Attribution in healthcare: what you can know
A patient sees a hoarding near the flyover in March. In June her father is diagnosed with something that needs a surgeon. She searches on her phone, reads three pages, asks a cousin, calls a number she finds on a listing site rather than yours, speaks to an agent, is told a consultant is available Thursday, and walks in on Thursday with a referral letter from a GP she had already visited in May. Her registration form records the source as “walk-in”.
Every attribution model in existence will get this wrong. The question worth asking is not how to get it right — you cannot — but how much wrongness you can tolerate and still allocate a budget sensibly. That is a different and much more answerable problem.
I have spent a fair amount of my career being asked for a number that does not exist, and a smaller but more useful part learning to replace it with three numbers that do. This is what I would tell a peer taking over digital at a hospital group tomorrow.
What you genuinely cannot know
Be specific about the holes, because vagueness here gets exploited by whoever has the prettiest dashboard.
- Offline conversion. The treatment happens in a building. Unless identity is matched from enquiry through to billing, the revenue end of your funnel is disconnected from the demand end, and identity matching in Indian hospitals is defeated routinely by shared phone numbers, the relative who registers on the patient’s behalf, and the number typed in wrong.
- Walk-ins. The largest single source category in most units and a near-total information void. “Walk-in” means “the front office did not ask” about as often as it means anything else.
- The contact centre gap. A call from a campaign that arrives on a published number rather than a tracked one is untraceable the moment it is answered. In most groups a meaningful share of campaign response arrives exactly this way, because patients call the number they already have.
- Long consideration cycles. Cookie windows are measured in days and weeks. Oncology and joint replacement decisions are measured in months. The attribution window expires before the decision is made.
- Cross-device and shared-device behaviour. One household, one phone, three patients. Platform identity graphs were not built for this and do not handle it.
- Influence without contact. The cousin who recommended you. The WhatsApp group. The apartment complex resident who had a good delivery. Unmeasurable, and frequently decisive.
What the platforms will tell you, and why it is wrong in a specific direction
Platform-reported conversions are not random noise. They are biased, consistently, in a direction that flatters the platform. Last-click over-credits search and especially brand search, which harvests demand other things created. View-through and engaged-view conversions over-credit video. Any model built inside one platform cannot see the other platforms at all, so the sum of all platform-claimed conversions routinely exceeds your actual appointment count, sometimes substantially.
This matters operationally because the channels that get over-credited are also the channels that are easiest to scale, so naive optimisation concentrates spend into brand search and remarketing — the two places where incrementality is lowest. You will show improving cost per acquisition and flat patient volume for two quarters before anyone connects the two.
Privacy and DPDP change the method, not the ambition
Health data is sensitive in every framework and India’s regime is explicit about purpose limitation and consent. Practically, three constraints bind you.
You cannot build audiences from clinical attributes without a consent basis that specifically covers it, and “they were our patient” does not cover advertising. You should not be pushing condition-level identifiers into advertising platforms at all, regardless of what a vendor tells you is technically possible — the reputational exposure of an oncology audience leak is not worth any amount of efficiency. And consent withdrawal has to propagate to every outbound system, which means your audience lists need rebuilding on a cycle rather than accumulating.
What this removes is the richest form of targeting. What it leaves is still workable: contextual and intent signals, locality, non-clinical behavioural segments, and aggregate measurement. The groups that handle this well treat the constraint as a design input from the start. The ones that handle it badly build a programme on patient-attribute targeting and then have to dismantle it, usually after a legal review rather than before.
The three numbers I trust
First, identified enquiries, deduplicated. A contactable human who asked about a service, with duplicates collapsed by phone number across form, call, chat and walk-in. This is measurable, defensible and close enough to the top of the funnel to be useful. It is also almost always lower than the sum of what each channel claims, and the gap is a useful hygiene metric in itself.
Second, enquiry-to-appointment and appointment-to-consult. Internal, unaffected by platform politics, and where most of your actual loss lives. You do not need attribution to measure these, and improving them raises the yield of every channel at once.
Third, total volume against total spend, at the unit-and-service-line level, over a rolling quarter. Blunt, aggregate, and honest. If spend rose and new patients in that service line did not, the channel reports are telling you a story about credit, not about causation.
Everything else I treat as directional evidence rather than fact.
Who is asking for the number, and what they actually want
When someone demands attribution they are rarely asking a measurement question. Learn to hear the real one, because the answer is different each time.
