TL;DR: Chiropractic and physical therapy are the two verticals where the document that gets you paid is the same document that gets you audited. US chiropractic is a $24.1 billion industry of 66,137 businesses growing 2.8% in 2026 (IBISWorld); US physical therapy services run $54.6 billion in 2026 heading to $76.6 billion by 2033, spread across roughly 37,000 outpatient clinics that are ~75% independent. Both are cheap to advertise into and brutal to retain out of: PT is the third-cheapest lead in healthcare at $32.79 with a 15.35% conversion rate and $4.95 CPC (LocaliQ 2026, 16 specialties) — and yet chiropractic care-plan completion runs 30–50%, with the steepest dropout at visits four through six, and PT patients in a 2024 JOSPT Open study self-discharged at 55%. The reason is not demand. It is that Medicare pays only for active/corrective treatment, requires the AT modifier on CPT 98940/98941/98942, and states plainly that once maximum therapeutic benefit is reached, maintenance therapy is not medically necessary. A filtered model writes “patient reports feeling somewhat better” — which is exactly the sentence that fails that test.
venice-uncensored-1-2 at $0.20 in / $0.90 out per million tokens costs a 900-patient, two-DC clinic about $0.35 a month for 790 documents, or $0.00036 per daily note.
What this covers
- The cheapest lead in healthcare is the leakiest
- The dropout is the business
- The line Medicare draws, in its own words
- Direct access is three questions, not one
- The five documents a filtered model ruins
- What you are allowed to say out loud
- The number you should stop printing
- PHI is the prompt
- The cost math at live Venice rates
- A seven-step action plan
The cheapest lead in healthcare is the leakiest
LocaliQ's 2026 search advertising benchmarks break healthcare into 16 specialties. Overall healthcare search sits at a $5.64 average cost per click, a 6.07% click-through rate and a $66.02 cost per lead — down about 6% year over year, which is unusual in a report where costs have climbed for a decade.
Physical therapy is the cheapest lead in the set after dermatology and ophthalmology:
| Healthcare specialty (LocaliQ 2026) | CPC | CTR | CVR | CPL |
|---|---|---|---|---|
| Physical Therapy | $4.95 | 6.61% | 15.35% | $32.79 |
| Dermatology | $4.90 | — | 25.33% | $18.54 |
| Ophthalmology | — | — | 18.29% | $30.88 |
| Healthcare average | $5.64 | 6.07% | — | $66.02 |
| Plastic & cosmetic surgery | — | — | — | $102.51 |
| Addiction recovery | $3.46 | — | — | $120.30 |
| Mental health | $4.22 | — | — | $141.17 |
Two things are true at once. PT's cost per click fell roughly 10% year over year — one of only three specialties to drop that much — and its conversion rate of 15.35% is the third highest in healthcare. That combination is why a PT clinic can generate a lead for $32.79 while a mental health practice pays $141.17 for the same click-shaped thing.
Independent benchmarks put PT non-branded CPL in a $28–$45 band, which corroborates the LocaliQ figure, and then add the number nobody quotes in a pitch deck: the qualified cost per lead — someone who has been treatment-planned and can actually pay — runs $175–$250 or higher. Same channel, same campaign. The gap between $33 and $200 is the entire business.
Chiropractic has the mirror problem in a different shape. IBISWorld sizes the industry at $24.1 billion in 2026, up 2.8% on the year and growing at a 2.3% five-year compound rate, across 66,137 businesses — and no single company holds more than 5% share. A fragmented market with sub-inflation growth means growth comes from share, and share in a local service business comes from three places: the phone being answered, the plan being finished, and the record supporting the claim.
The dropout is the business
Here is the number that should reorganize a clinic's entire marketing budget. Care-plan completion in chiropractic — the share of new patients who finish the recommended visit count — is generally reported between 30% and 50%, with the steepest dropout in the first four to six visits. Six-month return rates land 40–65%. Annual-active rates, the most flattering definition, sit 50–75%. The definitions are not interchangeable, and practices that track only annual-active overstate retention badly because it hides the early cliff.
