Fifty patient queries, four AI systems: the protocol, and a correction.
Published before the data was collected, because a study designed by an interested party to confirm that interested party’s thesis deserves exactly the scepticism it usually gets.
An earlier version of this page was headlined “We ran fifty patient queries through four AI systems. Doctors were named in under 20%.” It was dated 15 July 2026 and presented as completed research.
That study has not been run. No data was collected, and the figure in that headline was not measured. It should not have been published in that form and we are not going to describe it as anything milder than what it was.
We have replaced it with the protocol below, which is what actually existed: a pre-registration setting out how the study will be conducted and what would falsify it. The earlier headline is quoted here rather than deleted, because a correction that hides the original claim is not a correction.
Anybody who quoted the under-20% figure on our authority should stop. We will publish the results entry when the study has been run, and this notice stays on the page permanently.
This is a protocol, not a result. It sets out how the study will be run, what would falsify the hypothesis, and when the data publishes. Results, raw observations and the coding sheet publish as a separate entry.
Until that entry exists, no figure from this study should be quoted as a finding — by us or by anyone else.
When a patient asks an AI system to help choose a doctor, how often is an individual clinician named?
And the secondary question, which matters more commercially: is being named predictable, or is it noise?
If the clinicians who surface share identifiable characteristics, that is actionable and a practice can do something about it. If they do not, then a great deal of what our industry currently sells is unfounded — including some of what we sell.
The claim that AI systems rarely name individual doctors circulates widely. We have made it ourselves. We have not seen it measured in an Indian context with a published method.
Second-hand claims are how bad practice propagates. Somebody measures something once, in one market, and it becomes an axiom quoted for three years after the systems have changed underneath it.
Fifty queries, three specialties, four query types.
| Element | Specification |
|---|---|
| Queries | 50, written in the phrasing a patient would plausibly use, including imprecise ones. The full list publishes with the results. |
| Specialties | Oncology, cardiology, orthopaedics. Chosen for high search volume, high stakes, and genuine variation in whether patients travel for care. |
| Query types | Direct selection · diagnosis-stage informational · comparison · verification. Roughly twelve to thirteen each. |
| Systems | Four, run in a clean session with no history. |
| Location | Indian queries from an Indian IP address. Results are not portable between markets and we will not treat them as such. |
| Runs per query | One. See the limitations below — this is the study’s weakest point and we are stating it rather than hiding it. |
The falsification condition, set before the data.
Our hypothesis is that individual clinicians are named in a minority of responses, and that where they are named the pattern is explicable rather than random.
The study falsifies that if clinicians are named in a majority of responses, or if the named clinicians share no identifiable characteristics — in which case our position that this is addressable work would be unsupported.
We commit to publishing all outcomes with the same prominence as a confirming result.
If the results are unflattering, they publish on the same schedule.
And if we quietly fail to publish, that should be read as a finding in itself.
We are stating that plainly because it is the only enforcement mechanism available. Nobody audits marketing research. The sole protection a reader has is a protocol published in advance and an author who has made it embarrassing not to follow it.
A study designed after the conclusion is known is not a study.
It is a sales document with a chart in it. The difference is whether the method was published first, and whether the falsification condition was set before anybody looked at the data.