Abstract

• 1. Introduction
• 2. Goodreads as a proxy for ›quality‹
• 3. Data
  • 3.1 The Chicago Corpus
  • 3.2 The Goodreads ›mediocre‹ subset
  • 3.3 Other quality proxies
• 4. Method
• 5. Results
  • 5.1 Rating count of the quality / non-quality groups
  • 5.2 Comparison of rating distribution polarisation per group
  • 5.3 Quality titles at high and low rating count
  • 5.4 Quality defined on a title- vs. author-base
• 6. Conclusion & future works
• Bibliography
• List of Figures and Tables

1. Introduction

[1]The fields of computational linguistics and computational literary studies have in recent years converged on the development of a host of powerful new methods and tools for analysing text on a large scale. While computational linguistics and Natural Language Processing (NLP) research has increasingly researched literary corpora and modelled the way texts impact readers, computational studies of literature have embraced computational linguistics and NLP methods, allowing for the growth of more complex quantitative literary research. Since literary studies deal with social and cultural elements inherent to literary, reading, and book history as well as canon-formation, quantitative approaches have also borrowed systems from computational sociology, statistical and social science methods, enabling literary research on a scale that would be unfeasible ›by hand‹.

[2]Beyond upping the scale, computational methods have also made a shift in the object of literary studies research: from studies in literary history predominantly tracing the trajectories and influences of canonical works and authors, toward bringing into focus what Moretti has called ›the great unread‹: the forgotten 99 % of literary history.‍[1] Since no one can access (let alone process) everything that has ever been published, quantitative literary studies apply theoretical frameworks of corpus linguistics to try to create limited collections that are representative, i. e., reasonably mirror the whole population on a smaller scale. In traditional corpus linguistics, for example, a large diachronic collection of documents from different written and spoken domains could be considered a reasonable representation of a language in a given time span, so that the spreading of a new term at a given point might be reasonably considered to mirror the spreading of that word in the real world.

[3]A similar approach may be taken to the study of literature, where one may track the emergence and spread of concepts or stylistic devices across corpora,‍[2] and even study the popularity of a given work with readers, comparing its reception to that of other works in the corpus, extrapolating and correlating popularity to, for example, particular textual features. Still, representativeness is paramount: if that corpus consists only of bestselling sci-fi, we have gained insight only into what features contribute to popularity in this given niche population.

[4]While the ability to process and analyse large quantities of literary texts and perform complex statistical experiments on them has recently made new ways of studying literary appreciation possible, the question of how to assess (and recreate) literary success is probably as old as narrative itself. The question of quality has naturally been prominent in literary criticism, but its significance has often been eclipsed in the scholarly discussion. Disciplinary shifts, such as the debate on canon representativity and exclusion,‍[3] methodological shifts, such as moving the focus from evaluation towards interpretation in 20th century literary criticism,‍[4] as well as a modern expansion of literature’s conceptual boundaries to encompass texts with new experimental forms (i. a., hypertext fiction), or such that are ideologically opposed to the notion of ›pleasing‹ the reader,‍[5] have all played a role in making words like ›literary quality‹, or ›classics‹ appear to belong to the »precritical era of criticism itself«.‍[6] Nevertheless, literary cultures continue to establish and uphold estimations of literary excellence in practice, such as through literary awards, classics book series, or anthologies. It seems a disparity has arisen between a more modern literary studies inattention to, or »denial of quality«,‍[7] and the multitude of literary judgements prevailing within literary cultures – a disparity that is brought to a point in the context of more recent computational literary studies which show a consensus on quality judgements among readers at the scale of large numbers.‍[8]

[5]Most studies of literary success or quality have attempted to test links between reader appreciation and textual features: both stylistic features,‍[9] and more narrative characteristics, such as shape and dynamic of narratives’ sentiment arcs extracted by sentiment analysis.‍[10] In one sense, traditional and computational literary studies have similar tendencies: like traditional literary scholarship, computational studies of literature often focus on ›the excellent few‹ – the ›exceptional books‹ that do well among readers‍[11] – and seek to understand their exceptionality on the textual level.