The chief executive asking “is digital working?” wants reassurance that the spend is not waste and a basis for defending it. A total-volume-against-total-spend trend by unit answers that better than any channel report. The chief financial officer wants to know the marginal cost of the next patient and whether it is rising; give them cost per identified enquiry and the enquiry-to-appointment rate, and be honest that the product of the two is an estimate. The unit head wants to know why their footfall is flat while the group celebrates a campaign; usually the honest answer involves slot availability or a consultant’s calendar, and attribution is the wrong instrument entirely. And the agency wants credit, which is a legitimate commercial interest and a poor basis for your allocation.
I have made the mistake of answering all four with the same deck. It satisfies nobody and trains the organisation to distrust the numbers, because each audience can see the part that does not address them.
Holdouts are the only real answer, and they are politically expensive
The only way to know whether spend caused volume is to withhold it somewhere and compare. Geographic holdouts are the practical version in a multi-unit group: switch off a channel in one comparable catchment for a defined period, hold everything else constant, and look at total enquiries and registrations rather than platform conversions.
This is harder than it sounds for reasons that have nothing to do with statistics. The unit head whose catchment is the holdout will escalate within a week. The agency will explain that the test is methodologically unsound — they are not always wrong, but the objection arrives suspiciously fast. Seasonality in Indian healthcare is strong and regional, so a four-week test spanning a festival is worthless. And you need a comparable catchment, which in a group of a dozen units might mean exactly one valid pair.
I still think it is worth doing once a year on your largest channel, and that you should brief the unit head personally, in advance, with the duration in writing and a commitment to restore spend on a fixed date. The first time I ran one I did not do that and the test was cancelled in its second week by someone senior enough that I could not argue.
What to do about the measurement gaps you cannot close
- Dynamic number insertion on digital properties, with the caveat that it only catches the patient who calls from the page rather than from memory. Helpful, partial.
- Ask the question properly at the front office. Not a twenty-option dropdown. Five options, asked out loud, with “saw something online” as one of them. Crude self-report beats a source field nobody fills.
- One identity, enforced at registration. Phone number as the spine, validated, with the relative’s number recorded separately rather than instead. This single discipline fixes more measurement than any tool.
- Accept lag deliberately. Report long-cycle service lines on a rolling quarter, never month-on-month, and say so in the review so nobody reads noise as trend.
- Reconcile to one source of truth monthly. Registrations from the HIS. Not the CRM, not the platforms. Whatever cannot be reconciled to a registration is a claim, not a conversion.
How to make budget decisions under genuine uncertainty
Stop trying to allocate precisely and start classifying. I split spend into three buckets and argue about the proportions rather than about line items.
Harvest — brand search, your own listings, remarketing, anything intercepting demand that exists. Cheap per conversion, low incrementality, necessary defensively. Cap it rather than optimise it, because it will eat the budget if you let cost per acquisition decide.
Capture — non-brand search, condition and procedure intent, aggregator presence, local discovery. This is where attribution is least bad and where most of the real work sits.
Create — awareness in a new catchment, a new unit’s launch, a service line nobody knows you have. Measured by aided awareness, branded search volume and total enquiry trend over quarters, never by last-click. If you hold this bucket to a cost-per-lead target you will zero it out within two budget cycles and then wonder why the new unit never matured.
Then decide the split as a judgement, document the reasoning, and revisit quarterly. A documented judgement that the board can challenge is worth more than a precise-looking number that nobody can interrogate.
If you are starting this next quarter
- Enforce phone-number identity at registration in every unit. Dull, unglamorous, the foundation of everything.
- Collapse the source dropdown to five options and train the front office to ask the question aloud.
- Reconcile one month of platform-claimed conversions against HIS registrations and put the gap in a slide. This conversation changes how your group talks about attribution permanently.
- Split the media budget into harvest, capture and create and cap harvest explicitly.
- Move long-cycle reporting to rolling quarters.
- Plan one geographic holdout for the quarter after, with the unit head briefed in writing and a fixed restore date.
- Write down what you have decided not to measure and circulate it. Naming the unknowns is what stops someone inventing a number to fill them.
Attribution in healthcare is not a reporting problem you will solve. It is an uncertainty you will manage, and the operators who do it well are the ones who stopped pretending otherwise early.
Questions people ask
Attribution is the attempt to credit a patient’s arrival to the marketing that caused it. In a hospital it breaks because treatment happens in a building, months after a search, often booked by a relative on a shared phone, via a number the patient already had. Walk-in is the largest source category and usually means the front office did not ask. The question worth answering is not how to get it right — you cannot — but how much wrongness you can tolerate and still allocate budget sensibly.
Because each platform is biased in a direction that flatters itself, and none can see the others. Last-click over-credits search, especially brand search, which harvests demand other things created. View-through conversions over-credit video. Summed across platforms, claimed conversions routinely exceed real appointments. The operational danger is that over-credited channels are also the easiest to scale, so naive optimisation pushes spend into brand search and remarketing, where incrementality is lowest. You will see improving cost per acquisition and flat patient volume for two quarters.
