At the 2026 Charleston Premiers, Citations showed a reader asking an AI assistant a question, confirming the answer in the book, finding its shelf location and formats at nearby libraries, and borrowing a digital copy through the library’s reading platform. Attendees voted it Best New Product and Most Impactful.
A connector that can search a catalog and place a hold creates circulation events on the library’s behalf. Before patrons rely on it, the library needs to test each of those operations and check how each one is authorized and reported.
In short
An AI library connector chains several distinct events: a passage retrieved, a catalog searched, availability checked, a hold placed, a loan or renewal completed. Each event has its own permission and its own line in a usage report.
The five tests below cover the source, the circulation operations, the permission behind each operation, the usage report and the data that leaves the library. Each test uses a few complete journeys, failures included, with pass criteria agreed before anyone sees the results.
Test 1: Confirm the source behind the answer
A library connector should point the reader to the book and to the library. That works only if staff can follow the reference to the title, the edition and the passage that supports the claim. The reader catches a hallucination by checking the answer against that passage.
Test with questions whose answers staff already know. Each link should lead to the intended catalog record, and not to another edition or a similar title. For translated excerpts, the reader should see the original page beside the translation.
A plausible citation does not establish support, because a retrieved passage is not necessarily the one that answered the question.
Test 2: Run each circulation operation, including the failures
Catalog search, availability check, hold, renewal and loan are separate events. Test each with an authorized patron account in three situations: a title on the shelf, a title checked out, and a request the patron is not permitted to make, such as a blocked account, a reached hold limit or a non-holdable item type.
The connector should tell the patron what happened in plain language, including when an operation is refused or fails, and give the reason when the library system provides one.
The status of each operation in Citations differs, and the test plan should follow it. Loans, due dates, renewals and holds are live on Koha. Catalog search, holdings and links to e-resources are in the pilot. A library that takes part through a MARC export can be found, but the export does not create a live transactional connection, and its records are only as current as the export schedule. Koha exposes circulation operations, including renewal eligibility checks, through its REST API, so the test should confirm that each answer the patron receives reflects the library’s own circulation rules.
In the Charleston demonstration, the digital loan was completed on the library’s reading or listening platform. Borrowing directly through the connector is planned and not available. The walkthrough on the Citations Libraries page is illustrative, with a simulated library and reader. Confirm which operations are enabled in your configuration before announcing the service to patrons.
How each event should appear in a report
The table lists, for each event in the chain, what the reader should see, what the library should test and how a usage report should count it.
Event | What the reader should see | What to test | How a report should count it |
|---|---|---|---|
Passage retrieval | The supporting passage, with title and edition | Right passage and edition; original page beside any translation | Use of a passage from the work |
Catalog search | Matching records at this library and nearby | Known titles found; no records missing from the catalog | Discovery interest, including titles the library does not own |
Availability check | Shelf status, location and formats | Status matches the library system at the time of the request | Interest in a specific copy or format |
Hold | A confirmation, or a refusal with its reason | Allowed, blocked and limit-reached requests | A circulation event created through AI |
Loan or renewal | A due date, or a handoff to the e-book or audiobook app | Renewal eligibility; handoff to the reading platform | Completed circulation |
Test 3: Match the permission to the operation
Owning a physical copy, licensing a digital title and authorizing an AI service to expose passages from it are three different rights. Review the contract for the specific operation proposed, starting with who may bring which collection into an AI service.
In Citations, a reader can find a title and its holdings without a reading license. Passages are readable only when the publisher has granted the rights to the institution and the license has been accepted, and that licensed reading is in the pilot. A provider subscription does not give the right to read the title inside an assistant.
For special collections, archives and local history, record what the institution is authorized to share and whether access is open to the public or limited to members. Refer uncertain cases to the rights or legal team before the collection is made available.
Test 4: Read what the report actually counts
Search, retrieval and loan are separate counts, and a citation does not show that a passage shaped the generated answer. COUNTER, the standard libraries use for publisher usage reports, now makes similar distinctions.
COUNTER’s best practice on generative and agentic AI usage metrics, published April 20, 2026, adds an optional Agent access method to keep AI usage separate from human usage. It distinguishes AI responses generated, AI investigations (content an AI system selected for its final response) and AI requests (content whose full text the system could access).
Ask any provider for a sample report with event definitions, reporting period and coverage. In the Citations pilot, the library receives a monthly report, title by title, of what its readers discovered and read through AI, including titles the library does not yet own. The report carries no reader identity and never the text of a query. Selectors can use these counts as one input when deciding what to buy. The sample report on the Libraries page uses a simulated library.
A service reports only the activity that passes through it. Each count in the report answers its own question, which is also why separate layers of AI visibility should be read one at a time.
Test 5: Map the data that leaves the library
Ask which identifiers and request data move between the library system, the connector and the AI service, what is retained, and who can access it.
In Citations, licensed text is served from where its owner keeps it, passage by passage under the terms set by the publisher or library, and no book file is handed over for training. For e-resources, Citations links to the library’s provider and does not open the file. Sign-in uses the institution’s identity for access, and the monthly report shows titles and counts without reader names, individual reading histories or query text. As co-founder and CEO Jim Bryant put it at Charleston, “Your book stays home. The AI visits.”
For retention, deletion and subprocessors, ALA’s ten-item vendor review on logging, retention and deletion provides the checklist. The library should also ask how counts are shown when a collection or audience is very small, since a single reader can be identifiable in a short list.
Scaling the tests to a pilot
Five tests can look heavy for a service patrons may try a few times. Tests 2 and 3 are largely scripted pass/fail checks, and Test 1 can start with a dozen questions whose answers staff already know.
Agreeing on success criteria before anyone sees results is the step to keep in every case, as with acceptance tests agreed before results arrive.
Before announcing the service
Record the collection, the audience, the permissions, the supported operations and the success criteria. Run a few complete journeys from question to hold, including those that fail, then announce to patrons only the operations that passed in your configuration.
Publishers and libraries can test the connector through the Charleston Circle pilot, which is open now and free, with no membership cap and no fixed term. Libraries can open collections they are authorized to share, such as archives, local history and rare holdings, to the public or only their members, through the Citations Libraries connector. Publishers can open a private reading room where libraries preview new titles before they order.
Every member gets a seat at the table as Citations negotiates with AI companies. One model under discussion would have each authorized use pay both the publisher and the reader’s library from an escrow prefunded by the AI company. No such payment exists until AI companies adopt the model.
Request a pilot place or a demo
Related reading
The procurement questions libraries now put to AI vendors are mapped in ALA’s AI vendor checklist.
The same acceptance-test method, applied to a tutor built on licensed educational content, is set out in six acceptance tests for an AI tutor.
Why a retrieved passage does not show which source shaped the answer is explained in retrieved is not cited.
Sources
Library Technology Guides — "Citations wins Best New Product and Most Impactful awards at Charleston Premiers," November 2026
Citations wins Best New Product and Most Impactful awards at Charleston Premiers
COUNTER Metrics — "Best Practice: Generative and Agentic AI usage metrics," April 20, 2026
Best Practice: Generative and Agentic AI usage metrics
Koha Community — "Koha REST API: Checkouts" (checkouts, renewals, renewal eligibility)
Koha REST API: Checkouts
American Library Association — "Guidance on the Use of Artificial Intelligence in Libraries" (CD#44.2), July 2026
Guidance on the Use of Artificial Intelligence in Libraries (CD#44.2)
Citations — "For libraries," product statuses checked November 16, 2026
Citations for libraries
This article is operational analysis for publishing and library teams, not legal advice.