PT Rates · Twin Cities

Live: https://pt-rates.vercel.app

Competitive Market Analysis. What everyone in your sector is charging for a given service.

— Aidan, Business Prototypes, 2026-10-12. Sibling instance of the same product: Salon rates (live at https://salon-rates.vercel.app). This one is the physical-therapy sector, where the price is what each insurer pays a clinic per billing code.

A phone-first page for one kind of reader: the owner of an independent physical therapy clinic, between patients, who wants to know whether an insurer pays them less than it pays the clinic down the road. Type the clinic name or NPI and the page says, per payer, "paid below the market median on N of M codes," with the code-by-code table underneath for the few who scroll.

grounded: 2026-10-04 — built from Aidan's PT market analysis repo at his request

The contract (2026-10-24)

Live: https://pt-rates.vercel.app/contract/ · one page per recipient at /contract/for/<token>/

A fixed-price engagement for the private sector, off the work behind our AHRQ Healthcare Choice submission: the buyer's fee schedule against the Minnesota market, per payer and per CPT code, the dollar gap at their own volume, and a one-page renegotiation case per payer. One engagement = one billing group (tax ID). BCBS Minnesota and HealthPartners in dollars, UCare in percent of charges; Medica, UHC, Aetna and Cigna are named on the page as not covered (DATA_MODEL.md).

$2,500, fixed. A hypothesis, set 2026-10-24. The one published rate card found for the same work: Contracting Providers charges $5,000 for a review of up to four payer contracts plus one negotiation, and $2,500 per additional negotiation (https://contractingproviders.com/services/payer-contract-negotiations, read 2026-10-24). This is the review and the case without us sitting at the table, so it is priced at their per-contract add-on. Against the buyer's side: a clinic under a payer's median sits about $4 a visit below it (see Price below), about $5,000 a year per payer at 1,250 visits; two payers make the fee about a quarter of one year's gap.

How it sells: POST /api/checkout with {"offer":"contract"} charges offer.json → contract.price_usd, tags the session metadata.kind=contract (and metadata.for=<token>), and the existing webhook emails the buyer what happens next plus the full-table password, and sends Aidan a Sale: CONTRACT notice. Same Stripe router registration as the table (metadata.product=pt-rates). Funnel events: contract_open, checkout_open, paid (all with offer=contract).

Recipients: outreach/contract/ledger.csv (gitignored; name, title, org, email + status, the record quoted, token). python3 scripts/build_contract.py writes private/bundle/contract.json (served one token at a time by api/contract.js, never static) and one draft per row to outreach/contract/drafts/. Drafts are not sent from here.

The paywall

Free: any clinic's own rates against the market, per payer, per code and per visit, plus four neighbouring clinics; the market's medians and spreads. Paid: the full table, every clinic and every payer side by side, and your own visit mix priced against all of them. Same pattern as msp-house-index.

site/ ships only aggregates (data/market.json) and the clinic directory. Every clinic's own rates live in private/payers/ (gitignored); scripts/build_public.py bundles them into private/bundle/, which scripts/deploy.py uploads as api/_full/, read by the functions only.

route	does
GET /api/preview?clinic=<id> / ?for=<token>	one clinic's own rates plus four neighbours' (the free slice)
POST /api/checkout	Stripe embedded Checkout at the price in site/offer.json; mints the buyer's password
POST /api/stripe-webhook	emails the password (Mailgun) once per session, and a sale notice to Aidan
GET /api/session-status	the return page's check; sends the password if the webhook has not
GET /api/data	password in x-password: the full table

/for/<token>/ opens on one clinic; tokens come from outreach/round1/ledger.csv (gitignored, names and addresses never ship). Funnel events, to Vercel and PostHog: preview_open, lock_click, checkout_open, paid, unlock. Lost password: python3 scripts/fulfil.py <session>; a free copy: python3 scripts/issue_password.py "comp:<label>" (then redeploy so the label logs). Vercel env: STRIPE_SECRET_KEY, STRIPE_WEBHOOK_SECRET, ACCESS_SECRET, MAILGUN_*, SITE_URL.

The per-clinic payer files were public in site/data/payers/ until 2026-10-15 and are still in this repo's git history before that commit.

Price

$500, once. A hypothesis, set 2026-10-15, not a finding. A clinic under its payer's median sits a median $2.67 (HealthPartners) to $5.02 (BCBS) a 45-minute visit below it (measured from private/payers/ that day). Call it $4. A three-therapist clinic sees about 7,500 visits a year (3 x ~10 a day x 250 days); if one payer is a sixth of them, 1,250 visits x $4 = $5,000 a year from closing the gap in one renegotiation. A tenth of that is $500. The sibling msp-house-index drew 10 checkouts and 0 payments at $100 (2026-10-11 to 10-12), so the price is the thing under test.

