Skip to content

$2,500 fixed  ·  10 working days  ·  30 questions, 5 engines, 150 answers captured

Find out what AI tells your patients when they ask who to see.

Your patients and clients are asking ChatGPT and Google's AI which practice to go to. I measure the answers, show you the transcripts, and tell you what is holding you back. I ran it on my own site first, and I publish the half that went badly.

Book a 30-minute call Or see how I scored my own site

01   The problem

This does not
announce itself.

A patient used to type "hand surgeon near me" and pick from a list. Now they ask an assistant, and the assistant answers in a paragraph that names three practices. If yours is not one of the three, nothing in your reporting changes. Your rankings hold. Your traffic looks normal. The phone is quieter than it was, and there is no line item anywhere that explains it.

That is what makes this different from every other visibility problem I work on. A ranking drop shows up in a dashboard. This does not.

15→8% Searches producing a click once an AI summary appears
45% US adults who used AI for a local recommendation in 2026
6% The same figure a year earlier
0 Line items in your reporting that show this

Click-through figures from Pew Research, 900 US adults across 68,879 real searches. Local recommendation figures from BrightLocal's 2026 survey of 1,002 US adults.

02   The evidence

I ran this on my own site first,
and I will show you the bad half.

Every vendor in this category will show you a case study with a big multiple and no method. I would rather show you my own numbers, including the ones that did not go my way.

In June I audited my own site against an eight domain, hundred point AI readiness scale. It scored 71. After the fixes it scored 91.

The largest single cause was not content. My Cloudflare firewall was quietly returning 403 to OAI-SearchBot, Claude-SearchBot, PerplexityBot, ChatGPT-User and Perplexity-User. Five AI crawlers, blocked, with no error showing anywhere in my analytics. After the fix all five returned 200, re-verified across eleven user agents.

Then the part nobody publishes. I built a fixed panel of 25 buyer questions and I have re-run it every month since. Today I am visible on 2 of them, scoring 3 out of 50. At baseline I scored 4. Two months of work and no slots gained.

Meanwhile on Microsoft's side, my site went from roughly 2 citations a day in June to between 20 and 37 a day by late July, 817 in total, holding between 29 and 43 percent of the citation share on the questions I target.

So one engine started quoting me constantly and another did not move at all. Sitting underneath both numbers is the thing that decides the outcome, and I found it by counting: not one of my citations came from a question shaped like "who should I hire." I win the questions that ask what something costs or how something works. I do not win the question that asks who to pick.

That gap is the whole problem, and it is the reason this is an audit rather than a content retainer.

03   The diagnosis

The model already knows who you are.
It will not bring you up.

Once I had the pattern on my own site I went looking for whether it held generally. It does, and the independent research is more useful than anything a vendor in this category has published.

A study of 112 companies found that models recognized a named business between 94 and 99 percent of the time when asked about it directly, and volunteered it unprompted in only 3.3 percent of ChatGPT answers and 8.3 percent of Perplexity's. The strongest predictor of being discovered was community presence, at a correlation of 0.395. That study's measure of on-page optimization showed no meaningful correlation at all.

Ahrefs looked at 75,000 brands and found branded mentions across the web correlated with AI visibility at roughly 0.664. Backlinks managed 0.218.

And when researchers benchmarked the optimization techniques this industry sells, published at NeurIPS in 2025, only 3 of 54 method and domain combinations produced a significant positive effect. Several made rankings worse.

Put plainly: being findable is a content problem, and being recommended is a corroboration problem. Most of what is sold as AI optimization works on the first one and is priced as though it fixes the second.

I have seen this from three directions now. My own site, and two ecommerce clients who rank organically and appeared in zero buyer-intent AI answers.

04   The deliverable

Thirty questions. Five engines.
Every answer on the record.

Thirty questions your patients really ask

Written for your specialty and your market, and printed in the report word for word. Discovery, condition-led, comparison, and the practical ones about insurance and availability. They stay frozen so the re-measure means something.

Five engines, each one named

ChatGPT, Perplexity, Google AI Overviews, Microsoft Copilot and Gemini. If one of them cannot be reached on your market, that goes in the report instead of the engine quietly disappearing from the count.

Five runs of every question, from your city

Run the same question twice in an afternoon and the sources cited overlap by about a third. One run is an anecdote. Five runs across three days, collected from your metro rather than mine, is a measurement.

A score with a confidence range on it

Every rate carries an interval. Ask any other vendor for the range on their headline number. Most cannot produce one, because most are reporting a single run.

Where the answers come from

Which sites each engine drew on to build its answer, and how much of that is your own website talking about you. This is usually the section that changes the conversation.

Your entity and listing record

Google, Yelp, Apple, Bing and your specialty directories, checked for conflicts. Two listings that disagree about your name or address can make you resolve as more than one business, and that is invisible to any content audit.

It closes with a fix list capped at seven items, in priority order, written so your developer can act on it without calling me. Then a sixty minute readout call. Then a re-measure at 90 days on the identical questions, and I send you the difference whichever direction it goes.

05   The limits

Three things I could charge you for
and will not.

