What AI-mediated discovery changes for authors

For most of the modern history of scientific publishing, a reader arrived at a paper through channels an author could at least picture. A journal table of contents. A database search with a query the searcher typed and could refine. A reference list in a related paper. A colleague forwarding a PDF. Each of these was imperfect and several were quietly unfair, but they shared a useful property: they were legible. You could ask where your paper appeared, and get an answer that made sense.

A growing share of encounters with the literature now begins differently. Someone asks a question in natural language of a system that retrieves a handful of sources, synthesises them, and returns an answer. The reader may never see a result list at all. From the author's side, this collapses two previously separate stages — being found and being read — into a single opaque step, and it removes the feedback that used to tell you something was wrong.

Absence stops being visible

Under the old model, a paper that was not being read still showed up somewhere. It appeared in a search result the user scrolled past. It sat at position forty. It had download statistics that were low but non-zero. Low engagement was legible as low engagement.

Under the new model, a paper that is never retrieved produces no signal whatsoever. It is not ranked last; it is absent. From the author's point of view this looks identical to a paper nobody found interesting, and identical to a paper that does not exist. The three cases are indistinguishable from the outside, and only one of them has a remedy.

This is the single most important shift to internalise. The relevant question is no longer 'how well is my paper ranked', it is 'is my paper reached at all for the questions it actually answers'. That is a different question with a different set of causes and a different set of fixes.

Phrasing does more work than it used to

Retrieval systems match a question against the text and metadata of a record. If your abstract describes the population as it is described in your subfield, and the question uses the vocabulary of the person asking — a clinician, a patient advocate, a policy analyst, a researcher one field over — the match may not happen. Neither party is wrong. The vocabulary simply does not meet.

Historically this cost you a little. A determined searcher would try synonyms, follow a reference chain, or ask a librarian. In a single-shot question-and-answer interaction there is often no second attempt. The cost of a terminology mismatch has gone up substantially, and the cost of fixing one has not changed at all.

The same applies to structure. An abstract that states the population, the intervention, the comparison, the outcome, and the principal limitation in identifiable positions gives a retrieval system far more to work with than a narrative paragraph that mentions all five in passing. This is not a demand to write badly. Structured abstracts have been standard in clinical journals for decades precisely because they aid comprehension. The new pressure simply makes the existing advice more consequential.

Availability is now a discoverability question

Whether a lawful open version of your paper exists has always affected who could read it. It now also affects how much of your paper is available to be processed at all. A record with only a title and abstract in the open offers far less surface for a system to match against than one with full text available under terms that permit it to be used.

This is worth stating carefully, because it is easy to overclaim. There is a plausible mechanism, and there is a decision — deposit the accepted manuscript, register the identifiers, link the preprint to the version of record — that is entirely within an author's control and costs an afternoon. That is enough to justify doing it, without asserting a proven causal effect that the observational evidence does not yet support.

What this does not change

It does not change what makes research good. Nothing here is an argument for writing to be found rather than writing to be understood, and the two goals point in the same direction more often than not. Clearer statements of population and outcome help human readers first.

It also does not mean an author should optimise against a system. Keyword stuffing, inflated claims, and titles engineered for match rather than accuracy make the record worse and are the kind of tactic that any competent retrieval system eventually penalises. The durable version of this work is the boring version: say precisely what you studied, in the words your readers would use, and make the record complete.