Physical therapy shows the same shape from a different dataset. A 2024 JOSPT Open study of physical therapy patients with chronic pain found 73% average session attendance and a 55% self-discharge rate — and higher no-show rates were associated with higher odds of stopping before the plan was complete.
The utilization baseline, from the peer-reviewed Medical Expenditure Panel Survey analysis, is 8.3 visits per year at roughly $87 per visit, or about $721 per patient per year. That is the defensible floor. Patient visit average across US practices runs about 12 to 16; a well-run office targets 30 or more. Every point of PVA is visits from a patient you already paid to acquire.
Put the two numbers together and the arithmetic gets uncomfortable. A clinic with 900 active patients at 8.3 visits a year is doing roughly 623 visits a month and about $650,000 of annual revenue at the MEPS rate. A 5% improvement in retention is worth more than any single campaign, and the research on that is not subtle: Bain's classic finding is that a 5% reduction in defection produces 25% to 95% more profit.
Where the leak is physically located:
- No-shows. Chiropractic clinics commonly run 5–20%, and each missed slot costs roughly $200. Automated text reminders cut no-shows by about 38%. In a 2024 MGMA poll, 37% of medical groups said no-shows were getting worse year over year — usually the ones still making manual reminder calls.
- The perception cliff. A 2026 survey of 455 lapsed patients found 58% quit for perception-based reasons, not cost or access. They did not feel the care was doing anything they could see.
- The unbooked checkout. The patient who leaves without the next appointment is not lost at visit twelve. They are lost at visit five, standing at the front desk.
Which brings this to the uncomfortable conclusion: the document set is the retention system. Not the ad, not the offer, not the review count. The re-exam that shows a number moving is what gets the patient to visit twelve.
The line Medicare draws, in its own words
Chiropractic is the only Medicare benefit written as a single procedure. CMS billing article A56273 is explicit that coverage “is limited to manual manipulation for the treatment of subluxation,” billed as CPT 98940 (1–2 spinal regions), 98941 (3–4) or 98942 (5). Everything else a chiropractor might reasonably do — x-rays, exams, physiotherapy, traction, supplies, orthotics, nutritional counseling — is statutorily excluded when performed by a DC, regardless of medical necessity.
Then comes the sentence that decides whether a note is worth writing:
“Once the maximum therapeutic benefit has been achieved for a given condition, ongoing maintenance therapy is not considered to be medically necessary under the Medicare program.” — CMS, Billing and Coding: Chiropractic Services (A56273)
And the mechanical test that implements it:
“Claims submitted for Chiropractic Manipulative Treatment (CMT) CPT codes 98940, 98941, or 98942 … not containing an AT modifier will be considered not medically necessary.” — CMS, A56273
The AT modifier is a legal assertion, not a checkbox. It says: this is active or corrective care, and I expect measurable improvement. CMS's own documentation checklist says it directly — make sure the record shows the service is corrective treatment, not maintenance — and warns that the mere presence of the AT modifier does not make a service reasonable and necessary. Contractors can and do deny after review when the record does not carry the assertion the modifier made.
The documentation standard is specific enough to be checkable:
| Requirement | What Medicare accepts |
|---|---|
| Subluxation, imaging route | X-ray within 12 months before or 3 months after treatment start; CT or MRI showing subluxation; your review noting the level (older films acceptable for chronic conditions like scoliosis if the record shows it predates 12 months and is reasonably permanent) |
| Subluxation, exam route | At least 2 of 4 P.A.R.T. elements — Pain, Asymmetry/misalignment, Range of motion abnormality, Tissue tone change — and one of them must be asymmetry or range of motion |
| Diagnosis | Precise level of subluxation as the primary diagnosis on the claim; the neuromusculoskeletal condition as secondary |
| Acute care horizon | Care may last up to 3 months, with visits decreasing over time |
| Frequency | Typically one manipulation per patient per day is reimbursed |
| Expected denial | Signed ABN plus the GA modifier; an ABN signed after the service is invalid and the practice absorbs the cost |
The audit exposure is not abstract. Common denial reasons in the CMS checklist are missing or inadequate treatment plans, an unclear chief complaint, and failure to document all treated spinal regions. Three of those four are writing failures.