[6]Computational studies that seek to predict reader appreciation often set up the task as one of classification, dividing their corpus into ›successful‹ and ›unsuccessful‹ titles, using a threshold value of some proxy of reader appreciation, like the number of downloads on project Gutenberg.‍[12] Others, who seek to model the continuous scale of ›more or less successful‹ titles, often focus on discussing and contrasting the textual characteristics of the ›very good‹ against the ›very bad‹.‍[13] However, few studies take up the discussion of what is often a very large part of any corpus: the mediocrely estimated titles, neither decidedly good nor bad. This group is also important when seeking to assess the different ways literary cultures judge works. We would assume that the titles that are considered ›great‹ or ›bad‹ vary across proxies of literary quality. For example, the titles on bestseller lists are not necessarily those most often assigned in college syllabi. It should therefore not be a surprise if what is considered ›mediocre‹ varies greatly across different standards and tastes. As such, what is considered ›mediocre‹ may vary depending on what ›quality proxy‹ studies use to approximate the notion of ›literary success‹ or ›quality‹.

[7]In computational literary studies, a ›proxy‹ serves as a formal method for approximating abstract constructs or concepts through operationalization. Proxies bridge qualitative interpretation with quantitative methodologies, and are a translation of constructs, like a literary device, or concepts, like literary quality, into measurable variables. When speaking of a ›quality proxy‹, we mean a specific operationalization of quality or reader appreciation among many. For example, one might differentiate between literary ›fame‹ and ›popularity‹, since fame, such as that of James Joyce’s Ulysses does not necessarily mean that the book is widely read. These two different forms of quality may be measured in dissimilar ways – i. e., through different ›proxies‹ – for example by looking at how often a book is subject of literary scholarship, vs. how many copies it sells, or how often it is rated on Goodreads.‍[14] Consequently, the construction of proxies also assists in compelling us to precisely define more ambiguous concepts like ›quality‹.

[8]For studies seeking to assess factors contributing to reader appreciation, selecting a resource to use as such a proxy for reader appreciation or literary success is one of the great challenges, if not the greatest. Studies often use a single proxy for literary quality as a golden standard, which may not adequately capture the diverse – and, sometimes, opposed – preferences of various types of audiences.‍[15] Among various proxies for quality, such as a text’s presence in established literary canons‍[16], whether or not it was longlisted for awards,‍[17] sales numbers,‍[18] ratings on the large online platform Goodreads are widely used.‍[19] Specifically, the average ratings or stars of Goodreads are the most widely used, although Goodreads actually has two dimensions: stars representing various types of literary evaluations reduced to the 1–5 point ›stars‹ scale, and the rating count, representing how often a title is rated, indicating its circulation or fame among platform users.

[9]In the present paper we suggest that a closer inspection of ›mediocrely‹ or ›indecisively‹ evaluated texts may help studies to better assess the corpus, as well as the proxies of reader appreciation themselves, as it allows for a better understanding of the judgement behaviour and nature of the resource. Since it is the most widely used resource, we examine Goodreads ratings and rating counts as a proxy for literary ›mediocrity‹, specifically taking the ›mediocre‹ Goodreads’ ratings as an ideal subset for exploring the evaluative process within the proxy. Our question pertains to the nature of Goodreads mediocre ratings: do they signify a simply ›lukewarm‹ or tepid reception from readers, or are they the result of a readership divided into polarised opinions about specific titles? We describe this ›mediocrely rated‹ Goodreads subset in terms of how the titles are evaluated outside of Goodreads, and examine rating behaviour, specifically the distribution of raters’ evaluations on the 1–5 point scale, to explore the distinct types of literary ›mediocrity‹ formed on the Goodreads platform.

2. Goodreads as a proxy for ›quality‹

[10]Goodreads is a popular online social platform for readers that allows users, among other things, to comment, recommend, and review a book. It is, according to Nakamura (2013), a ›social cataloguing site‹, which links to other social networks (Facebook, Twitter / X, Instagram, and LinkedIn), and where social networking may, for some users, be just as important as book cataloguing and reviewing.‍[20] With its more than 90 million users, Goodreads arguably offers an insight into reading culture ›in the wild‹.‍[21] It catalogues books from a wide spectrum of genres and derives book-ratings from a heterogeneous pool of readers in terms of background, gender, age, native language and reading preferences.‍[22] Whether it may be followed that sites like Goodreads represent a ›democratisation‹ of literary criticism may, however, be questioned, not least because we see the continuation of established patterns on the platform. For example, works that are often assigned on college syllabi are also perceived as ›canonic‹ or ›classic‹ on Goodreads.‍[23] Moreover, while the site had mostly anglophone users at the outset – which was presumably reflected in the type of books rated (and highly rated) – the rapid expansion of the site makes it difficult to gauge user demographics, though it structurally must be presumed to reflect certain tilts, i. e., preferences of readers that review books online vs. those that do not.

