What actually changed
A patient used to type a phrase into a search box and receive ten links. They compared, clicked, and assembled a shortlist themselves. Being on page one was worth something because page one had room for ten practices.
Increasingly they ask a question and receive one answer, in prose, naming one or two practices. The competitive unit has changed from a position in a list to presence in a paragraph. There is no second page.
This breaks a familiar promise. An agency cannot get you to number one in a generated answer, because there is no ranking underneath it to be first in. Ask the same question twice with different wording and the answer can change. Ask from a different city and it can change again.
What has not changed
Almost all of the underlying work. Consistent business information, clear service pages, real answers to the questions patients ask, credentials that can be verified, reviews that accumulate. Anyone telling you this is an entirely new discipline requiring an entirely new budget is overselling. What is genuinely new is the measurement, and that is where the category is empty.
1. The vocabulary, briefly
Three labels circulate for roughly the same body of work, and the differences are mostly stylistic.
- AEO, answer engine optimization. Favored by people from a technical search background. Emphasizes structured data, entity consistency and directory presence.
- GEO, generative engine optimization. Favored by people emphasizing content and substantiation. Same destination, different route in.
- AI search. What the market actually says out loud. When we audited healthcare marketing agencies, the ones with a real offering overwhelmingly used this phrase or simply named the systems. Very few used the acronyms at all, which matters if you are searching for a vendor using them.
The label tells you nothing about capability. There is one question that does, and it is in section 8.
2. How an answer about your practice gets built
You do not need the internals. You need a working model good enough to decide what to change, and three ideas carry most of it.
Retrieval: can the system find anything about you?
Many answers are assembled by retrieving current information rather than recalled from training. That makes crawlability a live concern rather than a historical one. Google publishes guidance for site owners on how to be eligible for its AI features, which begins from ordinary indexability rather than from anything exotic: Google Search Central, AI features and your website. OpenAI documents its crawlers separately, and which one matters depends on whether you care about training, search, or live browsing: OpenAI bots documentation.
Entity resolution: can it work out that you are one thing?
This is the step practices most often fail and least often think about. Before a system can say anything useful about your practice it has to resolve your name, location, providers and services into a single coherent entity. A practice that appears under three slightly different names, at two addresses, with a provider who left in 2023 still listed on four directories, is genuinely hard to resolve. A system that cannot resolve you confidently will name someone it can.
Structured data is the most direct way to state this in machine-readable form. Google's introduction to structured data is the starting point, and the relevant vocabulary for a practice includes MedicalBusiness, Physician and the FAQ types.
Corroboration: does anything outside your website agree?
A claim that appears only on your own site is weaker than the same claim appearing on your site, your Google Business Profile, a specialty society listing and a hospital affiliation page. Agreement across independent sources is what moves an entity from known to recommendable. It is also why review depth and directory hygiene do more work here than most practices expect, and why a practice with a beautiful website and a neglected third-party footprint underperforms.
The practical consequence
Most AI visibility work is not clever. It is making one business legible, consistent and corroborated across the handful of places a machine looks. The practices that win are usually not the ones doing something sophisticated. They are the ones who cleaned up contradictions everyone else left in place.
3. Measure it yourself
This is the part the category does not publish, so here it is in full. You can run this without hiring anyone. It takes an afternoon.
Step 1. Write twenty questions a patient would actually ask
In their words, not yours. A patient does not search “rhytidectomy.” Cover four kinds:
- Category questions. “Best plastic surgeon in Bellevue for a facelift.” “Who should I see for veneers in Seattle?”
- Cost questions. “How much do veneers cost in Seattle?” Patients ask these first and most practices never answer them.
- Comparison questions. “Med spa or dermatologist for injectables?” “Invisalign vs braces for teens.” These decide the category before they decide the provider.
- Name verification. “Is [your practice] any good?” This is the query with the most exposure and the one nobody measures. It is what a patient runs after a friend gives them your name.
Step 2. Run them across named systems
ChatGPT, Perplexity, Google AI Overviews and Gemini at minimum. Name them in your record. “We checked AI” is not a measurement.
Step 3. Record more than whether you appeared
For each question and each system, write down:
- Were you named? Yes or no.
- Who else was named? This list is your competitive set as a machine understands it, and it is frequently not the list you would have written.
- Was anything said about you inaccurate or out of date?
- The date, the location context, and whether you were signed in. All three change answers.
Step 4. Interpret honestly
Four outcomes, and each means something different.
- Named consistently. Good. Check the accuracy of what was said before celebrating.
- A competitor named instead. Go and look at what they publish that you do not. In our experience the difference is rarely mysterious.
