DRAFT — attorney review required before publishing. This page has not been reviewed by a lawyer. Statements marked [ATTORNEY] need confirmation before the page goes live.
AI use policy at ClinicRankPro, Georgetown, TX
ClinicRankPro uses AI models to draft and check, never to decide or publish. Models draft report narratives, summarize audit data, and draft review replies and emails for a person to review. They never see patient data, never give clinical advice, and run behind a private gateway we host. Every published word is reviewed by a named person.
Reviewed by [OWNER NAME], last reviewed 16 September 2026. This page describes our working practices and is updated when they change.
What do we use AI for?
Four jobs, all of them producing drafts for a person to read.
- Drafting report narratives. The free reports and monthly client reports start as structured audit data. A model turns that data into plain-English findings, following a fixed template.
- Summarizing audit data. Picking the three biggest problems out of a long list of checks, with the evidence for each.
- Drafting review replies and emails. Suggested replies to public reviews, the Monday client update, and outreach emails, all held for approval.
- Classifying form submissions. Spam check and clinic type, from the practice name and website only.
What do we never use AI for?
- Clinical advice. No model writes a medical claim, a treatment recommendation or a diagnosis on a client’s behalf. Clinical statements on client pages are drafted from facts the practice supplied and reviewed by a licensed provider at the practice.
- Publishing without human review. Nothing a model produces is sent, posted or published until a person has read it. There is no auto-publish path.
- Handling patient data. Models never receive protected health information. See the HIPAA Statement.
- Writing reviews. No AI-written reviews, ever, for any client. It breaks Google’s and FTC rules.
- Making up identifiers. Any identifier written to a client site, such as a schema type or directory URL, is verified against the live source in the same run. Models never invent one.
Who reviews AI output?
[OWNER NAME], by name, on every report and every published word. Prospect reports are held until approved. Monthly reports are generated on the 3rd, reviewed on the 4th and sent by the 5th. Weekly updates sit in the outbox as drafts until read. Review replies and website content also go to the client for approval before they go live. Anything that fails our checks twice is marked for human review with notes on what to fix, and is never sent as is.
What data goes to the models?
Public business data and our own audit data, nothing else. That means practice name, address, phone, website, providers and credentials, services, review counts and public review text, rankings, competitor listings, and the results of our technical checks. What never goes to a model:
- Patient names, appointment data, call recordings, or anything from a booking or health record system
- Your name, email and phone from a form submission. The intake step strips them before the spam check and keeps only the email domain
- Page HTML, credentials, or account access of any kind
A rule check runs on every draft and fails it if it contains anything that looks like a Social Security number, a date of birth, or a patient name or record reference. The reviewing model is told the same rule.
Where do the models run?
Behind a private, self-hosted gateway on hardware we control. Our tools talk only to the gateway; the gateway holds the provider keys and routes each task to a model from [MODEL PROVIDERS] through their business APIs, or to a model running locally on the same machine. Each task has a preferred model and fallbacks, and a monthly budget cap stops spending when it is reached.
Under the business API terms of the providers we use, prompts and outputs are not used to train their models. [ATTORNEY] Confirm each provider’s current API data-use terms and record them in the sub-processor list. Providers may keep prompts for a short period for abuse monitoring; we record each provider’s stated retention period and prefer the shortest available. Our own logs keep the prompt data and outputs for each job so a report can be traced, for the retention periods in the Privacy Policy.
How accurate is it?
Every number in a report must trace to a line in the source data, or the draft fails. Rankings trace to a dated search result. Search volumes trace to the data provider and are labeled as estimates. Speed scores trace to a PageSpeed test with its date. Anything the audit could not check is written as “to be confirmed on the call”, never as a finding. Timelines and projections are labeled as typical ranges, not promises; see the Terms and Conditions.
How do we test for bias and hallucination?
- Two model families. The model that checks a draft is chosen from a different family than the one that wrote it, so it does not share the writer’s blind spots.
- Rule checks. Each draft is checked against a written rule set: numbers traceable, no banned words such as “guaranteed” or “best”, no patient-data patterns, no invented identifiers, estimates labeled.
- Bounded revisions. A draft gets at most two revision rounds. After that it is marked for human review rather than looping.
- Fixture tests. The pipeline is tested against fixed sample data where the right answer is known, before any change goes live.
- Human spot checks. [OWNER NAME] compares a sample of drafts against the raw data each month and logs anything a model got wrong, so the prompts and rules can be corrected.
Can clients see which parts were AI-assisted?
Yes, on request, for any deliverable. Each job keeps a trace: which sections a model drafted, which model, what the checker flagged, and what changed in human review. Ask and we show you. Clients are told at onboarding that drafts are AI-assisted, and this policy is linked from every report. If you would prefer a deliverable with no AI drafting at all, say so and we will do it by hand at the same price.
How do you contact us?
Questions about this policy go to [EMAIL] or the contact page. Related pages: Privacy Policy, Security, California Privacy Notice, how we handle compliance.
Frequently asked questions
Is my report written by AI?
Partly drafted, fully reviewed. Automated tools gather the data, a model drafts the narrative from that data, a second model checks it against our rules, and [OWNER NAME] reads and edits every report before it is sent. Nothing goes out unread.
Does any patient information go to an AI model?
No. Models see public business data and our own audit results: practice name, address, services, providers, public reviews, rankings. They never see patient names, appointment data, call recordings, or your booking or health record systems. Contact details from our forms are stripped before the spam check.
Where do the models run?
Behind a private gateway we host on hardware we control. The gateway routes each task to one of [MODEL PROVIDERS] through their business APIs, or to a local model. Provider keys live in the gateway, and no client data is used to train any model.
Can I ask which parts of my deliverable were AI-assisted?
Yes, at any time. Every report keeps a trace of which sections were drafted by a model, which model, and what was changed in review. Ask and we show you. Published website copy is disclosed the same way.
How do you stop AI from making things up?
Every number must trace to a source in the audit data or the draft fails. Anything the audit could not check is written as “to be confirmed”, not as a finding. A checker from a different model family reviews each draft, and a person reads the result. Estimates are labeled as estimates.
Will you write our medical content with AI?
We use AI to draft, never to decide. A model can produce a first draft of a service page from your approved facts. A person edits it, a licensed provider at your practice reviews the clinical statements, and you approve it before publication. AI never gives clinical advice on your behalf.