Will Patients Find Your Practice Through AI? What the Evidence Shows
Google states in writing that AI search needs no special optimisation. Here is the measured evidence on citations, schema and llms.txt, including the debunks.
How patients find a practice through AI search
Table of Contents13 sections
If you run a practice, you've probably had at least one email this month selling AI search visibility, maybe with a free AI visibility score attached. A lot of money is being made right now selling that to Australian practice owners.
The short version: there's no separate technical layer for AI search, and Google has said so in writing three times. For AI Overviews, organic rank is the strongest predictor: 76.1% of cited pages rank in the top ten. Schema doesn't get you cited, nothing reads llms.txt, and the one change with support is citing sources and numbers in visible text.
Before you buy any of it, read one sentence from Google's own documentation, last updated December 2025. The bold is Google's.
There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.
And from Google's AI optimisation guide, updated 10 July 2026:
From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.
So has nothing changed for you? A lot has, and some of it hits your practice in a specific and slightly strange way that I'll get to below, once we've cleared away what's being oversold. The gap between what the evidence supports and what's being sold is wide, and you should look at it closely before you sign anything.
Key Takeaways
- Google says there are "no additional requirements" to appear in AI Overviews, and no special optimisations. That's Google's wording, and the bold is Google's too.
- Rank is the lever for AI Overviews. 76.1% of cited pages sit in the top ten; position one is cited in 43% of queries, position seven in 5%.
- Rank barely carries over to AI assistants. Their citations overlap with Google's top ten about 12% of the time, and 80% of assistant citations don't rank at all.
- Being cited is worth about 120% more organic clicks than not being cited on the same page, but still around 38% fewer than if no AI Overview had appeared.
- Your own website is the largest single source category in every local study that measured it. It's the least fashionable finding here, and the best evidenced.
- 13.6 million Australians used an AI tool in an average four weeks (Roy Morgan, 14,646 respondents). How many used one to choose a dentist has never been measured.
Where AI Search Stands in Australia
AI Overviews reached Australia on 28 October 2024, when Google rolled them out to more than 100 more countries. AI Mode arrived in the October 2025 expansion, and Australia is in Google's current supported countries table for both AI Mode and Gemini in AI Mode.
Both work by what Google calls query fan-out, which sounds technical but is simple once you see it:
dividing your question into subtopics and searching for each one simultaneously across multiple data sources
Keep that idea in your head. It explains most of this article, and most of what ordinary local search work looks like now.
For usage, the only credible Australian primary source I could find is Roy Morgan, published 2 June 2026 from its Single Source survey of 14,646 Australians aged 14 and over. It found 13.6 million Australians, 58% of that population, used an AI tool in an average four weeks. ChatGPT accounted for 10.5 million.
That measures people using an AI tool. It doesn't measure your patients using AI to choose a dentist, and nobody has measured that in Australia.
Any figure about Australian AI search that sounds more specific than this is either American data with an Australian label on it, or a marketing agency's own small sample. I went looking, and I found several of both.
AI Search Exposure Is Real, and Not Where You Would Guess
Ahrefs analysed 146 million desktop search results pages in September 2025 and found AI Overviews on 20.5% of them, which is the kind of number you'll see quoted in every pitch deck you get sent this year. WebFX ran the same kind of count inside healthcare, which is closer to your world, across 130,070 US healthcare queries in July 2025.
| Search type | Shows an AI Overview | Study |
|---|---|---|
| All searches | 20.5% | Ahrefs, 146M results pages |
| Medical topics | 44.1% | Ahrefs |
| Health topics | 43.0% | Ahrefs |
| Non-local searches | 22.8% | Ahrefs |
| Local searches | 7.9% | Ahrefs |
| Health search, no modifier | 54.2% | WebFX, 130,070 US queries |
| Health search plus a location | 35.8% | WebFX |
| Health search plus a brand | 34.0% | WebFX |
| Health search plus both | 25.7% | WebFX |
Look at the bottom half first. Medical and health came second highest of Ahrefs' 27 categories. If you write about health, you're in one of the most AI-exposed corners of the internet.