A guarantee that AI will recommend you

Nobody controls what an assistant recommends, and I would be careful with anyone who says otherwise. Run the same question twice in one afternoon and the sources cited overlap by about a third. What I can promise is that every finding in your report is reproducible by you, and that I re-measure in 90 days on the same questions.

An llms.txt file as a lever

Google has said in writing that Search ignores them. When Ahrefs checked 137,210 domains, 28 percent had published a valid file and 97 percent of those received no requests for it at all. Mine scores 96 out of 100 and I still treat it as housekeeping. I will build you one. I will not bill it as a lever.

Schema markup as an AI citation fix

The one controlled test I trust compared 1,885 pages that received new markup against 4,000 matched pages that did not. AI Overview citations went down 4.6 percent. Schema earns its place for rich results and it genuinely helps on Microsoft's side. It is not the thing standing between you and a recommendation.

06   The engagement

$2,500. 10 working days.
The fee comes off the build.

The AI Answer Audit

I measure what ChatGPT, Perplexity, Google AI Overviews, Copilot and Gemini say when a patient or client asks who to see, then hand over the transcripts and a fix list capped at seven items.

$2,500 / project

10 working days

  • →30 questions, 5 engines, 5 runs each
  • →Every answer captured and dated
  • →Scored, with a confidence range on every rate
  • →Entity and listing audit
  • →Fix list capped at seven items
  • →Sixty minute readout call
  • →Re-measure at 90 days, included
See the AI Answer Audit

Implementation

Entity and listing correction, structured data, and the corroboration work that moves recommendation.

quoted from your fix list

Scoped after the audit

  • →Priced from the seven items, not before
  • →Audit fee credited within 30 days
  • →Prohibited Tactics exhibit signed by both sides
Talk it through

Ongoing

Quarterly re-measure and entity monitoring, offered after the 90-day result and not before.

not sold up front

Discussed at the re-measure

  • →I would rather earn the conversation than sell it
  • →Already on a monthly engagement with me? It is part of that, not a second invoice
Ask me later

07   Fit

Who this is for.
Who it isn't.

A straight read on whether this lands. If you fit the left column, book the call. If you fit the right column, I will say so rather than take the fee.

Who this is for

  • → You are a medical practice or a law firm, and a new patient or client is worth four figures or more
  • → You are already visible in ordinary search and cannot work out why the phone has changed
  • → You want to know what is true before you buy a solution
  • → You can give a decision in a room, without a committee

Who this isn't for

  • × You want a guarantee that ChatGPT will name you
  • × You want content volume. That is a different purchase, and on the evidence it is the wrong one to make first
  • × You are a California law firm. State law changed on 1 January 2026 and now carries damages of $5,000 to $100,000 per advertisement, so I do not sell a templated version of this into California firms. I will take it as a custom engagement with your own counsel signing off on every asset
  • × You need me to touch patient records. I never do, and this audit never requires it

Selected build

Greenacre Law.
363 structured data blocks.

I rebuilt Greenacre Law's site in English and Spanish: 224 pages live, all 220 legacy URLs still resolving, the homepage down from 763 KB to 34 KB, and 363 structured data blocks across the English pages. That is the retrievability layer this work depends on, built on a named firm and published with their permission.

Before that, on a real estate law firm I still work with: half the ad spend, twice the monthly qualified signups, and cost per qualified lead down more than 60 percent inside 90 days. I did not isolate which change produced which part of that, and I say so in the write-up.

Read the Greenacre build Read the real estate law case

Ready to look

Find out what
the answer says.

Thirty minutes, no deck. Tell me your specialty and your city and I will run three of the questions live on the call, before you have paid me anything.

Book a call Email me directly

Want the architecture behind the lead side of this? Read the Lead Quality Stack. It is the reference doc behind every practice engagement.

Or email hi@connercrowe.com.

FAQ

Before you ask.

Can you guarantee I will show up in ChatGPT?
No, and be careful with anyone who tells you they can. Nobody controls what an assistant recommends. What this gives you is where you stand today, what is holding you back, and which fixes are worth doing. Placement follows the work.

How is this different from SEO?
SEO decides whether you can be found. This decides whether you get brought up. They run on different signals. A study of 112 companies found models recognized a named business between 94 and 99 percent of the time when asked about it directly, and volunteered it unprompted in only 3.3 percent of ChatGPT answers. The same study found on-page optimization had no meaningful correlation with being discovered.

Do you need access to my patient records or my systems?
No. The audit reads answers any member of the public can see. There is no protected health information in it at any point, which is also why there is no data agreement to sign before it starts.

Will you write the content?
Not as part of this. The audit tells you whether content is even your problem. On most practices I expect it is not, and I would rather find that out before you pay anyone to write.

What happens after 90 days?
I re-run the same 30 questions on the same engines and send you the difference, up or down. That is included in the fee, and I publish the delta whichever way it goes.

My marketing company says they already do this.
They might. Ask them for three things: their question list, how many times they ran each question, and the confidence range on their headline number. If all three exist, you are in good hands and you do not need me.

What does it cost?
$2,500 as a fixed fee, over 10 working days. If you decide to have me implement the fixes within 30 days, the fee comes off that quote. There is no ongoing commitment attached to it and I do not quote one until the 90-day re-measure is in.