Finding 1: Availability (RQ1)
The conspiracy books availability varied across each Amazon 19 national marketplaces. Out of the 20 identified COVID-19 conspiracy books, six books were unavailable on all 19 national markets. All of them were available in GoodReads and none in Audible. Table 1 indicates with "x" unavailable books and with “eyes” the ones viewable on the websites.
An overall takeaway is that the availability of a book is independently determined by the platform. Despite Amazon being the parent company of Goodreads and Audible, the outcome of our research shows different results for all three platforms. This illustrates that each platform curates its policies towards conspiracy theory books differently.
Table 1. The availability of the books on Amazon’s 19 national marketplaces
Figure 1 shows the availability data sorted into a bar graph to illustrate the countries that have less available conspiracy books, on the left, compared to the countries that have the most available books, on the right. This ranking coincides with the general understanding of these countries' stances towards internet censorship (Bischoff, 2020). Sweden appears as an outlier because it does not as actively censor this type of content but may have fewer available books because of its unique approach to the COVID-19 pandemic (Molloy, 2020).
Figure 1. Amazon’s deplatforming of COVID-19 conspiracy books per country
The outcome of the varied Amazon national marketplace availability is not fully explained by Amazon's policies. Once a book is uploaded to an Amazon country website, a Kindle (digital) version of a book should be available across all Amazon country marketplaces (Chesson, 2020). For physical books, the uploader can select the desired “target region” for their book (e.g. Amazon Europe; Japan; Australia; North America) (Amazon Services, n.d.).
In both cases, distribution is dependent on publishing rights. However, as the Kindle Direct Publishing website states (Kindle Direct Publishing, n.d.), original unpublished content usually has worldwide rights. As it seems to be the case of the titles of our study, this factor may not fully explain discrepancies in availability across Amazon stores
Therefore, it is unknown if the books that are unavailable in specific Amazon country markets are inaccessible because the uploader chose to not list them in these regions, if Amazon actively removed them from these specific domains, if there are publishing rights restrictions involved, or if the algorithm for these Amazon domains does not encourage them to be listed. To answer this, further research and potential consultation with Amazon is required.
Regarding GoodReads book availability, the main finding is that the platform has available all the titles, including links to bookstores and spaces for user's reviews. Users can trace back to the original link of the book on Amazon as well as view other alternative online bookstores carrying the book, such as Barnes & Nobles and Book Depository.
Users who aren’t able to access the books on Amazon, thus use Goodreads as an outlet to express their concerns about the removal. The review section becomes an unmoderated space for users to share their inputs on the readings and alternative routes to access unavailable content.
Figure 2. Screenshot of a review on GoodReads, providing a free Kindle copy of a deplatformized book
Finding 2: Recommendation system (RQ2)
Although some conspiracy theory books were not available, Amazon's recommendation system displayed other conspiracy theory books for customers to purchase. By searching for unavailable books on different national Amazon stores, page results fitted into 5 categories. The first category has no results and additional information, the second category is sponsored recommendation, the third and fourth categories are automatically recommended books by the system, and the fifth category is related to products recommended by the system.
Figure 3 illustrates one of the books deplatformized on Amazon, The Great British Coronavirus Hoax: A Sceptics Guide, with category 3, “try these instead”, recommending another COVID-19 conspiracy book. Therefore, the removal of one COVID-19 conspiracy book may lead users to alternative options.
Figure 3. The deplatformized book ‘The Great British Coronavirus Hoax: A Sceptics Guide’ on Amazon with “Try these instead” algorithmically suggested
Through the analysis of the third and fourth categories, it was found that different countries have different system recommendation preferences. The title "The Great British Coronavirus Hoax: A Sceptics Guide" on Amazon in the U.S., for example, has different recommended categories from the ones that appear in countries like France, India and Spain. Image 3 shows that U.S. category recommendations revolve around "world" and "European" while in other national stores they are related to the title of the book, such as "British" and "coronavirus".
Figure 4. Book ‘The Great British Coronavirus Hoax: A Sceptics Guide’ on France, India, Spain and the US Amazon with different suggested categories
In the case of Audible platform, out of the 20 books, all of them unavailable, seven titles lead the user to a recommendation page. Most of the recommended problematic books had the QAnon conspiracy theory as the main subject. Similar to Amazon’s recommendation feature, conspiracy theory book unavailability on Audible does not diminish access to the titles, rather, it presents users with alternative problematic books.
