The Goodreads “Classics”: A Computational Study of Readers, Amazon, and Crowdsourced Amateur Criticism

hat is a classic? This is "not a new question," as T.S. Eliot acknowledged more than seventyve years ago. 1 More than simply "not new," this question now feels decidedly old, hashed out, and even passé. Perhaps most glaringly outdated is the word "classic." Literary scholars don't often use the term anymore, at least not as a serious label for literature "of the highest rank or importance." In 1991, John Guillory declared that the term classic was "all but retired." 3 The label, according to Guillory, signi ed not only a "relatively uncritical regard for the great works of Western literature" but the "precritical era of criticism itself." 4 Instead, in academic conversations, the ardent language of the "classics" has largely been displaced by the more critical vocabulary of the "canon," which frames literary signi cance more carefully as a product of cultural selection.

This ood of Goodreads classics content represents an excitingly large archive of amateur criticism and reader responses, an opportunity for scholars to hear nonacademic readers' voices in ways that were dif cult if not impossible before the internet. For example, understanding how readers felt about classics in the Victorian period is dif cult because there is little rst-hand evidence from Victorian readers, as Richard D. Altick once explained: "The great majority of the boys and girls and men and women into whose hands fell copies of cheap classic reprints did not leave any printed record of their pleasure. Only occasionally did the mute, inglorious common reader take pen in hand." 6 Far from this "mute, inglorious" Victorian common reader, the twenty-rst-century readers of Goodreads regularly publish records of their readerly pleasure and displeasure on the internet. Beyond providing a rich archive of reader responses, Goodreads also raises questions about whether its social network might enable a democratization of the classics. The classics, after all, have historically been de ned by those in power and excluded "the interests and accomplishments of minorities, popular and demotic culture, or non-European civilizations," as Ankhi Mukherjee describes. 7 To what extent are millions of Goodreads users from around the globe now remedying or replicating such historical exclusions?
Though Goodreads data is a boon for literary criticism and a potentially transformative development for literary culture, it is also a boon for corporations. Amazon's looming shadow over Goodreads data helps bring into focus a more nancially minded de nition of a classic, perhaps best summarized by poet and literary critic Mark Van Doren. A classic, Van Doren said, is simply "a book that remains in print." 8 For the twenty-rst century, we might update Van Doren's de nition and say that a classic is simply a book that continues to make money in whatever form it takes, whether as a print book, audiobook, e-book, screen adaptation, or as the subject of millions of online book reviews. In fact, it is clear, based on the Goodreads reviews that we analyze in this essay, that industries such as lm, television, publishing, e-commerce, and tech not only pro t from the classics but pro t from each other in a circular loop, bene ting from the reinforcement of works as classics in other mediums and domains. Considered together, this classics industry, as we call ita formulation inspired by and indebted to Simone Murray's "adaptation industry" as well as Pierre Bourdieu's theories of cultural production 9proves to be one of the strongest in uences on Goodreads users' perception of the classics.
The tensions between democratic potential and corporate exploitation that we observe in the Goodreads classics are characteristic of many social networks and Web 2.0 platforms, which fundamentally rely on user-created content. These dynamics have been studied extensively by scholars of fandom, new media, and digital culture, among other elds. 10 Yet how social network dynamics and the internet economy are reshaping literary culture, in particular, is still a relatively new conversation, led by critics such as Murray, Aarthi Vadde, Lisa Nakamura, and Mark McGurl. 11 By examining Goodreads reviews in this essay, we hope to contribute to this emerging conversation. We also hope to add a quantitative, data-driven perspective to the discussion by curating a collection of more than 120,000 Goodreads reviews and by using computational methods to study some of the most salient trends. We believe that digital literary culture is an area that especially rewards the convergence of digital humanities methods and contemporary literary criticism. The massive number of Goodreads ratings and reviews is part of what makes the platform worthy of study and nancially lucrative, but also what makes Goodreads dif cult to understand in broad strokes. Digital humanities and cultural analytics scholars have demonstrated, however, that computational methods can help us better understand cultural phenomena at scale. By employing these methods on Goodreads data in particular, we build on previous Goodreads-related DH research by Karen Bourrier and Mike Thelwall, J.D. Porter, Alexander Manshel, and Laura McGrath, James F. English, Scott Enderle, and Rahul Dhakecha, Allison Hegel, and Andrew Piper and Richard So, among others. 12 Scale is not our only motivation for using computational methods. The contemporary book world, including but not limited to Goodreads and Amazon, is increasingly governed by algorithms and data, which presents a number of challenges for contemporary literary scholars. "Clearly the leviathan that is Amazon exerts immense in uence on the global book trade," as Simone Murray contends, "but how are scholars to document, much less critique, algorithmic culture's self-reinforcing effects on cultural selection if denied access to the workings of the algorithm's engine-room?" 13 To provide one answer to Murray's urgent question, we believe that computational methods can supply a way of documenting, understanding, and critiquing algorithmic culture and its effects. By collecting and analyzing Goodreads data with computational methods, we are able to see, for example, that Goodreads only publicly displays a small fraction of its data. We are also able to detail some of the speci c social effects of the platform's default sorting algorithm, which prioritizes the most liked and most commented on reviews. This digital infrastructure produces a feedback loop among Goodreads reviews, in which reviews that receive attention continue to receive more attention ad nauseam. This feedback loop is a tting metaphor for the consecration of the classics on Goodreads more broadly. Though Goodreads users technically de ne the classics for themselves, their de nitions are clearly shaped by a reciprocal system of reinforcing in uencesold institutions like high schools, universities, and publishing houses as well as new institutions like Amazon. The result is a reader-produced vision of the classics that is surprisingly less diverse, in terms of authors' race and ethnicity, than those represented by U.S. literature syllabi, though more diverse in terms of genre, including more genre ction, young adult ction, and adapted ction. Though Goodreads users seem strongly in uenced by traditional institutions and the capitalist marketplace, they nevertheless demonstrate enormous creativity in nding ways to make this critical conversation their ownparodying and panning different literary styles, reliving and reimagining high school English classes, pushing back against the perceived arbiters of literary authority, and publicly changing their minds.
