more questions. A site that appears across questions or tools is where to start.

What to do about it

1. Write ten patient questions using the words from your intake forms, front desk calls and reviews.
2. Run each question in at least two tools, and run it again on another day.
3. Log every cited site in one sheet, with columns for the question, the tool, the date and the site.
4. Sort the sites by how often they appear.
5. For each repeated site, choose a move. Claim, correct, add or skip.
6. Fix wrong names, phone numbers, services and provider details first. See local citations for listings that carry your name, address and phone.
7. Earn mentions only where they are real, such as associations, local press and podcasts.
8. Do not buy placements, rewrites or reviews.
9. Do not paste patient information into any AI tool while testing. Use invented scenarios.
10. Rebuild the map every quarter, and record what changed.