Figure 5. Recommended alternatives for “Rises of the New World Order: Book Series Update and Urgent Status Report: Vol. 1" on Audible
Finding 3: Platform affordances (RQ2)
Despite Amazon owning Audible and Goodreads, each platform has its own affordances. As shown in Table 2, platforms have three similar affordances: rating, overall rating, and reviews. These features were taken into consideration when exploring how COVID-19 conspiracy books fostered conspiracy ecosystems.
Table 2. Platform features on Amazon, Audible, and GoodReads
The analysis has shown that 11 of the 20 books were listed in one or more of Amazon's best-seller categories, with some even listed above the #50th position. Table 3 shows that there are titles in #3 of 45-Minute Religion & Spirituality Short Reads, #5 of Business Planning & Forecasting, #6 of Contagious Diseases and #7 of Vaccinations categories.
The best-seller illustrates how popular the book is in a specific category but also portrays how legitimized it is in particular communities. Therefore, it exemplifies how conspiracy theories ecosystems are benefiting from Amazon's platform affordances.
Table 3. Books per position in Amazon's best-seller categories
Note. X= not included in the best-seller category. Star= Ranked above position #50.
Finding 4: Conspiratorial ideation styles and epistemic markers (RQ3)
When characterizing the nature of discussions on Amazon and GoodReads, with respect to indicators of epistemic beliefs comparable actors, actions and stances emerge. The co-occurrence network as well as Global Vectors for Word Representation i.e., GloVe embeddings for Amazon are based on n = 587 user reviews (comprising n = 46,811 lemmata) regarding n = 63 books. The following target terms of epistemic beliefs have been selected: ‘see’, ‘truth’ and ‘prove’, as well as more general concepts like ‘people’ and ‘vaccine’.
Based on the positive pointwise mutual information (PPMI) term-feature matrix these target terms are exemplified in a co-occurrence network (see Fig. 6) which show strong word associations between terms relating to truth propositions and causal assumptions (see Tab. 4).
Note. Node size indicates positive pointwise mutual information value (PPMI) between term and feature, node shapes indicate: triangle = target term, round shape = features. Target terms queried were ‘vaccine’, ‘people’, ‘prove’, ‘see’ and for each the 30 most associated terms. [Visualization with ggraph version 2.0.3 (Pedersen et al., 2017)]
Table 4
Example Target Terms from the PPMI-Term-feature matrix for Amazon reviews
Target term | Functional class | Example Features (PPMI) |
'see' | Qualifiers of relevance and urgency | refreshing (4.74) |
adequate (4.56) | ||
pressure (4.90) | ||
coming (5.03) | ||
happening (3.97)
| ||
'vaccine' | Causal speech acts (entities and outcomes) | pharmaceutical (3.51) |
proponents (3.83) | ||
program (3.35) | ||
gardasil (3.83) | ||
consequences (3.39) | ||
injury (3.81) | ||
injured child (3.51) | ||
ineffective (3.51) | ||
killer vaccine (3.27) | ||
'truth' | Truth propositions | truth reveals (4.34) |
scam truth (3.62) | ||
tell truth (5.23) | ||
For the GloVe word embeddings 'see' returned the following terms with the highest cosine similarity: statistics (cos(θ) = 0.27), lingo dumb (cos(θ) = 0.26) or consensus scientific (cos(θ) = 0.27). For 'truth' the highest values for semantic similarity are dangerous (cos(θ) = 0.30), investigation (cos(θ) = 0.22), power (cos(θ) = 0.25), claims (cos(θ) = 0.211), silenced (cos(θ) = 0.20). Further, 'prove' returns the highest cosine values for nazis (cos(θ) = 0.24), search (cos(θ) = 0.24) or tell truth (cos(θ) = 0.26). For 'people' close terms returned are suffer (cos(θ) = 0.28), koehnlein (cos(θ) = 0.28) or admit (cos(θ) = 0.23) whereas, lastly, 'vaccine' is closely related to viral (cos(θ) = 0.28), inconvenient (cos(θ) = 0.25), democratic (cos(θ) = 0.31) or causes (cos(θ) = 0.24).
Regarding the co-occurrence network for GoodReads ’ reviews (see Fig. 7), it is based on n = 113 user reviews (with n = 30,298 lemmata) relating to n = 12 books. Indicators for assertions, conclusions and truth assumptions were: 'see', 'truth' and 'mainstream', 'hidden', as well as 'vaccine'. Example target terms from the PPMI term-feature matrix are shown below (see Tab. 5).