To close this introduction, we foreground our own approach to Goodreads data, since the exploitation of user data is a central subject of this essay. We have chosen not to publicly share our dataset of Goodreads reviews, though we have shared the code that we used to collect data from the Goodreads website, and we have obtained explicit permission from each Goodreads user who is directly quoted in this essay. 14 We believe that ethical approaches to user data will continue to be one of the most important conversations for digital humanities and cultural analytics research, and we expand on our choices further in the Appendix.

The Classics "Shelf": Genre, Hashtag, Advertising Keyword
This essay understands Goodreads users to be readers as well as "amateur critics," a framing that we draw from Aarthi Vadde, Melanie Micir, and Saikat Majumdar, among others. 15 As Vadde explains, "The ease and ubiquity of digital publishing have enabled the 'mass amateurization' of the critical, creative, and communicative arts, allowing amateurs to bypass the gatekeeping practices of speci c institutions...and to perform acts of photography, journalism, or authorship without necessarily identifying with a specialized guild or bene tting from its resources." 16 The digital platform of Goodreads similarly allows amateurs to perform acts of literary criticism, to publish their own analyses and judgements of literature, without formal training and without access to traditional publishing venues. The three main ways that Goodreads users act as amateur critics are by rating books between one and ve stars, by reviewing books in 15,000 characters or less, and by "shelving" books into categories. We begin with an extended discussion of Goodreads "shelves" because they are one of the primary ways that users de ne the classics and that Amazon pro ts from the classics.
The rst telling clue about these shelves is that the Goodreads website uidly refers to them as "shelves," "genres," and "tags." This slippery relationship points to a signi cant evolution of genre among readers and amateur critics in the twenty-rst century: genre is being subsumed and reshaped by tagging. Tagging is a common system for classifying and organizing content on the internet, in which users tag digital content with their own freeform descriptions, keywords, and metadata (think hashtags on Twitter). The shelf system on Goodreads is a social or collaborative tagging system because users can apply different tags to the same content, essentially crowdsourcing book categorization. Prior computa-tional social science and natural language processing research has explored how these collaborative tagging systems produce "folk taxonomies" or folksonomies, classi cation systems built by communities from the ground up. 17 Literary genre, in the hands of internet users equipped with tagging systems, has similarly blossomed into a grassroots taxonomy that incorporates conventional genres but also splinters into new genres, microgenres, publishing industry categories, reception metadata, hashtags, and more. 18 For example, a Goodreads user named Candace tagged Margaret Atwood's The Handmaid's Tale (1985) as "classics" and six other distinct categories ( gure 1): "wtf-did-i-just-read," "kindle-unlimited," "dark-themes," "favorites," "listened-to-audio-version," and "age-difference." 19 Fellow Goodreads users tagged The Handmaid's Tale as "science-ction," "fantasy-sf," "man-booker-shortlist-longlist," "tv-series," "re-read," and "feminism," among many other tags. As these examples demonstrate, Goodreads users mold conventional genre to better represent their tastes, values, and cataloguing needs. Allison Hegel argues that Goodreads shelves may also help readers "articulate their identities to others and connect with larger communities." 20 According to Jeremy Rosen, most literary critics today understand genre "not as a rigid category that texts 'belong to' or a set of rules that one must abide by, but as a exible set of techniques that can be adapted according to the needs of its users." 21 While Rosen's "users" are mostly authors, who mold genre to create their own literary works, the ambiguous term suggests that others can use genre, too, including readers and amateur critics. Thus, "classics" emerges as an important contemporary genre for readers in addition to a label of literary value and publishing category. categorize books with their own personal shelves/genres/tags. Notably, Candace has shelved The Handmaid's Tale in "classics." The number of likes and comments that this particular review received is also underscored because it is the basis by which Goodreads sorts reviews by default, which we discuss in more detail in the section "The Goodreads Algorithmic Echo Chamber." Though "classics" is just one Goodreads shelf among thousands, it is one of the most important and foundational. In the website's earliest days, the company used "classics" as their rst anchoring example to introduce and explain the shelving system: "You can cre-ate your own personal bookshelves. From classics to canadabooks, to childrenslit and geek, you can create any category that suits your personal taste." 22 Ten years later, the classics remained Goodreads go-to example: "Shelf names range from classics and coffee-table-books to childrens-lit and sci-you can create any category that suits your personal taste." 23 Because "classics" supposedly represents the oldest and most traditional literary category, it serves as an effective foil for any unconventional literary category a Goodreads user might dream up, and it also invites a mass of readers and amateur critics to participate in a seemingly elite conversation. The classics thus make the entire shelf system legible and appealing.
Shelves are also nancially lucrative for Goodreads and Amazon, the classics shelf particularly so. Each time a Goodreads user shelves a book in their personal library, that user simultaneously shelves the same book in the platform's massive library of more than two billion books. 24 "Goodreads turns the reader into a worker," as Lisa Nakamura observes, and through shelves, the company crowdsources the enormous work of organizing two billion books to the masses.
