When a clinical AI tool answers a question without sending the user to a journal page, a missing click does not necessarily mean the journal was unused. Equally, a retrieved document has not necessarily been read by a clinician. Medical libraries need reporting that distinguishes machine activity, visible evidence and human engagement.
That distinction is timely. In its 18 September 2026 update, COUNTER described work on a second phase of its generative and agentic AI guidance, aiming for publication in October 2026. It was an update on proposals, not a finalised second-phase standard. The post referred back to guidance published in June and discussed off-platform use, third-party AI tools and possible broader retrieval reporting.
Follow one question through the system
Imagine a fictional hospital clinician asking an evidence tool a bounded question. The system fetches several documents, loads selected material into its generation context, cites two sources and displays a concise answer. The clinician opens one citation but not the other.
These events are not interchangeable. A document can be fetched and discarded. It can contribute context without appearing as a citation. A citation can be displayed without being opened. A source can be opened and found unhelpful. No single event establishes understanding, a change in management or clinical benefit.
COUNTER's September update discusses retrieval, grounding, citation, presentation and engagement as distinct telemetry stages, and proposed mapping work rather than a completed universal implementation. Preserve that status when discussing it with suppliers. Do not relabel a locally invented event as an official COUNTER metric merely because it sounds similar.
Ask for two views of the same activity
A library needs a contractual reporting view and an explanatory product view. The first should identify the applicable standard, report version and any compliance status. The second should explain what actually happened in the AI workflow, with enough detail to interpret the formal report.
For example, ask whether a reported source interaction means the content was fetched, included in the response context, cited or opened by the user. Ask how repeated fetches, cached material and a single answer using several sources are handled. Request the rule used to attribute off-platform activity to the relevant institution.
These are original procurement questions, not requirements claimed to appear verbatim in COUNTER guidance. Their purpose is to prevent two technically valid reports from being treated as measurements of the same behaviour when their definitions differ.
A proposed library dashboard
Keep the existing resource-use reports and add a clearly labelled explanatory layer rather than overwrite the historical series. A practical dashboard could contain:
| Reporting area | Interpretation question |
|---|---|
| Human access | Did people open or request source material? |
| AI-mediated evidence use | At which declared stage was content used? |
| Source diversity | Was use concentrated in a small subset of the collection? |
| Failed access | Were relevant sources unavailable because of entitlement or technical problems? |
| Verification behaviour | Did users inspect citations or report unsupported claims? |
| Method changes | Did the event definition or attribution rule change? |
This is a proposed local dashboard, not a new COUNTER report or a claim that iatroX already supplies it. Its categories should be mapped carefully to whatever official reporting the institution receives.
The library should also retain the distinction between activity and value. A resource used rarely may answer difficult questions that are not served elsewhere. A heavily retrieved resource may simply be easy for the system to access. Neither interpretation can be assumed from volume alone.
Avoid three budget mistakes
The first mistake is cancelling a source solely because direct clicks fell after AI deployment. Investigate whether some use moved into an AI-mediated route. The second is accepting every machine fetch as equivalent to a human reading event. The third is adding both measures together without considering overlap.
A fictional example makes the problem visible. An answer cites a paper and the clinician then opens it. The system may record an AI-mediated citation event and a human access event. These are two events in one journey, not necessarily two independent users or two separate clinical decisions.
Budget interpretation should therefore describe the journey as well as the count. Ask which questions the resource supports, whether the answer remains inspectable and whether loss of access would materially affect the service. This is a practical evaluation, not an argument that every licensed source must be retained indefinitely.
Protect confidentiality while measuring use
Detailed clinical prompts are not a necessary default ingredient of a library usage report. Start with the least sensitive data that can answer the evaluation question. Aggregate where appropriate, restrict access and avoid small-cell reporting that exposes individual clinicians or unusual patient circumstances.
The ICO's anonymisation guidance, checked on 19 September 2026, distinguishes genuinely anonymous information from data that remains identifiable through additional information. Calling a report aggregate does not remove the need to assess the actual disclosure risk.
Clinical query content may be useful for a specifically governed study, but that is a different purpose from routine subscription reporting. Keep the purposes, access arrangements and retention decisions explicit.
What this means for iatroX and its partners
This article is published by iatroX and includes its own source-grounded model in the discussion. The public methodology, checked on 19 September 2026, explains retrieval and citation-grounding design. It does not by itself establish a COUNTER-compliant reporting service, nor does this article claim that such a service is available.
For a library or publisher, the useful next discussion is concrete: which source-use events can be evidenced, which definitions apply and how can activity be reported without revealing clinical information? An answer replacing a click is a measurement problem to solve, not a reason to stop measuring.
Frequently asked questions
Was COUNTER's second-phase AI guidance final on 19 September 2026?
No. Its 18 September update described ongoing work and an intended October release, rather than a completed second phase.
Is an AI retrieval the same as a clinician reading a paper?
No. Retrieval, inclusion in context, citation, presentation and human engagement describe different events.
Does this article establish that iatroX provides COUNTER reports?
No. It proposes questions for source-use reporting and does not claim an undocumented institutional reporting feature.