- A directory or health system returned instead of any practice. Common, and better news than it looks. It usually means nobody in your market has established a clear entity, and an unclaimed position is much cheaper to take than a contested one.
- Refusal to recommend anyone. Also common in health contexts. Rephrase toward the procedure rather than the provider and see whether you appear as a source in the explanation. Being the source is how you get named in the follow-up question.
Step 5. Re-run the identical set in ninety days
Same questions, same systems, new date. Two dated documents is a measurement. One is an impression. This step is the one almost nobody schedules, and it is the only one that makes the previous four falsifiable.
On variance
If the same question gives five different answers, that is normal and it is itself a finding. Wide variation usually means no practice in your market has established a stable presence, so the position is still available. Narrow variation, with the same two or three practices recurring, means somebody has, and you are either one of them or you are working against an incumbent.
4. What to fix, in order
Ordered by yield per unit of effort, which is not the order most agencies sell them in.
First: stop blocking what you want to be found by
Check your robots.txt and any security plugin before anything else. We routinely find practices blocking the crawlers they are simultaneously paying someone to improve their presence in. Decide deliberately which systems you want access from, using the vendor documentation linked in section 2, and make the file reflect the decision.
Second: resolve the entity
One practice name, one address format, one phone number, current providers only, everywhere. Your website, your Google Business Profile, specialty directories, society listings, hospital affiliations, insurance directories. Note that an agency's own Google Business Profile eligibility differs from a practice's: Google requires in-person customer contact and excludes online-only businesses, per its guidelines for representing your business. A clinical practice with a staffed location qualifies comfortably.
Third: answer the questions, including the cost one
Every question from your step 1 list should have a real answer somewhere on your site. The cost question is the one practices avoid and the one patients ask first. If you do not answer it, the system answers it from a third-party source and names whoever did. Published ranges also qualify the inquiry before it consumes coordinator time.
Fourth: make credentials machine-readable
Board certification, training, society membership and hospital privileges expressed as structured data and consistent prose, not as a logo strip a machine cannot parse. In specialties where a system hedges toward caution, which is most clinical ones, this is what tips an answer toward a physician practice.
Fifth: build corroboration
Reviews that accumulate steadily, third-party mentions, and profiles that agree with each other. Slower than the others and it compounds, which is why starting it late is expensive.
What not to do
Do not generate large volumes of thin, near-duplicate pages to cover more query variations. Google's spam policies name scaled content abuse and doorway pages explicitly, and the pattern they describe is exactly what a templated location-times-service page matrix becomes. One page that genuinely answers the question outperforms forty that gesture at it.
5. What is specialty-specific
Generic AI search advice produces generic results because answer behavior is not uniform across specialties. A few examples of how the work actually differs.
- Orthodontics. The person asking is a parent, not the patient. Cost, payment plans and duration dominate the questions. Adult aligner queries and pediatric queries are separate populations, and a single page addressed to both is usually retrieved for neither. See the orthodontics page.
- Plastic surgery. Credential signals do disproportionate work. The name-verification query is the highest-stakes and least-measured one in the specialty. See the plastic surgery page.
- Dermatology. The category question, whether to see a dermatologist or a med spa, is decided before the provider question and is usually won by whoever published more about it. See the dermatology page.
- Cosmetic dentistry. Cost and provider arrive as a single query. Practices presented as general dentists get resolved as general dentists. See the cosmetic dentistry page.
- Vein and vascular. Symptom questions dominate and frequently name nobody at all, which makes the top of the funnel unclaimed rather than contested. See the vein page.
- Medical spas. The one aesthetics category where this ground is already contested by agencies with real offerings. Competitive density is extreme and safety hedging cuts against a non-physician practice. See the med spa page.
6. Healthcare constraints that change the work
Advice written for e-commerce does not survive contact with a medical practice. Four constraints matter.
- Patient imagery needs documented, channel-specific consent. Before-and-after work is among the most retrievable content a practice can publish and among the easiest to publish improperly. Consent for a website gallery is not consent for a paid campaign.
- Tracking is not uniform across your site. A public information page and an authenticated patient portal are different risk contexts, and a single site-wide analytics configuration is the most common mistake we see. See our HIPAA and patient privacy page for how we handle it.
- Claims have to be substantiable. This is a regulatory requirement before it is a visibility strategy, and it happens to align: a claim you can support is a claim a system can corroborate.
- Messaging consent is a separate regime. A HIPAA authorization is not TCPA consent. Relevant the moment anyone proposes automating outreach off the back of new visibility.
7. What nobody can promise you
Stated plainly, because the category does not state it and you will hear the opposite.