By shelving books, Goodreads users also (more unsettlingly) organize themselves into coherent audience categories that can be effectively targeted by advertisers. The same shelves that Goodreads users invent are sold as advertising target keywords, as Goodreads' informational material for advertisers shows in gure 2. These shelves represent not only books but also people: the Goodreads users who form communities around genres and subject areas, who read and discuss the books shelved into these categories. Browsing through the list of advertising "target values" reveals that some of these shelves are fascinatingly niche like space-opera, mermaids, and reformation-history. Yet other target values like mental-illness and abuse seem more serious and sensitive, raising the concerning possibility that vulnerable groups might be targeted by advertisers. Goodreads ags the "classics" as one of their top 10 most "prominent" genres for advertisers, putting it in the same company as "contemporary," "historical-ction," "fantasy," " ction," "manga," "mystery," "romance," "non-ction," and "youngadult." Looking at the top 10 most rated books across the entire Goodreads website offers one clear picture of this prominence: ve of the top 10 are classics (Table 1). Figure 2: The rst page of a four-page document titled "Genre List for Advertisers," described as "the master set of genres currently available to be used as target values for your ads on Goodreads." The "classics" is bolded as one of the top 10 most prominent genres near the bottom-center of the page. This "Genre List for Advertisers" document can be found under "Target Advertising" on the "Advertise with Us" section of the Goodreads website. To fully grasp the signi cance of Goodreads users' shelving labor, it is helpful to compare Goodreads to Net ix, the world's largest video streaming service. Like Goodreads, Net ix has a massive microgenre system for its video content, featuring hyper-speci c genres like "Deep Sea Horror Movies" and "Romantic Dramas Based on Classic Literature." To assemble these 70,000+ "altgenres," Net ix "paid people to watch lms and tag them with all kinds of metadata," as Alexis Madrigal reported in 2014. 26 "When these tags are combined with millions of users' viewing habits, they become Net ix's competitive advantage," Madrigal argues. "The data can't tell them how to make a TV show, but it can tell them what they should be making." 27 By tagging books with their own extremely detailed metadata, Goodreads' 120 million users perform a similar service for Goodreads and Amazon, but they do it for free. 28

The Classics According to Goodreads Users
When Goodreads users shelve books, they supposedly classify books on their own terms without direct intervention from the academy, the publishing industry, or Amazon. Technically, any of the two billion books in the Goodreads library could become a classic in users' hands. Yet when we collate the books that Goodreads users have collaboratively consecrated as classics, we nd the strong in uence of school curricula and what we call the classics industry, the interrelated network of businesses that generate and pro t from the classicssuch as publishing, lm and television, and internet corporations like Goodreads itself. To identify this list of Goodreads classics, we rst selected the top 100 literary works tagged as a classic the greatest number of times by Goodreads users throughout the site's history (2006-2019). We then added the top 100 literary works that were tagged as a classic and most read by Goodreads users in the rst week of September 2019 (the week when we collected our data). The homepage for popular shelves like the classics prominently features books that were "Most Read This Week," displaying them even above the most tagged books in the genre. We decided to include this second group of books because they are conspicuously promoted by Goodreads and provide a slightly different perspective on the Goodreads classicsnot only what users have tagged as classics but also which classics users actually seem to be reading. Many of the 100 most read classics overlap with the 100 most shelved classics, and in total the list includes 144 unique titles. 29 Many texts labeled as classics by Goodreads users seem to overlap with English literature curricula from U.S. grade schools, high schools, and colleges. Though the Goodreads platform has an increasingly global audiencewith notable emerging userbases in India and the UKmost of its users have historically hailed from the U.S. and still make up an estimated 40% of sitewide traf c. 31 For two rough estimates of how much the Goodreads classics overlap with school syllabi, we consulted a recommended reading list from the Advanced Placement (AP) English programa common literature curricula in U.S. high schoolsas well as a compilation of college-level English literature syllabi from the Open Syllabus Project, which draws on syllabi from many countries but predominantly from the U.S. 32 More than a third of the Goodreads classics authors are speci cally recommended by the AP English program, and about half rank within the top 200 most assigned collegelevel authors.
Yet the Goodreads classics depart from these school-sanctioned lists in two particularly striking ways. First, the Goodreads classics are considerably less diverse in terms of the race and ethnicity of their authors. Race is extremely complex and dif cult to reduce to data, especially because racial categories differ across different societies. However, if we acknowledge this reduction and use racial categories from the U.S. to re ect the perspective of the majority of Goodreads users, 33 then almost 94% of the Goodreads classics authors are white, which makes them whiter than both the AP recommended authors (70%) and the Open Syllabus authors (86% This lack of racial and geographic diversity in the Goodreads classics is not entirely surprising when one considers the user demographics of Goodreads. Beyond the platform's U.S.-centrism, the racial demographics of its user base skew overwhelmingly white-at least according to Quantcast, one of the ad industry's leaders for measuring online traf c and user demographics. As of June 2020, according to Quantcast, Goodreads users were 77% Caucasian, 9% Hispanic, 7% African American, 6% Asian, and 1% other. 34 It is crucial to note, however, that Quantcast uses statistical modeling techniques to predict demographic characteristics such as gender, age, ethnicity, and income, and, as sociologist Ruha Benjamin argues, companies that "create racialethnic data to be sold to others" deserve intense scrutiny. 35 Quantcast data is nevertheless used by many companies, including Goodreads, which makes it important to consider. 36 With these purported user demographics in mind, the predominantly white Western makeup of these reader-produced classics is not shocking but it is nevertheless startling, and it cautions any outsized faith in crowdsourced technologies as necessarily or predictably democratizing tools.