Identified enquiries, deduplicated by phone number across form, call, chat and walk-in — always lower than the sum of channel claims, and the gap is a hygiene metric. Enquiry-to-appointment and appointment-to-consult, which are internal, immune to platform politics and where most real loss lives. And total volume against total spend by unit and service line over a rolling quarter. If spend rose and new patients did not, the channel reports are telling a story about credit, not causation. Everything else is directional.
The CFO wants the marginal cost of the next patient and whether it is rising. Give them cost per identified enquiry and the enquiry-to-appointment rate, and be honest that the product of the two is an estimate rather than a fact. Reconcile everything to HIS registrations monthly, not to the CRM or the platforms; whatever cannot be reconciled to a registration is a claim. A documented judgement on budget split that the CFO can challenge is worth more than a precise-looking number nobody can interrogate.
It removes the richest targeting and leaves the method workable. You cannot build audiences from clinical attributes without a consent basis that specifically covers advertising — being a patient does not cover it. Condition-level identifiers should never be pushed into advertising platforms regardless of what a vendor says is technically possible; an oncology audience leak is not worth any efficiency. Consent withdrawal must propagate to every outbound system, so audience lists are rebuilt on a cycle. What remains: contextual and intent signals, locality, non-clinical segments and aggregate measurement.
It is switching off a channel in one comparable catchment for a defined period, holding everything else constant, and comparing total enquiries and registrations rather than platform conversions. It is the only real answer to whether spend caused volume. It is politically expensive: the unit head escalates within a week, the agency calls it unsound, and Indian seasonality can make a four-week test spanning a festival worthless. Run one a year on your largest channel, brief the unit head personally in writing, and fix the restore date.
Stop allocating precisely and start classifying. Harvest is brand search, own listings and remarketing — cheap per conversion, low incrementality, cap it rather than optimise it. Capture is non-brand search, condition and procedure intent, aggregators and local discovery — where attribution is least bad and most real work sits. Create is awareness for a new catchment, unit or service line — measured by aided awareness and branded search trend over quarters, never last-click. Decide the split as a judgement, document it, revisit quarterly.
Enforce one identity at registration: phone number as the spine, validated, with the relative’s number recorded separately rather than instead. That single discipline fixes more measurement than any tool. Then collapse the source dropdown to five options, with “saw something online” as one, and train the front office to ask the question aloud. Crude self-report beats a twenty-option field nobody fills. Both cost training time, not software. Dynamic number insertion helps at the margin but only catches the patient who calls from the page.
Because the decision cycle is longer than any attribution window. Cookie windows are measured in days and weeks; oncology and joint replacement decisions are measured in months, so the campaign that started the journey has expired before the patient walks in. Month-on-month reporting on those lines is noise read as trend. Report them on a rolling quarter, say so explicitly in the review, and accept the lag deliberately rather than letting someone invent a number to fill the gap.
Listen for the real question, because it is rarely a measurement one. The chief executive wants reassurance that spend is not waste and a basis for defending it, and a total-volume-against-total-spend trend by unit answers that better than any channel report. The unit head asking why footfall is flat usually has a slot-availability or consultant-calendar problem, where attribution is the wrong instrument. I once answered all four audiences with the same deck. It satisfied nobody and taught the organisation to distrust the numbers.
Credit is a legitimate commercial interest for an agency and a poor basis for your allocation. Expect them to report platform conversions, and expect those to be flattering. Insist that reconciliation to HIS registrations is done by your team, not theirs — if the agency owns analytics, you have marketing from your agency, not analytics. Expect resistance to holdout tests and weigh it accordingly. A good partner will accept a capped harvest bucket and a create bucket measured on awareness rather than cost per lead.
Most of it applies more easily to a single hospital, because there is one HIS to reconcile against and one front office to train. Identity at registration, the five-option source question, monthly reconciliation of platform claims to registrations, and the harvest-capture-create split all work at one unit. What a single hospital cannot do is a geographic holdout, since there is no comparable catchment. For that, a time-based pause on one channel with a written restore date is the imperfect substitute.
Reconcile one month of platform-claimed conversions against HIS registrations and put the gap on a slide. That single conversation changes how the group talks about attribution permanently. Alongside it, enforce phone-number identity at registration in every unit, collapse the source dropdown, split the media budget into three buckets with harvest capped, and move long-cycle reporting to rolling quarters. Then write down what you have decided not to measure and circulate it. Naming the unknowns is what stops someone inventing a number.