[11]Moreover, problems and pitfalls with Goodreads data should be acknowledged, such as the existence of fake or sponsored reviewing, influences of Goodreads algorithms on users’ reading and reviewing behaviour, and the general anglophone and Western literary tilt and biases of the platform.‍[24] Still, some aspects of Goodreads reviewer behaviour support the idea that Goodreads represents an alternative resource for studying literary evaluation next to newspaper and scholarly literary criticism. For example, Verboord (2011) has shown that Goodreads reviews distribute more equally across genres and exhibit more attention to mass-market paperbacks and titles by female authors than newspaper reviews.‍[25] While Goodreads ratings and rating count do not present an absolute measure of literary popularity or quality – or in the context of the present study, literary ›mediocrity‹ – they do offer a valuable perspective on a title’s overall reception among a diverse population of readers with preferences that vary from expert and newspaper critics. In other words, we should remember that Goodreads reflects preferences of a certain audience however heterogenous the user demographic – it is one ›quality proxy‹ among many.

[12]Despite their diversity, different proxies of quality may display significant overlaps.‍[26] Moreover, if a book comes to be included in one proxy, this might correlate to an increase in another: for example, studies have shown that winning an award may mean that a title gains in scholarly prestige or popularity.‍[27] Similarly, Kovács and Sharkey (2014) found that winning an award corresponds to an initial increase in book ratings on Goodreads, however, a winning books’ average rating tends to decline relative to books that were long-listed for an award but did not win.‍[28] They suggest that the decline in average rating is partly an effect of higher popularity, whereby a title also attracts readers that are predisposed to dislike it. We may presume then that a higher number of ratings could correlate with a more polarised rating behaviour (i. e., a higher number of very low ratings).

[13]Goodreads’ average ratings represent the averaged ratings of all users for a title on a Likert scale implemented as the number of stars on the platform interface, ranging from 1 (low appreciation) to 5 stars (high appreciation). While the average score provides a general indication of a title’s reception, it is problematic because it conflates types of literary appreciation, i. a., satisfaction, enjoyment, and evaluation to one scale, forcing users to cast their literary evaluations to a mono-dimensional scale. Moreover, averaging Likert scale data entails a conversion of ordinal to interval data, reducing uneven judgements to one seemingly unambiguous score, which is a well-known issue.‍[29] The averaged score potentially conflates genre-specific value judgements, which may obscure important differences in rating behaviour across audiences. For example, readers of sci-fi may be inclined to give a generally higher average rating on Goodreads, something that we do not take into account when using average rating as a quality metric. Additionally, the average score may be the result of numerous middle-ratings or many ratings at either of the extremes. The simplicity of the Goodreads resource is also why it is often employed in computational studies seeking to assess literary quality or success: it offers a streamlined approach to a problem that frequently proves too complex for quantitative analysis. While the Goodreads proxy thus has the benefit of aggregating readers’ opinions into one quantifiable scale, it also creates a grey area in the middle. Mediocre ratings on Goodreads may be averages of numerous extreme positive and negative ratings,‍[30] which poses challenges for researchers both when interpreting and trying to build systems to predict ratings, especially when they only have access to the average ratings without information on rating distribution.