The Goodreads Algorithmic Echo Chamber
Goodreads users have not, on the whole, disrupted or remade the traditional canon of classics in any clearly radical ways via their crowdsourced shelving practicessave perhaps for the incorporation of genre ction. From the perspective of race and ethnicity, Goodreads users in fact seem to be reinforcing an even whiter and less diverse canon of classics than one would nd in a typical high school or college classroom today. By analyzing Goodreads reviews in addition to shelf classi cations, we hoped to better understand the forces and in uences shaping this perception of the "classics"who and what "is responsible for maintaining them in their preeminent position," as Jane Tompkins once put it. 38 When we turned to collect and analyze Goodreads users' reviews, we recognized one clear answer: Goodreads and Amazon. In this section, we brie y discuss the challenges that we faced while collecting Goodreads reviews, which we hope will be informative for others who wish to work with Goodreads reviews in the future. But more importantly these challenges reveal key insights about Goodreads/Amazon's proprietary algorithms and management of user data.
The rst key insight is that Goodreads purposely conceals and obfuscates its data from the public. The company does not provide programmatic (API) access to the full text of its reviews, as some websites and social media platforms do. To collect reviews, we thus needed to use a technique called "web scraping," where one extracts data from the web, speci cally from the part of a web page that users can see, as opposed to retrieving it from an internal source. 39 The Goodreads web interface makes it dif cult to scrape large amounts of review data, however. It's not just dif cult for researchers to collect Goodreads reviews. It's dif cult for anyone to interact with Goodreads reviews. Though more than 90 million reviews have been published on Goodreads in the site's history, one can only view 300 reviews for any given book in any given sort setting, a restriction that was implemented in 2016. Previously, Goodreads users could read through thousands of reviews for any given book. Because there are a handful of ways to sort Goodreads reviews (e.g., by publication date or by language), it is technically possible to read through 300 reviews in each of these sort settings. But even when accounting for all possible sort setting permutations, the number of visible and accessible Goodreads reviews is still only a tiny fraction of total Goodreads reviews. This throttling has been a source of frustration both for Goodreads users and for researchers. Working within these constraints, we collected approximately 900 unique reviews for each classic book-300 default sorted reviews, 300 newest reviews, and 300 oldest reviews-for a total of 127,855 Goodreads reviews. We collected these reviews regardless of whether the user explicitly shelved the book as a "classic" or not. We also explicitly ltered for English language reviews. Despite this ltering, a small number of non-English and multi-language reviews are included in the dataset, and they show up as outliers in some of our later results. Compared to the archives of most readership and reception studies, this dataset is large and presents exciting possibilities for studying reception at scale. But it is important to note that this dataset is not large or random enough to be a statistically representative sample of the "true" distribution of classics reviews on Goodreads. We believe our results provide valuable insight into Goodreads and the classics nonetheless. and 2019 re ect the fact that, in addition to collecting default-sored reviews, we speci cally collected the "oldest" reviews, most of which were published in 2007, and the "newest" reviews, most of which were published in 2019. Though the constraints of the Goodreads platform distort our dataset in certain ways, we tried to use this distortion to better scrutinize the in uence of the web interface on Goodreads users. For example, the company never makes clear how it sorts reviews by default, but we found that reviews with a combination of more likes and more comments almost always appear above those with fewerexcept in certain cases when there is, perhaps, another invisible social engagement metric such as the number of clicks, views, or shares that a review has received. Since we collected data in multiple sort settings, we are able to go further than this basic observation and investigate how exactly this default sorting algorithm shapes Goodreads users' behavior, social interactions, and perceptions of the classics. Based on our analysis, we found that the rst 300 default visible reviews for any given book develop into an echo chamber. Once a Goodreads review appears in the default sorting, in other words, it is more likely to be liked and commented on, and more likely to stay there ( gure 6). Meanwhile the majority of reviews quickly age beyond "newest" status and become hidden from public view. These liking patterns reveal that Goodreads users reinforce certain kinds of reviews, such as longer reviews ( gure 7), reviews that include a "spoiler alert" ( gure 9), and reviews written by a small set of Goodreads users who likely have many followers ( Table 2). If a review is prominently displayed by the default sorting algorithm, its author may be more likely to go back and modify this review. More default-sorted reviews included the words "update" or "updated" than oldest or newest reviews ( gure 8). In one especially interesting updated review, a Goodreads user raised her rating of Toni Morrison's The Bluest Eye and apologized for the way that her original, more negative review offended others and re ected her white privilege, which other Goodreads users had pointed out.

Topic Modeling Goodreads Reviews
Looking at the list of most popular Goodreads classics and analyzing liking patterns can only tell us so much about how Goodreads users perceive, de ne, and discuss the classics.
