AI Search · LLM
What does ChatGPT already
say about your practice?
Ask it. Most practice owners have not, and a meaningful share of the ones who do find something wrong — a procedure they stopped offering, a surgeon who left in 2019, a location that closed, or a confident recommendation of the competitor down the road. This page is about the model’s existing beliefs: where they came from, and what can actually be done about them.
The problem
A model can be confidently wrong about your practice, and nothing tells you it is happening.
Assistants answer from two places: what the model absorbed during training, and what it retrieves live at the moment of the question. The retrieved half you can influence quickly — that is ordinary visibility work. The trained half is a snapshot of the web from some point in the past, and it can carry a version of your practice that is years out of date.
The failure is silent in a way that search never was. A bad ranking is visible on a results page. A patient being told your practice does not perform a procedure you have performed for three years produces no impression, no click and no complaint. They simply go somewhere else, and you never learn why.
The errors we find fall into a few recurring shapes: a departed physician still listed as practising, a discontinued or newly-added procedure described wrongly, a merged or relocated office given the old address, credentials attributed to the wrong person, and — the most commercially annoying — a confident recommendation of a competitor for a procedure the practice specialises in.
There is no correction form. You cannot email a model and ask it to update its beliefs. What actually moves the answer is changing what the systems can retrieve and verify now, consistently enough and in enough places that the live evidence outweighs the stale training impression. That is slow, it is mostly unglamorous reconciliation work, and it is the only mechanism that exists.
One honest caveat, because the field is full of people selling certainty here: nobody outside these companies can guarantee what a model will say. What can be done is measure it, fix what is fixable, and watch the answer change. Anyone promising more than that is describing something they cannot deliver.
What the work is
Four steps, in this order.
This work is diagnostic before it is constructive. You cannot fix what a model believes until you have written down what it currently believes, per platform, in its own words.
Audit what each assistant currently says
A fixed set of questions — who performs this in our city, does this practice offer that procedure, who is the surgeon, where are they located, are they board certified — run against ChatGPT, Gemini, Claude, Perplexity and Copilot, with the answers recorded verbatim. Platform-by-platform, because they disagree with each other constantly: one may have you right and current while another is describing your practice as it was in 2022.
Trace each wrong answer to its source
Most errors have a findable origin — a stale directory entry, an old press release, an abandoned profile on a platform nobody logs into, a society listing never updated after a departure, a scraped aggregator repeating something from years ago. Identifying the source matters more than counting the errors, because the source is the only thing you can actually change.
Correct the retrievable record, everywhere at once
Your site, your Business Profile, the medical directories, the society and board listings, the aggregators. Same name, same address, same phone, same roster, same procedure list, same credentials. Consistency across independent sources is what a retrieval system reads as confidence, and it is the mechanism by which a current fact eventually outweighs a stale impression. Tedious, and it works.
Re-test on a schedule and watch the answer move
The same question set, re-run monthly, with the answers diffed against the previous run. That is how you know whether a correction propagated, and it catches new drift — a model update can reintroduce an old error, and a practice that only checked once will not notice. This is measurement by observation, which is the only kind available here, and we label it as such.
What you get back
A transcript, not a score.
The deliverable is the assistants’ actual words about your practice, per platform, per question, month over month — with the corrections made and whether each one has taken hold yet. Where an error persists, we say so and say why, rather than reporting an improved composite number.
It is worth being clear about the size of this. For most practices today, assistant answers are a small share of patient research and a large share of the impression a patient forms before they ever visit your site. The reason to do this work now is not volume. It is that being described wrongly costs patients invisibly, and that the corrections also improve ordinary search, so the effort is not spent twice.
- Verbatim answers from ChatGPT, Gemini, Claude, Perplexity and Copilot
- Every wrong answer traced to the source that produced it
- Directory, society and aggregator records reconciled to one truth
- Monthly re-test with answers diffed against the prior run
- Persisting errors reported as persisting, not averaged away
Questions
LLM SEO, answered plainly.
Can you make ChatGPT recommend my practice?
No one can promise that, and you should treat anyone who does with suspicion. What is achievable: make sure the assistants can reach your site, that what they find is accurate and consistent with every other source about you, and that your credentials and procedures are legible as structured facts. That measurably changes how often practices get named. It is influence, not control, and we describe it that way in the reporting.
How do I find out what an assistant says about my practice right now?
Ask it directly — "who performs [procedure] in [your city]", "tell me about [your practice]", "is [your surgeon] board certified". Do it in ChatGPT, Gemini and Perplexity separately, because they will disagree. That five-minute exercise is genuinely worth doing before you talk to anyone about this, including us. The free AI Visibility Grader runs a structured version of the same check.
An assistant said something factually wrong about us. What do we do?
Find the source rather than the symptom. It is almost always a stale record somewhere retrievable — an old directory entry, a society listing not updated after a departure, an abandoned profile. Correct it at every source that carries it, then re-test monthly. Corrections to retrieved facts can show up in weeks; something baked into training data takes longer and may need the weight of consistent current evidence to override.
Is LLM SEO different from the GEO and AEO pages?
Same engagement, different question. GEO asks whether assistants recommend your practice. AEO asks whether your content is what the answer is built from. This asks what the model already believes about you and how to correct it when it is wrong. The third is the one most practices have never checked, and it is where the unpleasant surprises live.
Does any of this matter yet for a medical practice?
It matters more for medicine than for most categories, because patients ask assistants health questions constantly and the assistants are visibly cautious in how they answer them — which means they lean hard on verifiable credentials and consistent records. A practice with excellent credentials that are not machine-legible loses to one with equal credentials that are. Whether it justifies a full program today depends on your market; the audit is cheap enough to answer that with evidence rather than opinion.
Find out what the assistants are telling your patients.
The free AI Visibility Grader checks what the major assistants say about your practice and which of your records are producing it. A few minutes, and yours to keep.