AI Research Discovery: What the Data Tells Librarians

In the Ithaka S+R US Faculty Survey 2021, which drew 7,615 responses, only 14% of faculty said they most often begin looking for scholarly material on their library’s website or catalogue. 29% start at Google Scholar, and about a third at a specific database.

The split by discipline is wider still. 36% of social scientists start at Google Scholar, compared with 30% of scientists and 12% of humanists.

So for most researchers, the first search doesn’t happen on the library’s own surface. That matters for every conversation about AI research discovery, because the question for libraries isn’t only how AI finds research. It’s where researchers meet it, and what collection it is searching when they do.

The vocabulary problem, in numbers

Keyword search has a known weakness, and it has been measured.

In 1987, Furnas, Landauer, Gomez and Dumais studied how people spontaneously name the same thing across five domains. In every case, the chance that two people chose the same term was below 20%. Their simulations showed that a system relying on one designer’s preferred word would fail 80 to 90% of the time in many common situations.

Librarians have worked around this for decades with controlled vocabularies, subject headings and carefully built search strings. Researchers mostly haven’t. They type the words they know.

AI research discovery attacks the same problem from the other side. Instead of asking the user to guess the right term, it represents both the question and each document by meaning, so items that use different words for the same idea can still be matched.

Take a question such as “Does living near green spaces help children’s mental health in cities?” Relevant work is spread across public health (“greenness exposure”), urban planning (“urban tree canopy”) and paediatric psychology (“emotional and behavioural difficulties”). A single keyword string rarely reaches all three. Matching by meaning can.

What independent tests of AI search show

The most useful evidence so far comes from testing AI search against searches built by information specialists.

Oscar Lau and Su Golder (University of York and Hull York Medical School) took four published systematic reviews in public health, pharmacology and surgery. They asked Elicit, an AI research tool that searches the Semantic Scholar corpus, each review’s question, then compared what it found with the studies the reviews actually included.

Case study	AI search sensitivity	Traditional search sensitivity	AI search precision	Traditional search precision
Vaping harms	27.6%	96.6%	42.1%	14.7%
Physical activity	25.5%	98.0%	43.3%	1.03%
Fertility treatment	69.2%	92.3%	46.2%	13.8%
Breast reconstruction	29.3%	not reported	26.7%	0.65%
Average	37.9%	93.5%	39.6%	7.55%

Sensitivity is the share of relevant studies a search finds. Precision is the share of results that turn out to be relevant.

Three findings matter for libraries:

AI search was far more precise. On average, about 4 in 10 results were relevant, against fewer than 1 in 10 for the traditional searches. For the physical activity review, the original search screened 4,868 records to find 50.
It missed most of the relevant studies. Average sensitivity was 37.9%, well below the 90% usually expected of a systematic review search.
It found studies the experts missed. In three of the four reviews, the AI surfaced eligible studies the original searches had not found. In the fertility review, all four extra trials it found were not indexed in MEDLINE or Embase.

The authors conclude that AI search is a strong tool for scoping, preliminary searches and seed papers, and a useful adjunct, but not a replacement for systematic searching. That is a role librarians already understand: a fast way in, followed by a rigorous search when completeness matters.

Your collection is what the AI can see

The Lau and Golder results point to something simple: an AI search tool can only return what is in the corpus it searches. The extra trials it found came from outside the databases the experts used. The studies it missed were, in part, outside or poorly described in its own corpus.

For a library, that corpus is the collection. And most collections are already used very unevenly.

The University of Delaware Library published its usage analysis of a Wiley journal package for July 2024 to June 2025, at a cost of about $1,000,000 a year:

Measure	Value
Journals in the package	1,898
Total uses in the year	133,897
Journals producing 80% of all uses	about 378 (20%)
Journals with 1 to 10 uses in the whole year	740 (39%)
Journals with 11 to 50 uses	665 (35%)
Cost per use, all journals	$7.47

Delaware cancelled the package from 1 January 2026. Its report notes that 74% of titles received 50 or fewer uses in a full year.

Low use can mean low need. But it can also mean content that is hard to find: titles that describe their subject in a vocabulary researchers don’t search for. The Furnas data suggests that second group is not small.

This is where AI research discovery built on the library’s own collection changes the picture. When matching is by meaning, a relevant article in a low-use journal can surface for a question phrased in another field’s words. Zendy reports that institutions using Zendy AI Discovery often see database usage rise by 30 to 70%.

The practical value for librarians is a cleaner signal. If usage of a title stays low even when discovery works by meaning, low need is the likelier explanation. If it rises, the content was there but hidden.

What to measure in your library

Libraries are moving, but carefully. Clarivate’s Pulse of the Library 2025 found that 67% of libraries are exploring or implementing AI tools, up from 63% in 2024. In Asia and Europe, 37 to 40% are in initial implementation or beyond, compared with 14 to 16% a year earlier. Budgets are now the top concern for AI adoption, cited by 62%, up from 56%.

With budgets that tight, any AI research discovery tool has to prove itself in the library’s own numbers. Five measures make that possible, most of them already in your COUNTER reports and search logs:

Measure	Where to find it	What it tells you
Usage of low-use titles before and after	COUNTER title reports	Whether content was unused or unfound
Share of uses from outside the top 20% of titles	COUNTER title reports	Whether use is spreading across the collection
Zero-result and low-result questions	Discovery search logs	Real gaps in the collection, for collection development
Interlibrary loan requests for content you hold	ILL records	Researchers failing to find what you already pay for
Where sessions start	Web analytics, referral data	Whether discovery is moving back to the library’s surface

Together these turn a question of belief, “does AI discovery help?”, into one of evidence. They also give you the figures for the ROI conversation with leadership.

What the numbers add up to

The evidence points in one direction. Researchers mostly start their searches away from the library. Keyword search misses relevant work because people rarely use the same words. AI search finds more relevant results per search, but only within the corpus it can see. And much of what libraries pay for is barely used.

Put together, the case for AI research discovery in libraries is not about novelty. It is about making the collection the corpus the AI searches, and then measuring whether that changes what gets found.

The collection decides what AI can find. The data decides whether it worked.

See how Zendy AI Discovery works with your library’s systems