To know more, we needed to listen to readers' own critical voices. To understand the most consistent conversations and overarching themes in Goodreads classics reviews, we analyzed the reviews with topic modeling, speci cally a latent Dirichlet allocation (LDA) topic model: an unsupervised machine learning algorithm that essentially tries to guess the main themes of a collection of texts. 40 We pre-processed our reviews with Laure Thompson and David Mimno's "Authorless Topic Model" package to capture the most cross-cutting themes. 41 This package helps to remedy a common problem that occurs when topic modeling a collection of texts by multiple authorsor, in our case, a collection of reviews about texts by multiple authors -which is that the resulting topics often pick up on language speci c to individual authors, such as words unique to Shakespeare plays or to Jane Austen novels. Author-speci c topics can be desirable in some cases, but we wanted to reduce the signal of individual authors in order to amplify readers' collective voices across the reviews. The nal 30 topics produced by the topic model help us pull out some of the major threads in the Goodreads classics reviews, which we manually labeled and split into four categories: "The Classics Industry," "Literary Themes," "Literary Qualities," and "Linguistic Styles." "The Classics Industry" includes topics such as "Adaptations & Audiobooks" and "Editions & Translations" ( gure 10). The "Literary Themes" and "Literary Qualities" categories point to thematic or stylistic elements that readers' commonly discuss in their reviews, including topics such as "War & Adventure" or "Length & Pace" ( gure 11, gure 12) Finally, the "Linguistic Styles" category captures both Goodreads users' writing styles and literary authors' writing styles, which commonly appear in the form of quotations. Sometimes the topics even pick up on a fascinating blend of readers' and au-thors' styles combined. For example, the "Conversational & Slangy" topic sometimes identi es the quoted voice of Holden Caul eld, The Catcher in the Rye's angsty protagonist, but other times it identi es Goodreads users writing in a satirical Holden Caul eld-style voice ( gure 13). Figure 10: These are ve of the 30 topics produced by our topic model (based on 120,000+ Goodreads reviews of "classic" texts), which we labeled The Classics Industry. The table displays our hand label for the topic; the most probable words for the topic; the texts that are most probable for the topic (when we aggregate all the reviews for that text); and a sample review that ranked highly for the topic, with top words bolded. For readability, we remove a set of common stopwords from the most probable words. Figure 11: These are 11 of the 30 topics produced by our topic model (based on 120,000+ Goodreads reviews of "classic" texts), which we labeled Literary Themes. The table displays our hand label for the topic; the most probable words for the topic; the texts that are most probable for the topic (when we aggregate all the reviews for that text); and a sample review that ranked highly for the topic, with top words bolded. For readability, we remove a set of common stopwords from the most probable words.
Before fully diving into these topics, we want to brie y elaborate on the topic model to clarify this method and provoke a thought experiment. How might Goodreads and Amazon be extracting value from this data using computational methods? By demonstrating the kinds of patterns that our topic model can detect, we might better understand what's happening in Amazon's "engine-room," as Simone Murray puts it. 42 Because the topic model algorithm is "unsupervised," we do not specify in advance which topics to look for, only the number of topics to return. The number of topics that we decided on was a signi cant and subjective decision. The topic model is not an objective magic wand but an interpretive tool. We chose 30 topics because we experimented with different numbers and ultimately found that 30 topics produced the most coherent and compelling results.
Each topic consists of all the words in every recorded Goodreads review, ranked by their likelihood of appearing in a Goodreads review assigned to a particular topic. The most probable words for each topic typically represent a common theme, discourse, or linguistic style across the Goodreads reviews, such as "women," "men," "woman," "would," and "society," the ve most probable words for the topic that we eventually hand labeled "Gender & Sexuality" (all topics were similarly hand labeled by us). These topic words may seem, at rst glance, simplistic (e.g., "men" and "women") or even arbitrary (e.g., "eyes," "upon," and "long"). Yet when we read through the individual Goodreads reviews that rank highly for each topic, we can start to understand their signi cance and critical utility. Simple words, it turns out, can help detect complex discussions of gender and race, and seemingly random groups of words can be the unexpected trademarks of particular linguistic styles. The topic containing the words "eyes," "upon," "long," "light," "man," "heart," and "world," for example, ranks highly in Goodreads reviews that include a quotation from the book being reviewed ( gure 13). These basic words indeed identify the presence of literary language in a Goodreads review with remarkable regularity and accuracy, even across a wide range of source textsfrom Fitzgerald's The Great Gatsby ("And as I sat there, brooding on the old unknown world, I thought of Gatsby's wonder when he rst picked out the green light") to Morrison's The Bluest Eye ("God was a nice old white man, with long white hair, owing white beard, and little blue eyes") to Shakespeare's Macbeth ("cleanse the stuffed bosom of that perilous stuff which weighs upon her heart"). These results bolster our con dence that the model is picking up on signi cant threads even when the assemblages of topic words do not seem immediately coherent. This ability to nd signi cant threads playing out in individual Goodreads reviews is one of the major assets of the topic model for humanistic interpretation. We use the topic model not only to identify broad patterns in the collection but also to draw speci c and noteworthy examples to the surface and to our critical attention. Figure 12: These are eight of the 30 topics produced by our topic model (based on 120,000+ Goodreads reviews of "classic" texts), which we labeled Literary Qualities. The table displays our hand label for the topic; the most probable words for the topic; the texts that are most probable for the topic (when we aggregate all the reviews for that text); and a sample review that ranked highly for the topic, with top words bolded. For readability, we remove a set of common stopwords from the most probable words. Figure 13: These are six of the 30 topics produced by our topic model (based on 120,000+ Goodreads reviews of "classic" texts), which we labeled Linguistic Styles and Non-English Reviews. The table displays our hand label for the topic; the most probable words for the topic; the texts that are most probable for the topic (when we aggregate all the reviews for that text); and a sample review that ranked highly for the topic, with top words bolded. For readability, we remove a set of common stopwords from the most probable words.