Direct access is three questions, not one
Physical therapy has the mirror-image problem: the law says yes and the paperwork still says wait. As of July 1, 2025, APTA reports that patients in all 50 states, DC and the US Virgin Islands have either provisional or unrestricted direct access — 21 unrestricted, and 29 plus DC and USVI provisional, meaning a time limit, a visit limit, or a referral requirement for a specific intervention such as needle EMG or spinal manipulation.
The state board's reference guide counts the restrictions more granularly: 27 states impose a visit or time limit, 8 limit access to certain populations or circumstances, 6 require a referral for a specific intervention, 9 require certain qualifications of the therapist, and 13 carve out no-referral-needed fitness and wellness services.
So a front desk has three stacked questions, and only the first one is about the practice act:
- What does the state allow, for how long, and with what notice? California: 45 calendar days or 12 visits, with a physician or podiatrist signing the plan (Bus. & Prof. Code §2620.1). Texas: 10 consecutive business days, or 15 with a doctorate plus residency, fellowship or board-approved certification (Occ. Code §453.301). Florida: 30 days (Fla. Stat. §486.021(11)(a)). New York: 30 days and 10 visits (Educ. Law §6731(d)).
- Will this patient's plan pay without a referral? APTA's own guidance is that most commercial plans do not explicitly require a referral in payment policy, but providers should review each contract to confirm direct-access services are payable.
- If this is Medicare, who certifies the plan? Medicare does not require a referral, but it does require certification. Under 42 CFR 424.24, the plan must be certified by a physician or a nurse practitioner, clinical nurse specialist or physician assistant who knows the case, recertified at least every 90 days, and if the therapist established the plan, the written order must be in the record or the plan delivered to that practitioner within 30 days of the initial evaluation.
That third question is the one that generates unbillable visits. A patient walks in under unrestricted direct access on Monday, gets treated, and the plan sits unsigned for five weeks. The clinical encounter happened. The documentation obligation did not.
The five documents a filtered model ruins
This is where the censorship conversation stops being philosophical. A safety-tuned model asked to write a clinical note produces the hedge, because the hedge is what it was trained to prefer. The hedge is also what Medicare denies.
| Document | What a filtered model writes | Why it fails | What the record needs |
|---|---|---|---|
| Daily treatment note | “Patient reports feeling somewhat better today.” | No objective measure → AT modifier unsupported → processed as maintenance and denied | Measured change, spinal level treated, the P.A.R.T. element observed |
| Plan of care | “Continue care as needed and reassess.” | No type, amount, duration or frequency → not a plan → certification and recertification fail | Diagnosis, long-term goals, type/amount/duration/frequency of services |
| Re-exam report | “Patient continues to improve.” | No baseline, no instrument, no interval → nothing a reviewer can compare | Baseline vs. current, same measure, with the visit number |
| Recall / reactivation message | “Your health matters — come back whenever you're ready.” | No finding, no date, no reason to act now | The specific measurement that changed and a specific appointment window |
| Ad headline | “Realign your spine and transform your life.” | Unsubstantiated health outcome → Misrepresentation and state board advertising exposure | What you do, for whom, and what is measurable about it |
Every row is the same failure in a different costume: the model reached for the sentence that cannot be held against it, and the sentence that cannot be held against it is the sentence that cannot be used.
The correct output is not the aggressive version either. A model that will write anything writes “guaranteed relief in 6 visits,” and that is a different liability with the same invoice. What this vertical needs is a model that writes the specific, defensible sentence — the one with a number, an instrument and an interval in it — and refuses to inflate past what the record supports. That is a narrower ask than “uncensored” sounds, and it is the only version that survives an audit.
