The library provides access to over 3 million e-books (based on our 23-24 reporting to ARL). Most of that access is acquired through e-book packages and EBSCO is a big one, with over 370,000 titles in a range of subject areas. EBSCO sends us quarterly reports on the usage of those titles, with statistics on views and downloads by title, purchase type, publisher and subject area.
Last week I received our report for March through May of 2025. The csv file was 5620 rows, so I asked ChatCPT to help me make sense of it. I uploaded the data and used this prompt:
Analyze this file please. It is an excel spreadsheet that lists the usage of electronic books at an academic library. I would like you to provide top level insights into the usage by numbers and content trends.
Within seconds, ChatGPT provided me with this:
📊 Overall Usage Statistics Total Accesses: 15,892 Abstract Requests: 15,061 – indicates strong exploratory or research use Online Views: 13,363 – the dominant method of engagement Chapter Downloads: 1,984 Full Downloads: 545 – suggests some selective reading patterns Turnaways: 483 – may indicate unmet demand or access limitations Print & Email Requests: minimal (Print: 239, Email: 10) 📚 Top Subject Areas by LC Subclass These reflect areas with the highest interest: E – U.S. History LA – History of Education PN – Literature (General) ML – Literature on Music DT – African History HQ – Women’s Studies, Sexuality, Family HV – Social Pathology, Criminology, Social Work RC – Internal Medicine PS – American Literature LB – Theory and Practice of Education These trends suggest: Strong interdisciplinary interest spanning humanities, health, and education. Particularly high engagement with history, literature, education, and social sciences. |
It was interesting that the use of abstracts was so high, and I was surprised that U.S. history was the top subject. I was impressed with this output, and the speed with which the data was analyzed. But of course, there are experts here at the library who are in a far better position to evaluate the analysis – its accuracy and its value. I asked Karen Kohn, Collections Analysis Librarian, to review ChatGPT’s output with me.
Karen was patient. Typically she looks at usage data for an entire year rather than a quarter. That aligns with our renewal cycle, and of course we report these figures to organizations like ARL on an annual basis.
Let’s also keep in mind that EBSCO is one of dozens of e-book packages and platforms that we license. While EBSCO covers all subject areas, a trend within this package is not necessarily what we’d see in JSTOR or ACM (Association of Computing Machinery).
WIth those caveats, Karen and I continued our conversation with ChatGPT.
I asked to see an analysis of “turnaways.”
📉 Turnaways Analysis Turnaways occur when users attempt to access a title but are denied, often due to license restrictions (e.g., limited user seats). 🚩 Top Turnaway Subject Areas (LC Subclass) HF – Commerce / Business Administration (257 turnaways) BL – Religion / Mythology (105) DT – African History (33) E – U.S. History (27) RA – Public Health (15) This suggests a notable unmet demand in business, religion, and global history/public health content, which may merit review of current licensing models or potential acquisition of multi-user or unlimited access licenses. Let me know if you’d like a visual of subject turnaways or recommendations on mitigating high-turnaway titles. |
Karen says,
I can see places where it’s especially important to have the person writing the prompts think carefully about what to ask and look closely at the results.
There are a handful of serious outliers in this dataset, it doesn’t always know when to check for them. At one point ChatGPT points out that What the Best College Students Do accounts for 90% of Belknap Press’s total usage. Later the report notes that there are 257 turnaways in the HF section, which “suggests a notable unmet demand,” but it doesn’t point out that 256 of these are for a single book.
She’s right. There needs to be a human who is examining the data, looking for inconsistencies and prompting ChatGPT when there are results that need clarification. One of the ironies of using AI tools is this: we ask an AI tool because we don’t know an answer, but to best evaluate the output, we need some knowledge of the subject. Karen continues, explaining that:
This is analysis that I could do in Excel using pivot tables, but I recognize that not everyone has the same Excel knowledge, even people who look at collections data as part of their jobs. As someone who is comfortable in Excel, I really like to do things manually so I have more control over how the data is displayed and can look closer at specific details when I have questions, but not everyone wants to or can make their own Excel tables.
Fair point. Look out for Part II of this post, in which I ask ChatCPT to draft training materials on creating a Excel Pivot chart. And a big thank you to Karen Kohn who helped to remind us of the need for human intervention and a critical eye when incorporating AI tools into our work with library assessment.