By aggregating all ~900 reviews for each classic book, we can also identify the topics most associated with every book and, conversely, the books most associated with every topic. The classics that rank highest for the topic we have labeled "Gender & Sexuality"which includes words like "women," "men," woman," and "society"are literary works that explore subjects related to women's writing, feminism, misogyny, reproductive rights, and lesbian desire: Virginia Woolf's "A Room of One's Own" (1929), Charlotte Perkins Gilman's "The Yellow Wallpaper" (1892), Sylvia Plath's The Bell Jar (1963), Alice Walker's The Color Purple (1982), and Margaret Atwood's The Handmaid's Tale (1984) ( gure 11). The classics that rank highest for the topic we have labeled "Race"which includes words like "white," "black," "society," and "racism"revolve around issues such as  (1902) ( gure 11). These coherent clusters of literary works, grouped from within the broader 144 classic titles, are surprisingly intuitive classi cations for an unsupervised algorithm trained on readers' responses alone, with no access to the texts themselves or to any external metadata about author, publication, or reception. Further, these clusters paint impressionistic pictures of the collective reader response to each book. For the hand-selected group of texts in gure 14, we can see which books generated more discussion of classrooms and school and which books generated more discussion of life and death, which books were more likely to be quoted from and which books were more likely to inspire gushing declarations of love. By incorporating rating information, we can also identify which topics corresponded to more positive ratings, like "Beautiful Writing," and which to more negative ratings, like "Unlikeable Characters" ( gure 15). Using computational methods on Goodreads data, it is thus possible to learn a lot of information about readersthe kind of information that is ironically valuable both to literary critics and to corporations like Amazon. Figure 14: This heatmap represents the probability that Goodreads reviews for a given book would contain one of the 30 topics on the left. It can also be explored as an interactive data visualization. Darker tiles indicate a higher probability of containing the topic. Scanning left-to-right for the "School" topic, for example, reveals that To Kill a Mocking Bird, The Great Gatsby, and The Catcher in the Rye have the darkest tiles in this row, which indicates that reviews of these books are most likely to discuss school-related subjects. Scanning top-to-bottom for Pride and Prejudice, to take another example, reveals darker tiles for the topics "Audiobooks & Adaptations," "Marriage," "Re-Readable," and "Gushing & Loving Language." The heatmap rows have been normalized to highlight differences between the books. We check the signi cance of these results via 95% bootstrapped con dence intervals, and the majority of visible differences are signi cant. Figure 15: This gure shows whether Goodreads users were more likely to rate books positively (4-5 stars) or negatively (1-3 stars) when their reviews were likely to contain a certain topic. When Goodreads users published reviews that were likely to contain the "Unlikeable Characters" topic, for example, they tended to rate the text in question negatively.
Perhaps counterintuitively, when Goodreads users published reviews likely to contain the "Enjoyable & Interesting" topic, they were also more likely to rate the text negatively, because reviewers often discussed not enjoying a book and not nding it interesting. These results are based on the full set of Goodreads reviewsall books in all three sort orderings. The error bars indicate the standard deviation across 20 bootstrapped samples of the books, providing a measure of instability when a particular book is included or excluded in the dataset.

The Classics Industry
The rest of this essay focuses on the category that we have labeled "The Classics Industry," the set of topics that help point to some of the institutions and phenomena most responsible for reinforcing the classics in the twenty-rst century. This formulation is partly inspired by Murray's sociological account of the "adaptation industry," in which she maps "the industrial structures, interdependent networks of agents, commercial contexts, and legal and policy regimes within which adaptations come to be," mostly focusing on bookto-screen adaptations. 43 Though Goodreads users often allude to the academy and professional literary critics in their reviews, the prevalence of the term "classic" itself points to the shaping in uence of forces beyond the academy. To put this prevalence in concrete numbers, more than 15,000 Goodreads reviews explicitly mentioned the words "classic" or "classics," while just under 400 reviews mentioned the words "canon" or "canonical." This simple metric reveals a clear fault line in literary critical discourse between scholars and readers. It also indexes the power of the classics as a marketing brand. We detail how this brand functions in the sections below, and we also call attention to the ways that Amazon speci cally in uences and pro ts from this branding. Figure 16: This gure displays the number of Goodreads reviews that explicitly mentioned "classic" or "classics" vs.

The Classics Industry: School
Though Goodreads users rarely discuss the canon and scholars today rarely discuss the classics, the academy remains an important engine for the classics industry. The topic that we have labeled "School, "which includes words like "school," "high," "time," "class," " rst," "remember," "years," and "english," identi es the clear in uence of school systems on Goodreads users' conceptions of the classics, aligning with theories of cultural production and canon formation proposed by scholars like John Guillory and Pierre Bourdieu. 44 The Goodreads reviews that rank highly for this topic reveal a few key patterns. While some Goodreads users talk about recent experiences in English literature classes, many more discuss literature classroom experiences from the past or refer to more generalized conceptions of "required reading." "This was the rst Toni Morrison I read for 10th grade English while I was in high school," one Goodreads user re ected about Morrison's The Bluest Eye (1970), which she shelved under "classics." "I couldn't get into [it] at the timeand I think a good chunk of that had to do with how the story and it[s] dif cult subjects were handled in a classroom setting. Now that I can say I've read it again for Book Riot's 2018 Read Harder Challenge (an assigned book you hated or never nished), I could denitely appreciate it more." 45 When users catalogue their reading histories, high school and college reading often gures as an essential part of a fully comprehensive account.
Classics consumed from one's school days serve as something like a starter pack for a Goodreads catalogue, providing an easy way to rate and review a number of books immediately. Even Goodreads users who have wildly disparate genre inclinations will likely share these schoolbooks in common if they share common backgrounds. Because of these common shared experiences, schoolbooks foster social interactions between users, and communities commonly form around and through themwhether to read a classic for the rst time or to reread a previously hated classic á la Book Riot's Read Harder Challenge. Popular conceptions of school syllabi and required reading shape readers' habits long after their school days, and readers even self-assign books in order to belatedly join these communities. "Somehow I was never assigned to read this in high school, so I'm reading it now!" Goodreads user Edward Rathke exclaimed about The Grapes of Wrath. 46 "I had been planning to read '1984' for a long time," explained another Goodreads user named Andrew. "It's one of those books that you are supposed to read in high school. My high school AP Lit teacher had us read Aldous Huxley's 'Brave New World' instead." 47 These reviews may also explain why the Goodreads classics are less racially diverse than contemporary literary syllabi, since readers are clearly in uenced by historical and imagined literary curricula more than contemporary literary curricula.