3. Data

3.2 The Goodreads ›mediocre‹ subset

[17]To analyze ›mediocre‹ Goodreads ratings, we define a ›mediocre group‹ within the Chicago corpus as titles that constitute the middle 25 % based on average Goodreads ratings. In this context, ›mediocre‹ does not suggest poor quality or a title’s standing on the Goodreads platform, but rather reflects the average rating compared to the entire corpus. To distinguish the middle 25 % of our corpus, we distinguished titles that lie between the 37.5th and 62.5th percentile of the distribution of average Goodreads ratings in our corpus (see fig. 1). These percentiles divide the corpus distribution into parts, with 25 % of the corpus (the ›mediocre‹ group) occupying the middle segment (between the percentiles): 37.5 % of the corpus falls below it and 37.5 % falls above it. The middle segment comprises titles representing the middle 25 % of the corpus in terms of their Goodreads ratings. We thus compile a ›mediocre‹ subset using the titles representing the middle 25 % of the corpus in terms of their Goodreads ratings, which is equivalent to having a Goodreads’ average rating between 3.72 and 3.91. This ›mediocre group‹ comprises 2,150 titles in total.‍[32]

Figure 1: Histogram of Goodreads ratings of the Chicago Corpus, where the ›mediocre‹ subset is indicated. [Chart: Feldkamp et al. 2025]

4. Method

[20]As we seek to gain a better insight into types of titles, rater behaviour and judgements of the ›mediocre‹ Goodreads subset, we apply measures that can reasonably be expected to differentiate groups within the subset. These are, firstly, external proxies of literary quality which indicate whether a title is highly appreciated outside the Goodreads platform (table 2). Secondly, with the Goodreads data, we examine whether these ›quality titles‹ within the ›mediocre‹ subset differ from other titles in terms of rating count, examining whether the ›quality‹ titles are more popular (i. e., more frequently rated) in comparison to the ›non-quality‹ titles in our subset. Thirdly, we examine the distribution of ratings of the ›quality‹ titles in our subset, to assess whether some titles in the subset are subject to different rating behaviour and, specifically, whether titles that are ›quality‹ titles judged ›quality titles‹ by external proxies (outside of Goodreads) are subject to a polarised rating behaviour within the Goodreads platform itself.

[21]To assess the polarisation of the raters of individual titles, we calculated the standard deviation (SD) of their rating distribution, based on the assumption that fewer ratings around the mean indicates more polarized rater behaviour. Moreover, following Maity et al. (2019), we calculate the Shannon entropy of rating distributions as a proxy measure to assess the polarisation of raters.‍[34] If values are more ›polarized‹, they should also be more ›spread out‹ on the rating scale (1–5), which entropy should capture.

[22]To inspect how SD captures polarisation of reception of titles – essentially a ›sanity check‹ of our measures – we plotted the distribution of ratings of individual titles. For assessing and visualising, we keep a threshold of rating count (10,000). Firstly, because very low rating count may exhibit a ›polarised‹ readership because judgements have not converged at the larger scale or random patterns not corroborated by other raters. Secondly, to visualise more recognizable titles for the sake of our readers. Generally, we found both SD and entropy to distinguish the same titles at the highest and lowest end of the spectrum: the list and order of titles on the low and high end of SD are almost identical to those at the low and high end of rating entropy.

Figure 2: Titles among top 20 highest in SD to the left (SD descending) and titles among the 20 lowest in SD to the right (SD ascending). Note that titles to the right (blue) seem to have a slightly more ›normal‹ distribution, while rating distributions of titles to the left (pink) have more of a U-shape, with more low ratings (of 1 and 2) and very high ratings (of 5). [Chart: Feldkamp et al. 2025]

[23]From this inspection, it is clear that SD and entropy of rating distribution can adequately capture how controversial titles are: titles on the high end of SD and entropy (fig. 2, left side) are known controversial titles. Either because they are stylistically demanding (e. g., James Joyce’s Ulysses), or controversial in terms of the themes and worldview they treat (e. g., Vladimir Nabokov’s Lolita or Tim LaHaye and Jerry B. Jenkins’s Left Behind). Titles at the low end of these measures (fig. 2, right side) are more mainstream literature: widely read and not known to be very debated. In fact, 7 out of the 20 titles lowest in SD are novels by Agatha Christie (note that we are still looking at titles above the set threshold in rating count, i. e., above 10,000 ratings).

[24]Moreover, testing the collinearity of standard deviation and entropy by using a Spearman correlation, we found that these two measures are very highly correlated in our subset (coefficient = 0.9, p < .01). As such, we focus on reporting values for SD in the following, as values for entropy are very similar.