Now the part that changes the conclusion: local searches triggered an AI Overview only 7.9% of the time, and every modifier WebFX added pulled the rate down. Add a location to a health search and an AI Overview gets less likely, and add a brand on top and it gets less likely again.
So "why does my tooth hurt at night" goes through AI a lot, while "dentist Coburg" goes through it much less.
The searches that book your appointments are the ones AI has touched least. That surprised me. It might surprise you too.
One contradiction I won't hide from you. Seer Interactive reports 76.9% of informational "near me" queries carrying an AI Overview, ten times Ahrefs' local figure.
They're clearly measuring different things, and nobody has reconciled them. What holds across every source is the direction: add a location or a brand name, and an AI Overview gets less likely.
What the AI Search Click Data Says, Including the Bit Nobody Reported
You've probably already heard some version of "AI Overviews cut clicks by 58%". I had too. The real picture is messier, and more interesting, and you should see all four studies side by side before you believe any one of them:
| Study | What it measured | Finding |
|---|---|---|
| Pew Research Center, July 2025 | 900 US adults, 68,879 searches | Clicks on 8% of visits with an AI summary, 15% without |
| Ahrefs, February 2026 | 300,000 keywords, position one | 58% fewer clicks, after a 49% unexplained drop in its control group |
| Seer Interactive, April 2026 | 5.47 million queries | Click-through fell to 1.31%, then recovered to 2.36% |
| Google, August 2025 | No data published | Total click volume "relatively stable" |
Pew Research Center: 8% Versus 15%
Pew is the only non-commercial measurement in this field, and the only one with nothing to sell you, so I'd start there. In a study published 22 July 2025, it tracked 900 US adults through a metered browsing panel across 68,879 searches in March 2025.
Users clicked a traditional result on 8% of visits where an AI summary appeared, against 15% where one did not. Only 1% clicked a link inside the summary itself.
Ahrefs: The 58% Figure, and Its Control Group
Ahrefs produced the 58% figure in February 2026, from 300,000 keywords using aggregated Search Console data, desktop, position one only.
Almost every retelling you'll read leaves out one detail, which is that the 58% is a net figure:
- The raw drop on AI Overview keywords was 78%.
- Their control group (informational keywords that never showed an AI Overview) fell about 49% on its own over the same window.
- The 58% is what's left after taking out a control decline that Ahrefs can't explain.
Nobody has explained that 49%.
Seer Interactive: The Recovery Nobody Covered
Seer Interactive published the sequel in April 2026, and almost nobody covered it. Across 5.47 million queries and 2.43 billion impressions, organic click-through on AI Overview queries fell to a floor of 1.31% in December 2025, then climbed back to 2.36% by February 2026.
Seer won't name a cause, and it warns you against forecasting from two months.
Still, look at the dates. Ahrefs measured December 2025, the exact bottom, and hasn't re-run the study. Notice something?
Google Says the Opposite
Liz Reid, Google's head of Search, wrote in August 2025:
total organic click volume from Google Search to websites has been relatively stable year-over-year
She published no data with it. No figures, no method, no time window.
So which one do you believe? Nobody offers a way to square the two, so I'll give you mine.
The industry studies measure click-through rate on the affected queries. Google measures total click volume across all of Search.
More searches at a lower rate per search can still add up to stable total clicks. So both statements can be true at once, and part of the fight is about which number you divide by.
The most useful figure in all of this is Seer's citation maths. Being cited in an AI Overview gets you about 120% more organic clicks than not being cited on the same results page, but still around 38% fewer than if no AI Overview had appeared at all. So a citation is a win for you, inside a smaller pool.
What Predicts Being Cited in AI Search
There are two findings here, and they point different ways depending on which AI you mean, so keep them apart in your head.
For AI Overviews, Ranking Is the Whole Game
Ahrefs looked at 1.9 million citations from a million AI Overviews in July 2025. 76.1% of cited pages ranked in Google's top 10, and the median organic rank of a cited URL was three.