What a hedged default costs you
- Denied claims you cannot appeal, because the record never asserted improvement
- Re-exams that show the patient nothing, feeding the 58% perception-based dropout
- Recall messages with no reason to reply
- Notes that take 12 minutes to write and say nothing
What an unfiltered model costs you if you let it
- Cure and guarantee claims that trip Misrepresentation and state boards
- Invented measurements in a record that will be reviewed
- Fine-worthy superlatives on a landing page
- Comfortable, confident, fabricated specificity
The working rule that reconciles both: the facts are variables, the sentences are generated. Names, dates, visit counts, range-of-motion degrees, pain scores, spinal levels and sample sizes come from your practice-management system and are never invented. The model writes the paragraph around them. Prompt-level instruction is not enough on its own — wire the numbers, then let the prose be as direct as the numbers allow.
What you are allowed to say out loud
Three rule sets govern a chiropractic or PT ad, and they do not fully agree.
Google. Chiropractic and physical therapy fall under the Healthcare and medicines policy, which restricts targeting and creative for health services. Separately, the Personalized advertising policy has a health section that constrains audience building around health conditions, and Misrepresentation is the policy that catches an outcome promise. Note the common confusion in agency decks: 6020955 is Misrepresentation; 6020954 is Abusing the ad network. They are not interchangeable.
The FTC. Health claims require competent and reliable scientific evidence before the claim runs, and testimonials must reflect typical results or carry a clear disclosure. The Endorsement Guides live at 16 CFR Part 255 and the Consumer Reviews and Testimonials Rule at 16 CFR Part 465. The FTC's own business-guidance pages have been moving and 404-ing through late 2026, which is why the rule text is the durable link.
The state board. Every chiropractic and PT board has an advertising rule, and most prohibit claims of cure, guarantees of result, and misleading superlatives. A board complaint costs more in time than a Google suspension costs in revenue, and it does not come with an appeals process that runs on a two-week clock.
The claims that actually get clinics in trouble are rarely exotic. They are these:
- “We fix the cause, not the symptom.” An efficacy claim about mechanism, which needs substantiation you probably do not have.
- “Drug-free, surgery-free relief.” Fine as a positioning statement; not fine as an implied guarantee of outcome.
- “#1 rated in [city].” A superlative with no dated source. Replace it with the review count and the window.
- “Safe for everyone.” Unqualified safety claims are the riskiest line in this vertical, given what patients can find on their own about cervical manipulation.
Rewrite pattern: name the population, name the measurement, name the time frame, and leave the outcome where it belongs — in the patient's response, not your headline.
The number you should stop printing
The chiropractic industry quotes 35 million Americans annually. That figure traces to the 2016 Gallup-Palmer survey, which measured 35.5 million US adults who had seen a chiropractor in the previous twelve months — and it counted no children, despite being repeated across the web as an all-ages number.
The federal survey disagrees, and it disagrees in the other direction. The 2022 National Health Interview Survey put chiropractic use at about 11.0% of US adults in the past year. Gallup's own 2015 wave said 14%, or roughly 33.5 million. The two instruments ask differently and are not comparable year to year, which is exactly why quoting the friendliest one as though it were an annual census is a substantiation problem rather than a rounding error.
For a clinic the practical instruction is simple: use your own numbers. Active patient count, new patients per month, care-plan completion rate, re-exam attendance, review count and window. They are smaller, they are defensible, and they are the only claims a competitor cannot copy from a press release.
PHI is the prompt
A progress note contains protected health information. The moment you paste one into a hosted model you have made a disclosure to a third party, and the question is whether that party is a business associate.
45 CFR Part 164 governs this: a business associate agreement is required before a vendor creates, receives, maintains or transmits PHI on your behalf, and the minimum-necessary standard applies to what you send. “The vendor says they don't train on inputs” is a privacy posture, not a BAA.
The practical controls that make an AI note workflow defensible:
- De-identify on the way in. Send a patient identifier and the clinical variables, not the name, address, employer or the narrative of the car accident.
- Keep the record of record in your EHR. The model drafts; a licensed human signs. An unsigned model output is not a note.