The Classics Industry: Publishing
School syllabi feed the classics industry. They are undoubtedly one of the reasons, if not the primary reason, that the classics are a prominent advertising target value on Goodreads. But they also feed another major node in the classics industry: the publishing industry. High school and college syllabi, as Rebecca Rego Barry writes, "are pro table to the classics publisher because they have a known market. These titles are thus doubly promoted for entrance to the canon, in classrooms and bookstores, and it is interesting to note that professors and publishers are symbiotic in this respect." 48 The topic that we labeled "Editions & Translations," which includes words like "translation," "edition," "original," and "version," picks up on discussions about which edition or version of a particular classic Goodreads users have read, purchased, or borrowed. These are the Penguin Classics, the Signet Classics, and the Modern Library Classics, which make up a signi cant part of the literary market. "The classics market is huge," The Guardian reported in 2016. "There's been a noticeable upswing in the number of publishers doing the classics." 49 Though comprehensive book sales data is hard to come by, according to Publishers Weekly and NPD BookScan, the "classics" sold almost 3.6 million units in the rst half of 2018 making it the fth-best selling literary category behind "General Fiction," "Suspense/Thrillers," "Romance," and "Mystery/Detective," and ahead of genre ction heavyweights like "Fantasy" and "Science Fiction." 50 Even Amazon has now developed its own line of classics: AmazonClassics. In fact, almost every Goodreads classic currently in the public domain now has an AmazonClassics Kindle e-book for sale. As the series title AmazonClassics con rms, the publishing industry is one of the major forces that contributes to the gulf between "classic" and "canon" in readers' critical vocabularies. 51

The Classics Industry: Adaptation & Audiobooks
Various classics editions from disparate publishers reestablish the classics in concert. They are solidi ed, as Barry puts it, not through any one edition but "through the continuous promotion of a given title in more than one imprint, certi ed by more than one set of arbiters over a longer period of time." 52 This "continuous promotion" is not limited to print publishing. The proliferation of literary texts into other mediums further reinscribes certain books as classics, as Sarah Cardwell argues and as our analysis con rms. 53 The topic that we have labeled "Audiobooks & Adaptations," which includes words like "movie," "audio," "version," "seen," and "listened," captures how Goodreads users' sense of the classics is shaped by adaptations. The cluster of classics that rank highest for this topic -Truman Capote's Breakfast at Tiffany's, L. Frank Baum's The Wonderful Wizard of Oz, Charles Dickens's A Christmas Carol, and Stephen King's The Shiningall have major decades-old Hollywood lm adaptations. But many of the high-ranking Goodreads reviews in this topic also discuss audiobooks, which share a surprisingly strong relationship with lm and television adaptations and with Amazon. Audible, the world's largest producer of audiobooks, is yet another subsidiary of Amazon. In the last 10 years, Audible has invested in a series of classic literature audiobooks, "Audible Signature Classics," narrated by famous lm and television actors. Most of these classics are the same popular Goodreads classics that we have already identi ed, paired with a performance by a high pro le celebrity: The Great Gatsby narrated by Jake Gyllenhaal (2013) 54 This catalogue represents what Cardwell refers to as "circular af rmation," when a certain selection of books are reinforced as classics by being adapted and con rmed "across various areas of the public sphere"con rmed, in this case, not only through audiobook adaptation but also through association with Hollywood celebrities. 55 Based on our collection of Goodreads reviews, we nd that this circle of af rmation sometimes marginalizes the print text itself. For example, in one review of Truman Capote's Breakfast at Tiffany's, a Goodreads user named Jennifer Masterson shelved the novella under "classics" and gushed: 3 delicious hours of audio read by Mr. Michael C. Hall aka Dexter!!! What a wonderful performance of Truman Capote's novella! I saw the movie years ago but I've never read the book! I'm so happy to have listened to this edition of the audio! 5+++++Stars for the narrator! 5 Stars for the story! Highly highly recommended!!! 56 This review of an Audible Original audiobook, narrated by a television star, Dexter's Michael C. Hall, was inspired by watching a Hollywood lm. And though this Hollywood lm was originally based on a novella, Jennifer, this particular Goodreads user, never read the novella and did not need to in order to review the book on Goodreads and perpetuate the classics industry. This review also demonstrates that Amazon-af liated audiobooks inspire users to visit and rate books on Goodreads, to bounce from one Amazon subsidiary to another. We speculate that Amazon may also use Goodreads data to help determine which audiobooks, television shows, and lms to invest in. One of the earliest Amazon Studios television series was an adaptation of Philip K. Dick's The Man in the High Castle (1962), a popular Goodreads classic, and one of Amazon Studios' biggest investments is a television series based on the Goodreads classics with the highest average ratings in our dataset, J.R.R. Tolkien's The Lord of the Rings trilogy, the rights for which Amazon purchased for $200 million. 57 We are not claiming that Goodreads reviews and ratings directly motivated this decision. But it is important to recognize that Goodreads data is controlled by Amazon, a corporation that is making some of the most expensive and high pro le literary investments of our time.

