The question I get asked most often by strangers when they find out what I do: “What should I read next?”
The question is asked eagerly, and yet we are supposed to have solved this problem by now through the power of algorithms that ingest reader habits and learn reader behaviors and deliver book recommendations precisely calibrated to sate reader hungers.
Are these algorithms giving me the kind of life-changing book recommendation that I have received from other readers from time to time?
Is technology helping readers find better paths from book to book, with fewer false starts and pitfalls and more transformative and transporting experiences along the way?
The best book recommendation engine is the knowledgeable clerk at a well-stocked, well-curated independent bookstore. To this recommender you verbally input the last few books you read and liked, and she outputs a title, physically handing you the book which you can buy and read alongside a cup of coffee in the café next door.
This recommendation engine has been replicated in the online space via the very low-tech Biblioracle, an occasional feature of magazine themorningnews.org. In this feature, author John Warner, the son of an independent bookstore owner, gives bespoke recommendations to online commenters. They input the last five titles they read and enjoyed, he spits out a recommendation. To this eye, his recommendations are quite good.
Like the real-world experience it replicates, however, it is not scalable.
The question that I get asked so fervently from time to time—“What should I read next?”—is surprisingly fraught. Books represent a large investment for readers in money and especially time and emotional energy. Acquiring a book and investing the time to read 25 or 50 or 100 pages only to cast it aside is a souring experience, maybe enough to sour certain readers on reading entirely.
The stakes are high.
Part of Amazon’s business model hinges on the notion that it can mine your behavior to suggest products—for our purpose, books—that you will like and want to read.
In the real world space, this function is served by the “featured” front table in the bookstore, or by the books face-out on the shelves.
But these efforts are laden with commercial conflicts that seem bound to get in the way of providing a useful recommendation.
Publishers and bookstores engage in “cooperative advertising” by which publishers pay bookstores to secure prime shelf space and placement on front tables.
Amazon engages in similar practices, with promotion in its online bookstore often contingent on payments from publishers. Whether or not these considerations come into play with regard to Amazon’s book recommendations, they are opaque to the reader, and a temptation to push books or categories based on outside factors is undoubtedly strong.
Amazon’s recommendations are also curious in that they are, by default, based on what readers have bought and not necessarily what they have read and loved.
What should a recommendation engine strive to do?
- Be transparent
- Ignore retail considerations
- Base recommendations on a reader’s reading habits
- Seek clues to what factors might make reader enjoy a book that they wouldn’t otherwise pick up
Neither a human nor an algorithm can meet these requirements perfectly, but a human is better suited to grasp the intangibles in play.
So what can algorithms strive to do?
Cataloging sites like Goodreads and LibraryThing seem best placed. The sites give the reader control over which books they catalog and therefore which books are the basis for the recommendations. The sites also do not have an explicitly retail function (though Goodreads is now owned by Amazon), hopefully lessening the possibility of conflicts of interest.
But the human element shouldn’t be dismissed as unworkable in the digital era:
Book communities may hold the most promise. Like-minded readers can offer recommendations that have the human touch, while crowd-sourcing makes the process scalable.
These idea may have to suffice until technology allows us each our own personal Biblioracle.