- A guaranteed recommendation. Not from us, not from anyone. Answers vary with phrasing, context, location and time, and the systems change without notice.
- Control over model training or retrieval. You influence what is findable, accurate and corroborated. That is the whole lever.
- Even movement across systems. Improvement typically shows up unevenly. Any report averaging four systems into one flattering number is hiding something.
- A credible usage statistic. There is no independent source for how many patients use assistants to choose a provider in healthcare specifically. Every figure in circulation comes from someone selling the service. We would rather say so than repeat one.
- That visibility fixes conversion. If inquiries sit for hours before anyone responds, being named more often makes the leak bigger. Measure response time before buying visibility.
8. How to evaluate a vendor
Do not ask whether they do AI search. Everyone says yes, which makes it a useless question. Ask these instead, and ask them of us too.
- Can you show me in writing where my practice stands today, before I sign anything? A firm doing the work has a document or can produce one. A firm selling the category has a service page.
- What questions did you test, in which systems, on what date? If the answer is not a specific list, there was no measurement.
- Will you re-run the identical set later and show me both? This is the step that makes everything else checkable, and it is the one most commonly missing.
- What did you most recently tell a client to stop paying for? Tells you whether you are talking to an advisor or a vendor.
- Whose name is my domain registered in? Domain, website, Google Business Profile, ad accounts and analytics should all be yours, and should stay yours if the relationship ends.
A 90-day sequence
If you do nothing else, do this in this order.
| When | Do this | Why it is here |
|---|---|---|
| Week 1 | Run the twenty-question baseline. Record everything from section 3. | Everything after this is unfalsifiable without it. |
| Week 1 | Check robots.txt and security plugins for accidental blocking. | Costs an hour. Occasionally explains the entire result. |
| Weeks 2-4 | Reconcile the entity: name, address, phone, providers, services, everywhere. | Highest yield per hour, and almost always neglected. |
| Weeks 3-6 | Answer the top questions publicly, including cost ranges. | The questions patients ask first are the ones practices answer last. |
| Weeks 4-8 | Add credential and organization structured data. Fix contradictions. | Makes the credibility signal machine-readable rather than decorative. |
| Ongoing | Review velocity, and measure inquiry response time. | Corroboration compounds. Response time usually matters more than any campaign. |
| Day 90 | Re-run the identical twenty questions. Compare, and publish the comparison internally. | Two dated documents is a measurement. One is an impression. |
None of this requires an agency. It requires someone to own it and a calendar reminder ninety days out. If that person does not exist in your practice, that is the actual problem to solve first, and it is worth saying out loud before anyone sells you anything.
Questions we get asked
Can an agency guarantee that ChatGPT will recommend my practice?
No. Answers vary with phrasing, location context, account history, the specific product and version, and the date. There is no ranking underneath to be first in. A practice can improve how discoverable, accurate and well corroborated its public information is. Nobody can guarantee inclusion or recommendation, and a vendor promising it is describing a mechanism that does not exist.
What is the difference between AEO and GEO?
In practice, very little. Answer engine optimization and generative engine optimization are used almost interchangeably for work that makes a business discoverable, accurate and well supported enough that AI systems can represent it correctly. Neither term is standardized. Most agencies with a genuine offering say AI search or simply name the systems.
How many patients actually use AI assistants to choose a doctor?
There is no credible independent figure for healthcare specifically, and we will not repeat one. Every percentage in circulation traces back to a company selling the service. The honest argument does not need the number: measuring costs very little, entity and content work compounds, and the cost of being early is low while the cost of being late is not.
Is AI visibility just SEO with a new name?
The underlying work overlaps substantially and any vendor claiming it is entirely new is overselling. The genuine difference is what gets measured. Traditional SEO reports positions and traffic. Answer engine work has to report whether the practice was named in a generated answer, in which systems, and next to whom.
How do I evaluate an agency selling AI search visibility?
Asking whether they do it is useless because everyone says yes. Ask instead for a baseline: the questions they tested, the systems they tested in, the date, and whether your practice was named in each. A firm doing the work has a document or can produce one before you commit. A firm selling the category has a service page.
Should we block AI crawlers from our website?
That is a real decision with a real trade-off, and it should be made deliberately rather than inherited from a plugin default. If you want your content considered for inclusion in AI answers, blocking the relevant crawlers works against that. If you have content you would rather not have used, blocking is the lever. What you should not do is block them by accident and then pay someone to improve your presence in the systems you are excluding.
Would rather we ran it?
We will produce the baseline across four systems, with the question set published so you can check it yourself. You keep the document whether or not you continue with us.