5. Results

5.1 Rating count of the quality / non-quality groups

[25]In comparing titles in the ›mediocre‹ subset that appear in other proxies of literary quality (i. e., one of the proxies in table 2), we find that there is a difference in the frequency that titles in these groups are rated (fig. 3, 4, see also table 1), even if there is no difference in average rating between the groups (fig. 3, table 1).

Figure 3: Histogram of Goodreads rating in ›mediocre‹ subset. [Chart: Feldkamp et al. 2025]

Figure 4: Histogram of Goodreads rating count in ›mediocre‹ subset. [Chart: Feldkamp et al. 2025]

Figure 5: Boxplot showing the rating count per group. [Chart: Feldkamp et al. 2025]

	Avg. Rating count	Std. deviation
All	14365.27	3121538.20
Subset	9638.1	163201.66
Q. titles	26767.76	110124.60
Non-q. titles	1680.0	4685.72

Table 3: Rating count and standard deviation for the whole corpus, the ›mediocre‹ subset, as well as the quality and non-quality groups.

[26]We find that the mean rating count of the ›quality group‹ in our subset is higher than the non-quality group, but it is also higher than the mean rating count in the whole corpus. Note that the standard deviation is, however, extremely high, indicating that the ›quality‹ group is by no means homogenous. This would suggest that the quality group contains several titles that are rated extremely often, even when compared to the ratings in the whole corpus. In this sense, several works that are considered of ›quality‹ outside of the platform Goodreads seem to garner more attention than those that are not.

5.3 Quality titles at high and low rating count

[30]In figure 2, we saw that standard deviation clearly indicates works in the quality group that have a polarised distribution. As in the case of rating count, however, standard deviation of rating distributions seems to be very heterogeneous within the quality group to the point where variance makes it difficult to distinguish from the non-quality group. Following the suggestion of Kovács and Sharkey (2014) that higher rating count is linked to higher polarisation of the raters,‍[37] we examine marginally more rated titles within the quality group itself. For this, we split the quality group into a group of titles with high and one with low rating count, selecting the subset mean rating count as a somewhat arbitrary threshold. The aim here is not to distinguish clear groups, but simply to examine the characteristics of some portion of quality titles that have a very high rating count.

Figure 6: Pieplot showing the sizes of groups below and above subset mean rating count (RC) within quality and non-quality titles. [Chart: Feldkamp et al. 2025]

[31]In figure 6, we see that a much larger portion of quality titles place above the subset mean compared to the non-quality group. Still, a considerable part of the ›quality‹ group – approximately ⅓ of the group – does not have a high rating count, indicating works of what we might call ›indecisive quality‹, That is, they are valued highly beyond Goodreads (contained in other quality proxies), but they do not seem to garner much attention on Goodreads, neither in terms of rating nor in terms of higher rating count.

[32]While there are no differences between the high RC and low RC ›quality‹ groups in terms of publication date or average rating, we find different types of quality proxies more prominent in one group compared to the other (fig. 7). It seems that more canonical works are more prevalent in the high rating count compared to the low rating count ›quality‹ group. Proxies that are institutionally oriented (OpenSyllabus, the Norton Anthology) as well as the GoodRead’s Classics list,‍[38] but also titles that won prestigious awards (NBA, the Nobel Prize) are more frequent in this group. Conversely, titles longlisted for genre-fiction awards appear more prevalent in the low rating count ›quality‹ group (fig. 7).

Figure 7: Barplot showing the number of titles from different proxies within the above and the below subset mean rating count (RC) groups. [Chart: Feldkamp et al. 2025]

[33]To further assess this low rating portion of the quality group, two literary scholars manually inspected the titles it contained. They found that many of the titles appear to be less appreciated (not the most prominent) works by rather famous or canonical authors. This includes, for example, The Little Lady of the Big House by Jack London, who is most prominently known for works like The Call of the Wild, Flush and The Years by Virginia Woolf, primarily known for her works Mrs. Dalloway or To the Lighthouse; The 42nd Parallel by John Dos Passos, primarily known for his Manhattan Transfer; Steps by Jerzy Kosinski, primarily known for his The Painted Bird, as well as, for example, Murphy by Samuel Beckett, primarily known for his theatre plays, and so on.

