Double Dipping, Now with AI?

Usage is the second currency of scholarly publishing. The first is money. The second answers whether that money was well spent.

Since scholarly journals and books went digital, libraries have been able to see with considerable precision how often an article is accessed, a chapter downloaded or a database searched. COUNTER was founded in 2003 to make that usage comfortably and reliably comparable across publishers and platforms. Subscription prices and downloads could be combined into a cost per use. Usage became an argument in both directions: libraries could challenge unprofitable packages, while publishers could justify price increases with more content and greater use. Usage became a second currency for an entire industry.

That logic is beginning to run into a problem.

Younger students in particular are increasingly starting their research not in the library catalogue, a specialist database or a publisher platform, but by putting their question to a large language model. The model returns an answer, not a list of results. Perhaps with sources, perhaps without. Even if the underlying article was paid for by the library, the user may never open it.

Usage does not necessarily disappear. But it becomes invisible.

COUNTER itself now reports indications that libraries licensing AI tools are seeing larger declines in their conventional usage figures. Readers of a generated research summary often do not go on to consult the source material. And when an external system draws on licensed content, that use frequently does not appear in the publisher’s COUNTER report.

Outside scholarly communication, the replacement of clicks by AI answers is already clearly visible. The Pew Research Center analysed almost 69,000 Google searches. When an AI summary appeared, users clicked a conventional search result in only eight per cent of visits, compared with 15 per cent when there was no summary. Just one per cent clicked a source cited inside the AI answer. An Ahrefs analysis found that an AI Overview was associated with a roughly 58 per cent lower click-through rate for the top-ranked result.

Applied to scholarly content, the implication is simple: an article can be indispensable to an answer and still generate no download.

Not every development leads users away from the original, however. A study published in July 2026 on AI-mediated access to library resources found that ChatGPT, Perplexity and Gemini can also bring new visitors to openly accessible institutional repositories. Content with structured metadata, stable links and open access was particularly discoverable. AI can therefore displace scholarly content and make it newly visible at the same time.

A New Form of Double Dipping?

An uncomfortable question nevertheless remains for the economic relationship between libraries and publishers.

Several scholarly publishers already earn direct revenue from making their content available to AI companies. Wiley reported $40 million in AI licensing revenue for fiscal 2025, up from $23 million in the previous year. Informa expected more than $75 million from its Taylor & Francis partnerships in 2024. These arrangements included access to archive material used to improve large language models.

These agreements are not, of course, illegitimate. So far, they have been an important part of publishers’ AI strategies. Human access, model training and the integration of current content into AI products are different forms of use. Publishers are also investing part of the additional revenue in technology, open research and new products. Yet an economic conflict is emerging: the same content is licensed to libraries and sold to companies whose products may reduce the direct use of those library collections.

In open access, a comparable practice was debated under the label “double dipping”: publishers received subscription revenue and an additional publication charge for the same article. A new version may now be emerging. Libraries continue to pay for conventional access, while AI providers pay for machine use — and the human usage financed by libraries subsequently falls.

Usage is still too incomplete a measure for this new environment. An AI answer can rely on licensed scholarship without triggering a COUNTER download. COUNTER’s updated recommendations for AI and agent usage, published in June 2026, are therefore an important first step. They are intended to make human and machine access separately visible.

Better measurement, however, does not resolve the distributional question.

If publishers generate new revenue from the machine use of scholarly content while libraries record less direct usage, prices can hardly continue rising on the basis of the old arguments alone. In the next round of negotiations, libraries will ask who actually owns the productivity gain created by AI.

The publishers, because they package and license the content? The AI companies, because they turn it into answers? The scholars who created it? Or the libraries that financed access for decades?

And if a scholarly answer is sold twice but never clicked once, who, exactly, has used it?