Modeling is a legitimate tool. It is also a very different thing from knowing — and if you run a practice group, the difference shows up in real budget decisions every month. This piece lays out what modeled ROI actually is, what it's genuinely good for, and why a practice group (unlike a soda brand) doesn't have to settle for it.

What "modeled ROI" actually means

Media-mix modeling (MMM) and its modern descendants work roughly like this: take aggregate inputs (spend per channel per week, seasonality, promotions), take an aggregate output (revenue or visit volume), and fit a regression that estimates how much of the output each input "contributed." Dress it up with Bayesian priors and holdout validation, and you get statements like "paid search drove an estimated 22–31% of new patient revenue last quarter."

Understand what that sentence is: a statistical inference over aggregates, with assumptions baked in (how fast ad effects decay, how channels interact, how much of baseline demand exists anyway) and a confidence interval wide enough to matter. Change the priors, get a different answer. It observes correlation between spend curves and revenue curves — it never observes a single patient's journey, because it was designed for advertisers who can't observe one.

That's the crucial historical point: MMM was invented for CPG brands. Coca-Cola cannot know which TV ad produced which can of soda sold in which store. Estimation was the only option, so a rigorous estimation discipline grew up around that constraint.

A medical practice does not share that constraint. Its conversion isn't a can of soda in a distant store. It's an appointment on its own EHR schedule — a discrete, timestamped, location-tagged event in a system the practice owns.

Why healthcare agencies model anyway

If the ground truth sits in the EHR, why do traditional healthcare agencies report modeled ROI? Three honest reasons:

1. HIPAA locks them out of the patient-level path. Following a click from ad → website → form → booking → kept appointment means handling identifiable patient data at several steps. An agency orchestrating third-party tools — your CMS, a rented tracking vendor, a call-tracking platform, ad accounts — often has no BAA chain that permits it, and the consumer-grade tools in the middle can't legally carry the data anyway. Aggregate modeling is the workaround: if you only touch weekly spend and weekly visit counts, there's no PHI to mishandle. The compliance instinct is right. The measurement downgrade is the price.

2. The stack is fragmented. When the website, the forms, the call tracking, the analytics, and the scheduling system are five vendors with five data models, nobody can stitch a click to a kept appointment even if they were allowed to. Modeling papers over the seams.

3. Estimates are kind to retainers. This one deserves saying plainly, without cynicism: a modeled contribution number is unfalsifiable in any given month. "Our model estimates brand search contributed $340k" cannot be checked against anything. "You paid $214 per kept appointment at the Plano location" can. At $10k+ per month on a 6–12 month contract, which number would you rather be accountable to?

None of this makes modeling malpractice. It makes it a ceiling imposed by an architecture — the multi-vendor, no-BAA, orchestration architecture. Change the architecture and the ceiling goes away.

What closed-loop attribution gives you instead

Closed-loop means the measurement follows the patient's actual path, deterministically, from click to chair:

1. The click arrives carrying a click ID and UTMs. An owned first-party tracking script — one script, on infrastructure under your BAA — captures them against an allowlist and assigns a first-party visitor ID. No third-party pixel sees anything.
2. The visitor books through a widget that reads live availability from your actual EHR and writes the appointment back to it — provider, location, and appointment type mapped. The visitor ID is stitched to the booking at the moment of conversion.
3. The appointment resolves on the EHR schedule: kept, cancelled, or no-show.
4. Ad platforms get exactly one thing: a conversion event uploaded server-side against the click ID. Google and Meta learn "this click converted." They never receive a name, a condition, a form field, or anything else — so their bidding algorithms improve without a single byte of PHI leaving your control. (Full architecture: Cross-Location Attribution Under HIPAA.)

Out the other end comes the metric modeling can only gesture at:

CAC per kept appointment, by campaign, by location, no-show-aware.

Not "estimated contribution." A count. Campaign A produced 41 booked appointments at the Frisco location last month; 34 were kept; spend was $7,300; cost per kept appointment: $214.71. You can compare that to the visit's revenue, kill the campaign or triple it, and re-check the number next month. The healthcare PPC piece walks through running paid search on exactly this loop.

What each approach can and can't do

To be fair: what modeling is still good for

Honesty cuts both ways. Keep modeling (or at least directional judgment) for the places determinism can't reach:

• Non-clickable media. If a group runs TV, radio, or out-of-home, contribution there is inherently estimated.
• Long-horizon brand effects. Closed-loop credits the last measurable touch; a years-long brand campaign's halo is real and hard to observe per-patient.
• Very large systems whose demand drivers (payer contracts, referral networks, health-system politics) dwarf digital marketing.

The trap is using those legitimate cases to justify modeling everything — including the large majority of a typical practice group's acquisition budget — often most of it — that flows through search, social, and maps, where every conversion is individually observable on your own schedule. Estimating what you could simply count isn't sophistication. It's a workaround wearing a lab coat.

The attribution audit: 8 questions, no email required

Put these to whoever reports your marketing ROI:

• Is our "conversion" a form fill / phone call, or an appointment that landed on the EHR schedule?
• Can you tell me last month's cost per kept appointment (not booked — kept), per location?
• Is that number counted or modeled? If modeled, what's the confidence interval?
• What data, exactly, do ad platforms receive from our site — and can you show me the payload?
• Is every system in the measurement path covered by a BAA? How many BAAs is that in total?
• When a booked patient no-shows, does our reported ROI change?
• If we ended the engagement tomorrow, would the tracking and booking infrastructure leave with you?
• Which of the tools in our stack do you own, and which are resold?

A partner running a genuine closed loop answers all eight in one meeting. A partner running modeled attribution on rented tools will answer three and reschedule.

How to actually get closed-loop measurement

You can't buy this as a dashboard, because the dashboard is the last step, not the first. The prerequisites, in order:

1. EHR-integrated booking — real slots in, real appointments out, so the conversion event is the schedule entry. (Why EHR-integrated booking wins.)
2. One owned tracking layer — a single first-party script with click-ID capture and server-side conversion upload, replacing the pixel patchwork, under the same BAA as everything else. (HIPAA-compliant by default, including keeping GA4 safely.)
3. One accountable operator — because a loop stitched across five vendors breaks at every seam and every contract renewal.

That's the architecture PilotPractice ships as one system: the HIPAA-safe site, the booking widget with 30+ EMR integrations, the owned first-party tracking, and the AI voice receptionist and chat that capture the calls the loop would otherwise miss — one platform, one BAA, with our team running the marketing on top of it. It's the infrastructure enterprise agencies reserve for health systems, built for groups of 1 to 50 locations.

We'll happily show you the difference on your own numbers: book a demo and the first report you'll see is cost per kept appointment, by campaign, by location — counted, not modeled.