[34]Conversely, in the high rating quality group, we tend to see more typically canonical works by famous authors, such as Woolf’s To the Lighthouse and Mrs. Dalloway, or The Portrait of a Lady by Henry James, as well as works that are very known and sometimes controversial. The controversy of such titles may lie in their political aspects, like Ayn Rand’s The Fountainhead and Tim LaHaye and Jerry B. Jenkins’s Left Behind, or their dealing with controversial themes such as Vladimir Nabokov’s Lolita, but also in being stylistically experimental, such as Joyce’s Ulysses or William Faulkner’s The Sound and the Fury, or having a renowned difficult style, such as Malcolm Lowry’s Under the Volcano.

[35]Moreover, there appears to be a difference in the polarisation of rating distributions between the low and high rating count ›quality‹ groups (fig. 8), where we see a small successive rise from the mean standard deviation of the non-quality group, the quality group with low rating count, to the mean standard deviation of the quality group with higher rating count. By using a Spearman correlation, we find that there is a medium strength correlation between standard deviation of rating distribution in the quality group with high rating count (table 5).

5.4 Quality defined on a title- vs. author-base

[36]An important consideration following from our qualitative inspection of titles in these groups, is that our selection of proxies of literary quality may go some way in explaining why the low rating count ›quality‹ group may contain so many titles that are not very well known but are written by well-known authors. Particularly considering that some of our proxies are author-based, so that every title by an author mentioned in the proxy (e. g. a Nobel winning author) is included in the ›quality‹ group. This inclusion of titles that are not properly considered quality may also make it harder to distinguish between the quality and non-quality group. To test for this issue, we tried using exclusively title-based proxies of quality against using all proxies.

[37]While the difference between using title and author-based proxies is small (37 titles less), and the mean rating count of the quality group only goes up slightly, to a mean rating count of 27,723.82 compared to 26,767.76 when using author-based proxies, we seem to see a slightly larger difference when examining the polarisation of rating distributions, so that quality groups selected via title-based proxies show a stronger correlation than when using all proxies (table 5).

Figure 8: Boxplot showing the distribution of rating count values in the three groups. [Chart: Feldkamp et al. 2025]

	Q. titles > mean RC	Q. titles < mean RC	Q. titles
Authorbased proxies	0.30** (0.27**)	0.08 (0.06)	0.08* (0.07*)
Titlebased proxies	0.32** (0.29**)	0.09 (0.09)	0.09* (0.07*)

Table 5: Coefficients of Spearman correlations between rating count and standard deviation of rating distributions per group, when using all or exclusively title-based proxies (with entropy values in parentheses). *p < .05 **p < .01

[38]Combining qualitative and quantitative levels of analysis, we visualise the correlation, as well as placement of individual titles within the high rating count ›quality‹ group in fig. 9. Here we tend to see controversial titles like The Fountainhead, or stylistically controversial titles like Ulysses at the extreme end of standard deviation. Less polarised works, such as Cujo, Dolores Claiborne and Firestarter by the widely popular Stephen King appear at middle values of standard deviation, while works of a mass market type literature such as Agatha Christie’s The Body in the Library and The Mystery of the Blue Train, part two of Frank Herbert’s bestselling Science Fiction series, Dune Messiah, and Truman Capote’s popular classic Breakfast at Tiffany’s appear at the very lowest end of standard deviation (fig. 9).

Figure 9: Titles in the quality titles above subset mean rating count group according to their rating count and SD of rating distributions. [Chart: Feldkamp et al. 2025]

6. Conclusion & future works

[39]We have run an examination of a subset of contemporary English language novels falling into the ›least decisive‹ section of our corpus in terms of Goodreads’ average ratings, namely the 25 % of titles in our corpus that fall in the range of 3.72 to 3.91 in average rating. Intriguingly, this quartile range includes a notable proportion of works that have received accolades in alternative indices of literary merit, such as inclusion in canonical lists or receipt of esteemed literary awards. For brevity, we called this the ›quality group‹, as it represents several novels highly regarded by alternative evaluative lenses.