A separate study by Kurt Fischman, posted to SSRN in April 2026 and the only one I found that controls for both domain authority and organic rank, found rank beat everything. A page at position one was cited in 43% of queries. A page at position seven, in 5%.
Ahrefs' own author adds the caution you need, and I'd keep it in mind: ranking first means "that chance is a coin flip at best."
So if a page ranks in the top ten for a question, it has a real chance of being cited in the AI Overview for that question, and in Fischman's study nothing else he tested came close to rank as a predictor of who gets cited.
| What predicts a citation | AI Overviews | AI assistants |
|---|---|---|
| Cited pages in Google's top ten | 76.1% | About 12% on average |
| Best and worst performers | Position one cited in 43% of queries, position seven in 5% | Perplexity 28.6%; ChatGPT, Gemini and Copilot around 8% |
| Citations that don't rank at all | Not reported | 80% |
For AI Assistants, Ranking Barely Transfers
Ahrefs ran 15,000 long-tail queries (the long, specific kind your patients type) in August 2025 and found AI assistant citations overlapped with Google's top 10 only 12% of the time.
Perplexity was the exception at 28.6%. ChatGPT, Gemini and Copilot all sat around 8%. Eighty percent of assistant citations didn't rank anywhere in Google for the original query.
Query fan-out explains it again, and this is where you'll see why I asked you to remember it. An assistant breaks your question into a dozen smaller questions and merges the results, so a page ranking around sixth for several related queries can beat a page ranking first for one.
What that means for you: the work that gets you into AI Overviews is ordinary search work. The work that gets you into ChatGPT is being solidly decent across a whole topic, instead of dominant on one phrase.
(You can see what assistants actually name by asking them; I asked ChatGPT, Perplexity and Gemini 12 questions about Bali clinics and logged every answer.)
Schema Does Not Do What You Have Been Told About AI Search
This is the most oversold item in the whole category, so I'll give you all the evidence, and you can judge it yourself.
Here's Google, under a heading that literally reads "Mythbusting generative AI search: what you don't need to do":
Overfocusing on structured data: Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add.
The measured studies agree with Google, again and again, and I couldn't find one that doesn't:
| Study | What it tested | Result |
|---|---|---|
| Ahrefs, May 2026 | 1,885 pages that added JSON-LD, against roughly 4,000 matched controls | AI Overview citations down 4.6%; AI Mode up 2.4% and ChatGPT up 2.2%, both no different from zero |
| Fischman, April 2026 | Schema presence, controlling for rank and authority | Null, p = 0.296; schema on 43.1% of cited pages and 44.8% of uncited ones |
| SE Ranking, December 2025 | 2.3 million pages | FAQ schema had no impact on AI Mode citations, about three times weaker than real FAQ sections in visible content |
| searchVIU, December 2025 | Test values planted in eight conditions across five systems | "JSON-LD Schema Markup is NOT extracted by ANY system during direct fetch" |
Ahrefs' study is the closest thing to a controlled before-and-after that you'll find. And in searchVIU's test, every value the systems found was in visible content.
No study has found that schema causes citations once rank is controlled for. Ahrefs names the confounder itself: schema tends to live on better-maintained sites, and those sites publish better content and earn more links.
Two claims you'll hear in the Australian market didn't check out at all when I traced them. BrightEdge gets credit for a study showing a 44% citation lift from schema. As far as I can tell, that study doesn't exist, and BrightEdge's real 44% figure is about something else entirely.
And the Princeton and Georgia Tech generative engine optimisation paper is widely cited as showing that structured data improves AI visibility by 30 to 40%.
I read it. It never tested schema. Search the full text and you get no hits for "schema", "structured data", "JSON-LD" or "markup".
What that paper (published at KDD 2024) did find is more useful anyway. The 30 to 40% lifts came from:
- citing sources
- adding quotations
- adding statistics
- improving fluency
- writing with an authoritative voice
Keyword stuffing did worst of everything tested. A 2026 follow-up presented at SIGIR, running 252,000 controlled trials across six models, found that visible price and recency information act as citation gatekeepers while formatting-only changes do nothing.
