A usage report can look perfectly ordinary while the research session behind it has changed. A researcher asks an AI assistant to compare papers before a meeting. The system searches, checks several articles, draws on some of them, and returns an answer. The researcher follows one source. The rest of the session may leave almost nothing that looks like reading in a publisher report.
That is the awkward part. Publisher usage has never captured every encounter with an article, but it has relied on a relationship between research activity and visible events. People search, open abstracts, read HTML, download PDFs. Those actions leave traces.
AI can now do useful scholarly work before the researcher reaches the article, and sometimes no visit follows.
Counting machine activity does not solve the problem. An automated request may mean an article was examined closely. It may also mean a system checked a fragment, found it irrelevant, and moved on. Machine traffic can rise sharply without telling a publisher how much of that activity mattered to the research task.
The quieter problem runs the other way. An article may help shape a comparison or supply evidence for a synthesis yet never receive a direct visit from the person who asked the question. A citation may send the researcher back to the source. It may not. The platform record captures only the part of the session that reaches the platform.
That makes familiar numbers harder to read. Fewer downloads can still mean lower use, but they no longer tell the whole story. More automated requests may indicate growing machine interest, but they cannot simply be treated as additional readership. As AI tools move closer to authenticated researcher workflows, even an apparently ordinary interaction may become harder to interpret with confidence.
This is already the kind of mismatch that can surface in routine reporting: direct use softens, researcher reliance appears steady, and the numbers do not quite explain one another.
None of this makes established usage measures less useful. They remain the clearest account of direct activity publishers can observe. What has changed is the amount of research work that may happen around them. Searching, comparison, preliminary reading, and evidence selection can increasingly take place inside systems that reveal little about how an article contributed.
For analytics teams, the answer is unlikely to be another grand metric. The signals will remain uneven. Some machine activity can be identified. Some downstream use may never be visible. Reports will need more interpretation alongside totals, especially when platform activity and researcher behavior seem to be telling different stories.
A paper can matter to a research task without becoming a destination. As AI takes on more of the work between search and reading, publishers will increasingly need to understand the part an article played in the research session, even when the click never comes. Read more
Knowledgespeak Editorial Team