Who it was built for

Mason Richlen, DPT, co-owner of Maverick Physiotherapy (155 Eaton St, Saint Paul, inside Los Campeones Gym; org NPI 1073185393). A three-therapist, direct-access clinic. The page opens on Therapy Partners, Inc. (Saint Paul; a large independent practice that appears in all three payers' files, so every tab has an answer) with nothing typed; the search box is for any clinic.

The question this answers was first asked on 25 Jan 2026: how do people find out what a payer reimburses under Transparency in Coverage? The answer became three CSVs handed over on 2 Feb 2026 (clinic vs every local clinic per code, payer medians per code, data coverage) plus a markdown list of codes where the clinic sat below the market median. That last report is the product here. Its three uses: anchor a fee-schedule negotiation with each payer, prove underpayment against the local median per code, and show an employer or workers'-comp case manager what the market actually pays.

Per visit (2026-10-13)

The headline is now what each payer pays for a visit, not code by code: the people who negotiate PT contracts reason in dollars per visit against their own cost per visit ("If we can't get $81 per visit, then we're losing money" — a payer-contracting consultant on StrataPT, cited in customer-lab/maps/pt-rates.json). The page prices one visit per clinic per payer, puts it beside the market's per-visit spread and Medicare, and takes the owner's cost per visit (pre-filled with that $81 example) to show where a contract loses money and how much it trails the market per 100 visits.

• The mix is measured, not chosen. Codes and proportions are the share of treatment units Twin Cities PTs billed Medicare in 2024 (pipeline/medicare_pull.py → pipeline/visit_mix.py → site/data/visit.json): 44% exercise, 23% neuromuscular re-ed, 15% manual, 14% therapeutic activities, 2% gait, 1% self-care (98% of units).
• The length is the owner's. Minutes become units by CMS's 8-minute rule (45 min = 3 units, the default). Medicare's rows only bound a visit to 1.4–3.3 units, so no number is invented for it. Anyone can type their own units per code.
• UCare has no per-visit dollar (percent-of-charges); the page says so.
• Ideal model and what reality fills: DATA_MODEL.md. Distribution and sales rules: GO_TO_MARKET.md.

What is in, what is not

Payer	Rate rows kept	Twin Cities PT clinics	Codes with 3+ clinics	File dated
BCBS Minnesota (Aware Network)	254,112	195	39	2026-07-27
HealthPartners	4,060,908	134	39	2026-10-01
UCare	235	23	5	2026-09-16

278 clinics in the search box, from 4,812 NPPES PT/PTA records. Built 2026-10-14.

Not in, and why (each cost a day to learn in Feb 2026):

Payer	Why not
UnitedHealthcare	NPIs appear in its files, but none are linked to PT rate entries.
Medica	HealthSparq portal with bot protection.
Aetna	HealthSparq portal; national file structure.
Cigna	CAPTCHA in front of the file list.
Humana	Bot protection and CAPTCHA.
Medicaid	Published separately by Minnesota DHS in another format. (Medicare is on the page as a per-visit reference line, from CMS's 2024 billing data.)

Rules the numbers obey:

• The payer is always an axis. In the Feb 2026 data HealthPartners' medians were the lowest and BCBS's the highest; a blended number would lie to everyone.
• Clinic level before any median. HealthPartners publishes rates by individual (Type 1) NPI only, UCare by organisation (Type 2) only, BCBS by both. Individuals are rolled up to a clinic first (override map → shared TIN → shared NPPES address → TIN group → solo), or a three-therapist clinic is counted three times. pipeline/precompute.py prints the roll-up tally.
• One rate per clinic per code per payer = the median of the distinct rates seen for that clinic's NPIs (the same fee schedule repeats across networks).
• A market median needs at least 3 clinics, otherwise the code is not shown.
• Only negotiated / fee schedule professional rates are kept.
• Base prices only. Each NPI/code carries the base rate plus modifier variants (52/53 reduced service, CO/CQ assistant at 85%). Maverick's 97012 under BCBS read as eight "rates" from $9.73 to $22.88 until the modifiers were separated. Modifier prices are tagged at ingest and dropped.
• BCBS = the Aware network (its broad commercial PPO, file group 000000011, 31,950 plans cite it). The other 13 BCBS file groups are custom or narrow networks (Allina, PEIP, High Value, ...); they are in the DuckDB, tagged by network, and not on the page.
• UCare's PT procedure codes are published as a percentage of charges (65%, institutional), not dollars, so only its dollar-priced codes (97014 and the office-visit codes) appear. UCare is in, but it answers little.
• Every number on the page comes from site/data/. Nothing is typed.