In other words, you get cited by writing pages with specific, current, sourced facts, prices included. Nobody can sell you that as a technical trick or a monthly add-on, which may be why nobody pitches it.
Structured data is still worth having for ordinary search, and Google says so in the same breath. Just don't buy it as AI visibility. And Google retired FAQ rich results on 7 May 2026, so anyone selling you FAQ schema for rich snippets is selling something Google has switched off.
The One Genuinely New AI Search Thing, Aimed at You
Google's AI Mode help documentation describes what happens when a business hasn't published enough information:
AI Mode simplifies your search by gathering information from across the web, completing forms, or calling businesses directly... When businesses have limited information available online, Google can make calls on your behalf.
There's even a help page for business owners like you about managing these calls.
For a dental practice, this is the most concrete thing in the whole article. If your fees, appointment availability, emergency policy and Saturday hours aren't published anywhere a machine can read them, the failure is no longer just a patient who can't find the answer.
It's your front desk taking an automated call about it in the middle of a busy morning.
The fix is boring. Publish it in plain visible text, on your website and on your Google Business Profile. Google's own best-practice list for AI features names keeping your Business Profile current.
Related, and worth watching (not acting on yet): Sterling Sky reported in January 2026 that an AI-generated local pack is starting to replace the traditional three-pack on around 8% of keywords. It often shows one or two businesses instead of three, surfaces 32% fewer unique businesses, and in many versions drops the click-to-call button while paid ads keep theirs.
llms.txt Is a File Nobody Reads
Someone will offer you this, so I'll cover it. llms.txt is a markdown file at your site root, proposed in September 2024, and its own specification makes no claim about AI search visibility. The SEO industry added that story afterwards.
Here's Google, in writing:
You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them... Doing so will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them.
John Mueller of Google compared it to the keywords meta tag, and added that "none of the AI services have said they're using LLMs.TXT (and you can tell when you look at your server logs that they don't even check for it)".
Ahrefs tested that in June 2026 across all 137,210 domains in its analytics product. About 28% published a valid llms.txt, and ninety-seven percent of those files received zero requests in May 2026. No AI bot ever requested an llms.txt URL that returned a 404 either. So there's no getting-on-the-radar effect.
If your CMS makes one automatically, it costs you nothing and you can leave it. If someone is charging you for it as an AI visibility service, you're paying for a file the companies it's named after have never agreed to read. The full server-log breakdown, including who does fetch these files, is in the llms.txt piece.
Two AI Search Crawler Facts Usually Stated Backwards
People get these the wrong way round all the time, including the people setting up your site. Do you know which ones yours blocks?
| Crawler | What it governs | What blocking it does |
|---|---|---|
| GPTBot | Training for OpenAI models | Keeps you out of training; ChatGPT can still cite you |
| OAI-SearchBot | Citation in ChatGPT answers | Removes you from ChatGPT answers |
| Google-Extended | Training and grounding for Gemini apps | Nothing for AI Overviews |
| Googlebot | Google Search, AI Overviews and AI Mode included | Removes you from Google Search entirely |
Blocking GPTBot doesn't stop ChatGPT citing you. OpenAI documents that its crawlers are independent: "a webmaster can allow OAI-SearchBot in order to appear in search results while disallowing GPTBot to indicate that crawled content should not be used for training." GPTBot governs training and OAI-SearchBot governs citation. Block the wrong one and you vanish from ChatGPT answers.
Google-Extended doesn't control AI Overviews. Google says it governs training and grounding for Gemini apps, and that it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal".
AI Overviews and AI Mode run on ordinary Googlebot indexing. The only robots.txt lever over them is blocking Googlebot, and that takes you out of Google Search entirely.
If your site sits behind Cloudflare, check its managed robots.txt block. It can put rules that contradict your own at the top of your file.