The Classics Industry: Goodreads Users
The classics are clearly perpetuated by many powerful institutions as well as the market economy. When Goodreads users shelve, rate, and review classics, they contribute to this system and help sustain it. Making this point forcefully, Murray argues: The [Goodreads] website's beguiling abundance of actual reader responses to books has obscured for scholars the limited extent to which users either understand or can in uence its algorithmic operations, leading to overblown claims of readerly empowerment. Compelling evidence of reading's contemporary resilience and freely available research archive though it may be, Goodreads is above all else a node in platform capitalism. 58 Goodreads is indeed "a node in platform capitalism," but we believe it is important to engage with how "beguiling" Goodreads reviews are and how empowering the platform can feel for some Goodreads users. In Aarthi Vadde's study of "amateur creativity" on the internet, she argues that it is not possible to "make a blanket case for or against the emancipatory potential of participatory culture on the Internet." 59 Instead Vadde suggests thinking of the public sphere as "an always already commercialized, industrialized, and pluralized space." 60 We believe this framing is helpful for teasing out how Goodreads users sometimes explicitly resist Goodreads and produce remarkably interesting amateur criticism all while being exploited by Goodreads.
One of the most tting metaphorical representations of this ironic tension manifests when Goodreads users bash the classics, because, in doing so, they simultaneously reject and reinforce books as classics in the same stroke. The topic that we have labeled "Goodreads User Criticism," which includes words like "stars," "give," and "rating," picks up on a common rhetorical tropethe justi cation of a user's rating for a given textand it includes a signi cant amount of classics bashing. We nd that Goodreads reviews that rank highly for this topic are, overall, more likely to rate a text negatively ( gure 15). Negative ratings seem to demand lengthy, re exive justi cations in their accompanying reviews. For example, a Goodreads user named Bren, mentioned in the introduction of this essay, shelved Nabokov's Lolita as a classic but rated the novel only three out of ve stars. Though three stars was already a low rating (particularly within the Goodreads community), she later returned to the review and lowered the rating still further. In her updated review, Bren explained that she originally gave Lolita a higher rating "in deference of its classic status." 61 But as she watched other Goodreads users openly panning books, including Lolita, she gained new con dence to dissent from Lolita's perceived reputation and from its imagined community of fans, whom she dubbed "book snobs." This retroactive rating is a triumphant moment that Bren jokingly compares to winning an Academy Award: This review is inspired by some of my GR [Goodreads] friends whose fearlessness about giving low stars to books they do not like has inspired me to change my rating of Lolita from three stars to two stars as that is what I really feel . . . I get that this a classic and book snobs who read this will sig[h] in indignation but I do not care. I just did not get it and still don't. I'd like to thank anti book snobs everywhere for giving me the courage to rate Lolita two stars. I will never forget you. Wow..is this what an Oscar speech feels like? 62 Many Goodreads users like Bren seem to feel liberated when they reject the classics and express honest negative opinions about exalted books. When we reached out to Bren to seek her permission to publish this review, she further elaborated about what the Goodreads community means to her and even alluded to its special signi cance during the COVID-19 pandemic: "There is something about speaking against a Classic that can be very intimidating. People on here are fearless and, at least for me, I never feel judged . . . When I rst joined I was too shy to talk to people but years later, I have connected with wonderful people and it has become a wonderful source of comfort to me, especially in trying times like these." 63 For Bren, the Goodreads community is sincerely meaningful, and the ability to speak out against a classic is genuinely empowering.
Another Goodreads user, Peter Derk, re ected about the joys of publishing "really nasty review[s]" of the classics, but his joy, unlike Bren's, was premised on the perceived powerlessness of his Goodreads review in the face of a classic: Every so often I'll get into a classic. I guess because I feel like writing a really nasty review. Classics are great fodder for nasty reviews because 1. The people who made them are LONG dead . . . Saying bad stuff about a classic novel doesn't hurt the creator's feelings . . . 2. Classics have such a pedestal in the literary world already that the opinion of one lone weirdo . . . is pretty irrelevant. It's not like bashing on this book is suddenly going to render it a Not A Classic or affect its sales. Frankly, I think that about everything I read, but with classics, it's a pretty rock solid premise. 64 Rather than an emboldened community taking on Lolita's classic reputation, as Bren framed herself and her "GR friends," Derk describes himself as "one lone weirdo" who couldn't possibly make a dent in a classic book's reputation. Far from being able to hurt a classic's sales, as Derk acknowledges, his colorful, vehement 2000-word takedown of The Phantom of the Opera likely only contributes to its contemporary value by contributing to its continued discussion. This paradox is one of the reasons that the classics remain so powerful. Love them or hate them, the classics sustain themselves by staying in print, remaining a topic of conversation, and enduring as a commodity.

Conclusion
So what is a "classic" in the twenty-rst century? Based on our analysis of 144 Goodreads classics and 120,000 accompanying reviews, there are at least a few clear answers. For Goodreads and Amazon, a classic is a prominent advertising target value, a marketing tool, and a source of lucrative adaptation material. For Goodreads users, a classic is a book read in high school, a book that inspired a TV show, or a book that other Goodreads users have tagged as a classic. As we have shown, the classics industrythe collaborative forces of publishing, lm, television, Amazon, and morede nes the status of popular classics to a large extent. Yet for Goodreads users, a classic is also an invitation to become amateur critics and creative writers, a chance to re ect on their lives and relationships to power, a conduit for connecting to others, and an opportunity to enter a critical conversation that has long excluded them. Literary history lives both pro tably and vibrantly in the world under the moniker of the classics. To recognize the signi cance of the term is to recognize some of the places where literary criticism is most alive, relevant, and valuable.
Beyond the classics, this essay also points to major trends in contemporary literary culture that pose data-related challenges for literary criticstrends such as the rise of reader social networks, online amateur criticism, and Amazon. We believe that computational methods like the ones used in this essay can play a signi cant role in facing these challenges. When combined, computational methods and internet data can help literary critics simultaneously capture the creative explosion of reader responses as well as critique algorithmic culture.