- Log the template version. When a claim is reviewed eighteen months later, you need to know what the prompt said at the time.
- Retention should be describable in one sentence. If it takes a paragraph, it is too long to defend.
The cost math at live Venice rates
Rates checked on the live Venice price list today, November 14, 2026: venice-uncensored-1-2 at $0.20 in / $0.90 out per million tokens on a 128K context, deepseek-v4-flash at $0.138 / $0.275 on a 1M context, and gemma-4-uncensored at $0.1625 / $0.50 on 256K. The role-play variant sits at $0.50 / $2.00.
Model a two-DC clinic with roughly 900 active patients and 623 visits a month:
| Document | / month | Tokens in / out | venice-uncensored-1-2 | deepseek-v4-flash |
|---|---|---|---|---|
| Daily progress / SOAP note | 520 | 450 / 300 | $0.187 | $0.075 |
| Re-exam + outcome report | 60 | 1,600 / 800 | $0.062 | $0.026 |
| Plan-of-care narrative / ABN letter | 90 | 900 / 500 | $0.057 | $0.024 |
| Recall, reactivation + ad copy | 120 | 400 / 350 | $0.047 | $0.018 |
| Total per month | 790 | 459k / 291k | $0.354 | $0.143 |
Per document, that is $0.00036 for a daily note on the uncensored model. Put against the cost of one missed slot — about $200 — the entire month of AI drafting is 0.18% of a single no-show.
The comparison that matters is not model to model. It is against the alternative: a provider or PT spending 12 minutes per note. At 520 notes a month that is 104 hours of licensed clinical time writing sentences about numbers that already exist in the chart. The margin question in a rehab clinic in 2026 is not whether AI writes the note. It is whether the note contains the measurement that keeps the AT modifier standing and gives the patient something to see at visit five.
A seven-step action plan
- Split every patient by payer at intake. Medicare chiropractic is one procedure with one modifier and one denial mode; cash-pay wellness is a membership business; commercial PT is a plan-of-care certification business. One template cannot serve all three.
- Build the five templates before you evaluate a model. Daily note, plan of care, re-exam, recall, ad copy. Specificity is enforced by the template, not by the practitioner's mood at 6pm.
- Wire the facts from the EHR and never let the model invent them. Spinal level, P.A.R.T. findings, ROM degrees, visit number, outcome-measure score and interval are variables. If a number is not in the source data, the sentence does not get written.
- Audit ten notes a week against the AT standard. Ask one question of each: could a reviewer point to the measurable improvement this sentence claims? If not, the note is maintenance regardless of the modifier attached.
- Fix the re-exam before you fix the ad. The re-exam is the retention event. If it does not show a number moving against a baseline, you are paying $32.79 for a patient who quits at visit five.
- Rewrite every ad against the three rule sets. Google's healthcare and misrepresentation policies, the FTC's substantiation standard, and your state board's advertising rule. Replace superlatives with a count and a window.
- Get the BAA before the first PHI paste. Then de-identify anyway, keep the record of record in the EHR, and have a licensed human sign every note.
Chiropractic grew 2.8% this year while adding businesses at half a percent, which means the average clinic's growth came from taking patients from another clinic. Physical therapy has the cheapest qualified click in healthcare and a 55% self-discharge rate. Neither problem is solved by more traffic. Both are solved by a document that says something specific enough to be believed by a patient on visit five and by a reviewer eighteen months later — and a model that will write that sentence instead of the one that cannot be held against it.
Write the note that keeps the modifier standing
RawDialog runs venice-uncensored-1-2 and deepseek-v4-flash at $0.20 / $0.90 and $0.138 / $0.275 per million tokens — no filters sanding your progress notes into “patient reports feeling somewhat better,” no per-seat pricing, no training on your patient data. Bring your template library; keep your EHR as the source of every number.
Related reading on this blog: the Venice AI uncensored review, the developer API guide, uncensored AI for med spas and aesthetic clinics, gyms and fitness studios, and UGC video ad creatives.