[40]This group has on average a higher rating count than its ›non-quality‹ counterpart. When looking at the titles in the quality group that are above the mean rating count of the whole subset, and that push the whole group’s average up, we found a weak-to-medium correlation between the rating count of the titles and the standard deviation of their ratings, as well as the entropy of the rating distribution. This pattern suggests that titles with higher readership counts exhibit greater polarisation in ratings, corroborating the tendency suggested by Kovács and Sharkey (2014).‍[39] In other words, there is a portion of titles in the ›quality group‹ whose ratings on Goodreads are also polarised, so that their mean rating count obscures the extreme evaluations that are their basis.

[41]Qualitative analysis of these polarising works revealed distinct characteristics. These titles are often notable for their stylistic experimentation (e. g. Joyce’s Ulysses), possibly leading to mixed feelings in readers unfamiliar with or appalled by their style, or controversial in terms of their message (Rand’s The Fountainhead), topic or point of view (Nabokov’s Lolita). It is noteworthy that this highly-read subset predominantly comprises works considered canonical, with limited representation of genre fiction (fig. 7). Contrastingly, an investigation into the novels within the ›quality group‹ that have lower than average rating counts yielded a cluster of works by renowned or canonical authors that did not gain extensive appraisal, distinguishing them from the more polarised, widely-read titles.

[42]In summary, we have observed that the most mediocre section of Goodreads rating in our corpus comprises at least three populations:

• Absolute Mediocrity. These are titles that the majority of readers on the platform rated as neither great nor terrible, and that were not picked by any of the other quality proxies we consulted for having outstanding merits. They were not ›good enough‹ to score higher than 3.91, nor ›bad enough‹ to score below 3.72 on Goodreads, though being included in more than a few libraries around the world (thus making the Chicago Corpus’ original selection). Among such works we find, for example, the forgotten classic Susy, a Story of the Plains by Bret Harte set in the California Gold Rush, or the dystopian novel by Paul Auster, In the Country of Last Things.
• Relative Mediocrity. These are titles that most readers on the platform rated as neither great nor terrible, but that were considered by other quality proxies as having outstanding merits. While these titles’ average score on Goodreads reflects an actual ›non-decisive‹ judgment from the reader’s majority, they appear to be polarizing when we consider a wider evaluative landscape in literary culture. An example in this subgroup is The 42nd Parallel by John Dos Passos, i. e., a book by a famous author who is most known for other works.
• Apparent Mediocrity. These titles are polarising both inter- and intra- proxies, in other words, they are widely read, they are considered outstanding in other quality proxies, and they are also considered to be great by a portion of the Goodreads’ community itself. Their apparent mediocrity is an effect of two very decisive readerships pulling the rope in opposite directions. Examples of these titles are Joyce’s Ulysses or Nabokov’s Lolita.

[43]What is ›mediocre‹ on Goodreads is thus a particular flavour of mediocrity, composed of different subpopulations. As such, it is advisable for studies that use Goodreads ratings as a proxy for literary quality to assess the amount of their corpus that consists of such works and to use Goodreads ratings to predict quality while keeping this aspect in mind. It’s notable that in this respect, the number of ratings can be more telling than the average score, as titles categorised as ›apparently mediocre‹ based on Goodreads ratings exhibited a consistently elevated volume of readership compared to titles that were genuinely indeterminate in their reception.

[44]It is important to underline that the strength of Goodreads’ rating system – its reductionist approach that distils reader judgement into a mono-dimensional metric – is also its limitation. The most ›mediocre‹ titles in our selection still have very positive and negative ratings, and the non-decisive or ambivalent judgement of many reviews often emanates from a composition of perceived strengths and weaknesses perceived by the individual readers. Looking forward, our ambition is to conduct a more extensive exploration of rating polarisation within the Goodreads community. While the standard deviation and entropy of individual book rating distributions serve as satisfactory preliminary metrics for gauging polarisation, the employment of more sophisticated analytical techniques is warranted. At a more complex level, applying techniques such as aspect-based opinion mining on Goodreads’ reviews could return a much more nuanced (but, again, harder to summarise) landscape on the reception of a given novel. Conversely, the presence of titles held in high regard by other evaluative mechanisms suggests that Goodreads serves as a viable alternative resource for literary assessment, corroborating the assertions of Verboord (2011).‍[40] Thus, future research endeavours that compare the mass appraisal methodologies of Goodreads with other literary evaluative proxies could contribute significantly to the discourse on literary evaluation.