What Australian Patients Are Doing in AI Search
Nobody knows. So read the American data below with that in mind.
| US finding | Figure | Source |
|---|---|---|
| Used AI tools for local business recommendations | 45% (ChatGPT 31%, Google AI Mode 23%) | BrightLocal, March 2026, 1,002 adults |
| Most trusted part of a results page for healthcare queries | AI Overviews 37%, organic links 20%, map pack 13% | rater8, 2026, about 1,000 patients |
| Came across wrong provider information from an AI tool | 66% | rater8 |
| Still use traditional search engines regularly | More than 95% | SparkToro clickstream |
| AI Mode share of all Google searches, January to April 2026 | 0.34% | Similarweb |
You'll see BrightLocal's jump from 6% the year before quoted a lot. I'd be suspicious of it, because the question wording seems to have changed between waves.
The rater8 figures are the worrying ones for a practice: in that same sample, 60% still trusted the AI summary without checking it. It's a healthcare vendor's survey of US patients, with no Australian sample.
Then the counterweight, which rarely makes it into the same article. SparkToro found no real decline in traditional search in two and a half years, and Google still runs roughly 210 times ChatGPT's query volume.
Both are true. People trust AI answers a lot where they appear. And they still appear on only a small share of the searches that bring patients to your practice.
One more, because it's the only dental-specific one I found. BrightLocal ran 20 local searches across four AI engines in July 2025, and noted:
In our dentistry searches, ChatGPT exclusively sourced information from ten different dental directories.
That's a sample of twenty, from a company that sells citation building, so hold it loosely. It also contradicts BrightLocal's own December 2024 study, which found Yelp, Facebook and Google Maps missing from ChatGPT's local sources, with business websites at 58%.
The mix of directories shifts over months. Anything you build on whichever directory is in favour right now won't last long.
What to Do About AI Search, Which Is Boring
Across every local study that measured it, the best-supported finding is that the business's own website is the largest single source category. Least fashionable, best evidenced.
| Supported by evidence | Not supported |
|---|---|
| Citing sources, quotations and statistics | Schema markup as an AI tactic |
| Fluency and an authoritative voice | The BrightEdge 44% claim |
| Visible price and recency information | The Princeton 30 to 40% schema claim |
| Ranking normally | llms.txt |
| Real FAQ sections in visible content | FAQ schema, keyword stuffing and formatting-only changes |
So, in order:
- Publish your real information in plain visible text: your fees or fee ranges, what each treatment involves, your hours with holidays included, your emergency policy, what a first visit looks like and which health funds you take.
- Write it as prose a person would read, because the systems pulling it out read the visible page.
- Cite your sources, and put specific numbers in wherever you have them. That's the one change with published experimental support.
- Rank normally. For AI Overviews, rank is your strongest predictor, and you have no separate lever.
- Keep your Business Profile current, which Google names in its own AI guidance, and keep reviews coming in within what AHPRA allows for Google reviews.
- Publish prices and dates. The SIGIR work found both act as gatekeepers for citation.
Then stop. There's no extra technical layer, and Google has now said so in writing three times.
Got a proposal with a monthly line item for AI visibility? Ask which of the above it is, and what it costs. I cover what search work should cost in SEO pricing in Australia, we check the AI-citation layer in a fixed-price marketing audit, and the monthly version is tracking what the assistants say about you.
Common Questions
Do we need to do anything special to appear in AI Overviews?
Google says no, in bold, in its own documentation. A page needs to be indexed and eligible to show with a snippet. There are no extra technical requirements.
Does schema markup get us cited by AI?
The controlled studies say no. Ahrefs found a small decline for AI Overviews, and no significant effect for AI Mode or ChatGPT, on pages like yours that added it. Keep structured data for ordinary search, and don't treat it as an AI tactic.
Is llms.txt worth adding?
Google says it ignores the file, and of the sites that publish one, 97% had it fetched by nothing at all within a month. If yours came free, leave it, but I wouldn't pay for it.
Are AI Overviews killing our traffic?
Exposure is high on informational health questions. On location-based searches, which are the ones that bring you patients, AI Overviews show up far less. Measured click effects vary widely by study, the most recent data shows a partial rebound in early 2026, and Google disputes the whole framing without publishing data.
Should we block AI crawlers?
First work out which crawler you'd be blocking. GPTBot governs training, OAI-SearchBot governs whether you appear in ChatGPT answers, and Google-Extended has no effect on AI Overviews at all.