Data sources

Source	What	Date
HealthPartners TiC in-network files (mrfproddestinationdata.blob.core.windows.net/mrf-output/2026-10-01_HealthPartners_*_in-network-rates.zip)	negotiated rates by individual NPI	file dated 2026-10-01
UCare TiC table of contents (ucm-p-001.sitecorecontenthub.cloud/.../ucare_toc.json)	negotiated rates by org NPI	file dated 2026-09-16
BCBS MN index (mktg.bluecrossmn.com/mrf/2026/2026-01-01_..._index.json, 645 "Local" files)	negotiated rates by provider group	file dated 2026-07-27
NPPES registry API, taxonomies 225100000X / 225200000X, zips 550, 551, 552, 553, 554, 556	the provider list and clinic names	pulled 2026-10-04
config/cpt_codes.yaml in the pipeline repo	the PT CPT codes	—
CMS Medicare Physician & Other Practitioners by Provider and Service (data.cms.gov dataset 92396110-…)	the visit mix and the Medicare reference line	2024

Plain-language code names are in pipeline/cpt_gloss.json.

How it is built

Physical-Therapy-Market-Analysis (DuckDB)  ──precompute.py──▶  site/data/*.json  ──deploy.py──▶  Vercel

• pipeline/refresh.sh drives Aidan's Physical-Therapy-Market-Analysis clone at ~/Desktop/workspace/pt-rates-data (patches in pipeline/patches/), then runs pipeline/precompute.py, which writes:
  • site/data/meta.json — payers, counts, file dates, code gloss, the default clinic
  • site/data/clinics.json — the search list
  • private/payers/<payer>.json — market stats per code + one rate per clinic per code (gitignored; served by /api/data behind the paywall, see "The paywall")
  • site/data/market.json — the public aggregates scripts/build_public.py derives from them
• site/ is static plus five small functions (site/api/). No database at request time; the page fetches one payer slice at a time.
• scripts/check_slices.py recomputes every median from the per-clinic rates and fails if meta.json disagrees with a slice.
• scripts/deploy.py posts site/ to Vercel over the REST API (project pt-rates). Needs VERCEL_TOKEN.

How to refresh

# 0. one-time: the pipeline clone
git clone https://github.com/2016judea/Physical-Therapy-Market-Analysis ~/Desktop/workspace/pt-rates-data
cd ~/Desktop/workspace/pt-rates-data && uv venv .venv && VIRTUAL_ENV=$PWD/.venv uv pip install -e .
git apply ~/Desktop/workspace/pt-rates/pipeline/patches/*.patch        # Oct-2026 HP URLs, tolerant UCare TOC
cp ~/Desktop/workspace/pt-rates/pipeline/user_config.json data/        # Twin Cities zips, Maverick as primary

# 1. NPPES + UCare + HealthPartners (19 networks x 700 MB, ~5 min each), then the slices
~/Desktop/workspace/pt-rates/pipeline/refresh.sh

# 2. BCBS MN: 342 Local parts from the newest monthly index (~40 s each; Aware is the first 25)
~/Desktop/workspace/pt-rates/pipeline/refresh.sh bcbs

# 3. slices only (after any ingest finishes), check, README table, deploy
~/Desktop/workspace/pt-rates/pipeline/refresh.sh slices
python3 scripts/check_slices.py && python3 scripts/update_readme.py && VERCEL_TOKEN=... python3 scripts/deploy.py

The two big ingests stream with ijson in ~170 MB of memory (pipeline/ingest_stream.py); the pipeline's own ingesters json.loads() whole files and swap the machine. To run both payers at once, write BCBS to a side file (ingest_stream.py bcbs --db data/rates_bcbs.duckdb --nppes-db <copy of rates.duckdb>) and pass --extra-db to precompute.py: DuckDB allows one writer per file. To build slices while an ingest is still writing, copy the .duckdb and its .wal and point --db at the copy.

Build of 2026-10-04 (evening): all 19 HealthPartners networks and all 342 BCBS parts are ingested. Networks 7–19 added rate rows but no new Twin Cities clinic; the slices keep BCBS's Aware network only (see above), so the BCBS counts did not move when the other 13 groups landed.

If a HealthPartners URL 404s, bump the YYYY-MM-01 prefix in config/payers.yaml to the current month — the blob store keeps only the latest month (Feb's files were gone by Oct 2026). If BCBS's index 403s, the newest mktg.bluecrossmn.com/mrf/2026/<YYYY-MM-01>_..._index.json that answers 200 is the one to pass as --index (2026-09-01 on 2026-10-04; the 2026-10-01 file was not yet public).

Next build

The vision is everyone in your sector. Today that means Twin Cities PT clinics on three payers; each new sector is a new instance of the same product (Salon rates is the second). UnitedHealthcare, Medica, Aetna, Cigna and Humana are the PT gaps listed above.

Not this

An AI intake agent for the clinic was considered and killed on 2026-08-23 for four independent reasons. This page is rates only.

MIT.