What Is Grounding in AI?

Grounding is the practice of tying an AI model's answer to real outside sources, such as current web pages, instead of letting it answer from memory alone. For a health practice, grounding is the step where your pages can become the evidence behind what an AI tells a patient. Google's developer documentation describes grounding with Google Search as connecting the Gemini model to real-time web content.

How grounding works

A large language model is the software behind ChatGPT, Gemini and similar tools. It learns patterns from a huge body of text and then predicts what words should come next. That learning stops at a point in time. If your practice launched a weight loss program last spring, a model working only from memory may know nothing about it. It can fill the gap with something that sounds right.

Grounding closes that gap by giving the model something to read before it answers. Google's documentation for grounding with Google Search lays out the steps.

• The patient asks a question.
• The model decides whether a web search would help.
• If it would, the model writes one or more search queries and runs them.
• It reads the results.
• It writes an answer from what it read, and it can attach citations that link specific text to the source URLs.

Google says basing responses on real-world information lowers the chance that the model invents facts. Lowers is the honest word. Grounding reduces the problem of a hallucination, where an AI gives a confident but false answer. It does not remove it, because the answer is only as good as the pages the tool found and how it read them.

Three details matter to a practice owner.

First, the tool makes the call about when to ground. A question about the capital of France may not need a search. A question about a functional medicine doctor near you that takes new patients depends on facts that change, which is the kind of question where a search helps.

Second, grounding depends on access. Google Search Central says there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimizations are necessary. A page has to be indexed and eligible to show with a snippet. OpenAI runs a separate crawler, OAI-SearchBot, that surfaces sites in ChatGPT search. Its settings are independent from GPTBot, the crawler OpenAI uses for training. You can allow one and block the other.

Third, something other than your own choices can block the door. Cloudflare, a company that sits in front of many websites, announced in July 2025 that it was changing its default to block AI crawlers. If your site runs through Cloudflare, ask your web person what your setting is today.

A useful way to hold the two layers: the model's memory gives the background, and grounding gives the freshness. Your pages can only change the second layer quickly, because the memory (see training data) updates only when a new model version ships. Some tools may also reuse saved copies of pages, so a fix on your site can take a while to show in answers. I have not seen vendors publish consistent timing, so in my practice I recheck after a few weeks and do not undo a correct change early.

Grounding is a close relative of retrieval-augmented generation, the research name for retrieving documents first and writing from them. Grounding is the goal. Retrieval is one common way to reach it.

Why does this matter more for health than for, say, a pizza shop? Patients ask health questions when they are worried, and the answer can steer who they call. A tool that reads your page can pass along your actual approach, your credentials and your location. A tool that works from memory passes along whatever it half remembers, mixed with what it knows about other clinics. The practices that get described accurately are usually the ones whose pages say plainly what they do.

Example: grounding at a health practice

Picture a hormone clinic with three providers in a mid-size metro. It has a page titled "Bioidentical hormone therapy for perimenopause." The page opens with two plain sentences saying who it treats, names the clinician who prescribes, lists the lab panels it orders and gives the city and the booking line.

A patient asks a search-enabled tool, "Who offers bioidentical hormone therapy for perimenopause near me?" The tool decides a search would help, runs a few queries and reads the results. This example is illustrative, not a transcript, but the grounded answer might read as follows.

Several clinics in the area offer hormone therapy for perimenopause. One lists physician-led care with baseline lab testing and a consultation line on its service page. [1]

The number in brackets is the citation, a link back to the clinic's page. Now picture the same clinic with a page that says only "We offer personalized wellness solutions" over a stock photo. The tool has no facts to write from and nothing worth linking. It picks a different source or gives a generic answer.

Here is a real data point for contrast. Health Biz Scale checked 764 established US health practices between May 11 and Aug 20, 2026, with one check per practice. We asked GPT-4o-mini, through OpenAI's API, 2,292 patient-style questions. It answered from memory with no web search, so nothing was grounded. Of the 3,554 practice websites it named, 1,153 (32.4%) did not load, and only 6 of the 764 practices (0.8%) were named. The ChatGPT app searches live and can answer differently. The test shows what happens when a model has no sources to read. Grounding is the difference between guessing and reading, and reading only helps you if your page is there to be read.

How to check yours in 60 seconds

1. Open a tool that searches the web, such as Perplexity, ChatGPT with search turned on, or Google AI Mode.
2. Ask a patient-style question about your main service in your city, for example "Who offers thyroid testing in functional medicine near [your city]?"
3. Look for source links or numbered citations under the answer.
4. Open every one and see whether any is your page, or a page that lists you.

Also read the written answer itself, not only the source list. A tool may not fact-check what it retrieves, so a stale price or a provider who left can be repeated as if it were current. Some of these errors are self-inflicted, such as two different prices on two of your own pages. Fixing them is covered under AI brand accuracy.

Write down what you find. If no source is yours, a tool answered this question without reading your site, and that is the gap to close.

What to do about it

1. Put the facts a tool would need on one page for each service. Cover who treats, what is treated, where, how it works and how to book.
2. Write in plain, specific sentences. Replace slogans with the actual service, provider and city.
3. Keep key pages public and indexed. Check that no login wall, noindex tag or script is hiding the text.
4. Open your robots.txt file (the file that tells crawlers what they may visit) and confirm it does not block OAI-SearchBot. Ask whoever runs your site about any Cloudflare setting too.
5. Make your hours, providers and services match your listings and profiles, since a tool may read those too. See AI citation for how sources get shown.
6. Update pages the week a service, provider or price changes.
7. Retest every month with the same question and write down the cited sources. Grounded answers change as tools and pages change, so one check is a snapshot and a monthly log is a trend.