How do we get ChatGPT to recommend our practice?
Nobody can promise this, and anyone who does is guessing. What the research supports is ranking consistently across a topic (more than dominating one phrase), publishing specific and current facts in visible text, and being mentioned on other sites. That's a description of ordinary good marketing.
Are Australian patients really using AI to pick a dentist?
Nobody has measured it in Australia, and I looked. Roy Morgan found that 58% of Australians use an AI tool in an average month, but that's tool use, and choosing a provider is a different question. Everything more specific is US data or an agency's own small sample.
Why is Google phoning our practice to ask about our prices and availability?
Because that information isn't published anywhere a machine can read it. Google's AI Mode documentation says that when businesses have limited information online, Google can call on a searcher's behalf. The fix is to publish your fees, availability, emergency policy and Saturday hours in plain visible text on your site and your Google Business Profile.
Should we be paying an agency for AI visibility or GEO services?
Ask what exactly they'll do that ordinary search work doesn't already cover, and ask them to show you the evidence. Google's own documentation says there are no additional requirements and no special optimisations for appearing in AI Overviews or AI Mode. If the proposal is llms.txt, AI-specific schema and "AI-optimised content", the published evidence is against the work being sold.
Does adding FAQ or dental schema get our practice quoted in AI answers?
The evidence says no. Ahrefs tracked 1,885 pages adding schema and found no citation effect you could tell apart from zero. Structured data does useful jobs for search features. Nobody has shown it gets you cited in a generative answer.
Is llms.txt worth adding to our website?
Only if your CMS makes it for free. Ahrefs tested every domain in its analytics product in June 2026: about 28% published a valid llms.txt, and 97% of those files received zero requests in the month measured. Google says it ignores the file, so nobody should be charging you for one.
Will blocking AI crawlers protect our practice, and which one do we block?
Blocking GPTBot keeps your content out of training, but it doesn't remove you from ChatGPT answers, because OAI-SearchBot governs citation. Google-Extended covers Gemini apps and has no effect on AI Overviews, which run on ordinary Googlebot. Blocking Googlebot is the only robots.txt lever over them, and it takes you out of Google Search entirely.
Are AI Overviews going to stop patients clicking through to our website?
For some of your searches, yes, and the click studies disagree on how much. What matters for a dental practice is where AI Overviews actually show up: far more on medical and health topics than on searches with a location in them, and most patients search with a location.
How do we get ChatGPT or Gemini to recommend our practice?
There's no submission process and no setting. The citation research says generative answers lean heavily on pages that already rank well, so the work is the ordinary work: be findable, be accurate, and publish the specific things a patient asks about. Anyone promising you a place in an AI answer is promising something they can't control.
Sources
- Google: AI features and your website, updated 10 December 2025
- Google: Optimizing your website for generative AI features, updated 10 July 2026
- Google: AI Mode help and managing calls from Google
- Pew Research Center, 22 July 2025, 900 US adults, 68,879 searches
- Ahrefs: AI Overviews reduce clicks, update, February 2026, 300,000 keywords
- Seer Interactive: AIO impact on CTR, 2026 update, 24 April 2026, 5.47M queries
- Ahrefs: What triggers AI Overviews, November 2025, 146M SERPs
- Ahrefs: 76% of AI Overview citations pull from the top 10, July 2025
- Ahrefs: AI search overlap, August 2025, 15,000 queries
- Ahrefs: We tracked 1,885 pages adding schema, May 2026
- Ahrefs: llms.txt study, June 2026, 137,210 domains
- Aggarwal et al., "GEO: Generative Engine Optimization", KDD 2024, arXiv:2311.09735
- "What Gets Cited: Competitive GEO in AI Answer Engines", SIGIR 2026
- Roy Morgan: AI tools usage, 2 June 2026, 14,646 Australians
- BrightLocal: Consumers asking AI for business recommendations, March 2026, 1,002 US adults
- rater8: 2026 Patient Choice Report, June 2026, ~1,000 US patients
- OpenAI: Bots documentation
- Google: Common crawlers, updated 14 July 2026