Figure 7
GoodReads Review Co-occurrence Network based on Term-feature Matrix
Note. Node size indicates positive pointwise mutual information value (PPMI) between term and feature, node shapes indicate: triangle = target term, round shape = features. Target terms queried were ‘vaccine’, ‘truth’, ‘mainstream’, ‘hidden’, ‘see’ and for each the 30 most associated terms. [Visualization with ggraph version 2.0.3 (Pedersen et al., 2017)]
Table 5
Example Target Terms from the Term-feature matrix for GoodReads reviews
Target term | Functional class | Example Features (PPMI) |
'see' | Qualifiers of relevance and urgency | see happening (3.51) |
see fail (4.29) | ||
'vaccine' | Causal speech acts (entities and outcomes) | contains (3.16) |
causes (2.89) | ||
behind (3.16) | ||
death (2.78) | ||
dangerous (2.84) | ||
propaganda vaccine (3.16) | ||
'mainstream' | Entities | media mainstream (5.60) |
narrative (5.89) | ||
'truth' | Truth propositions | deeper truth (4.66) |
'hidden' | Truth propositions | hidden behind (5.33) |
decades hidden (5.06) | ||
decades propaganda (4.83) | ||
Visualizing term-feature relations of terms with related semantic meaning in the field of potential conspiratorial worldviews offers valuable insights on possible sources of these similarities and recurrent motives (e.g. the combination of entities like ‘mainstream media’, actions and outcomes e.g. death). It allows a focus on speech acts and social meaning construction making use of epistemic verbs like ‘see’ instead of being confined to predetermined literal topics and narratives and helps to understand interactions in a community culture. One of the main speech acts relates to the veracity aspect of information. In this argumentation mainstream media or proponents of vaccines are distorting and hiding information. As meaning construction depends on time, on the one hand dynamic word embeddings (e.g., RoBERTa) might prove to be a valuable future avenue, as they capture polysemy. On the other hand, the way respective books are received on other (fringe) platforms outside the Amazon sphere might offer even more differentiated views on a time lag in adoption as well as the significance and reach of sales platforms in the proliferation of potential conspiratorial worldviews.
From the research, it is possible to conclude that the unavailability of conspiracy theories books on Amazon does not fully diminish the accessibility to these and other problematic titles. Indeed, some of the platform affordances, such as recommendation features and users' reviews, serve the purpose of feeding users with alternative routes to conspiracy theory books even when the title of the original search is not available.
Additionally, discrepancies in availability across marketplaces and platforms do not encounter suitable explanations in Amazon's policy guidelines. Therefore, the lack of equitable and transparent moderation policies and practices may be argued as factoring in thespread of infodemic.
Further research on the role of Amazon as a social media company fostering potential conspiracy theorising could benefit from an analysis into the network of GoodReads concerning other bookstores it has affiliated with. Unlike Amazon's 19 national marketplaces, GoodReads does not adjust to the user’s location. Therefore, most of the reviews were written in English and published by users in English-speaking countries. In China, for example, Douban would be the alternative website for GoodReads. As mentioned, a bigger picture of COVID-19 conspiracy books online can be researched through a different angle by looking closely at each country as a case study.
Amazon Services. (n.d.). How to sell internationally with Amazon Global Selling. Retrieved from https://amzn.to/3pzv0aG
Bischoff, P. (2020, January 15). Internet Censorship 2020: A Global Map of Internet Restrictions. Comparitech. https://www.comparitech.com/blog/vpn-privacy/internet-censorship-map/
Chesson, D. (2020, November). Amazon International Markets Explained. Kindlepreneur. https://kindlepreneur.com/selling-kindle-books-internationally/.
Chollet, F., & Allaire, J.J. (2018). Deep learning for text and sequences. In Chollet F. & Allaire J.J.(Eds.), Deep learning with R. Manning. https://jjallaire.github.io/deep-learning-with-r-notebooks/notebooks/6.1-using-word-embeddings.nb.html
Firth, J. R. (1957). A synopsis of linguistic theory, 1930-1955. Studies in linguistic analysis.
Humphreys, A., & Wang, R. J. H. (2018). Automated text analysis for consumer research. Journal of Consumer Research, 44(6), 1274-1306. doi:10.1093/jcr/ucx104
Infodemic (2020, October 21). What are COVID-19 Conspiracy Theories? [Blog post] Retrieved from http://infodemic.eu/2020/10/21/what-are-covid-19-conspiracy-theories.html
Kindle Direct Publishing. (n.d.). Paperback Distribution Rights. Retrieved from: https://kdp.amazon.com/en_US/help/topic/G201834280
Molloy, S. (2020, May 3) Coronavirus: Inside Sweden’s unique and controversial approach to Covid-19. nzherald.co.nz. https://bit.ly/3owGc6I
Nord, A. (2021). Amazon Product Api. Amazon-buddy. 2.2.33. https://github.com/drawrowfly/amazon-product-api
Pedersen, T. L., Pedersen, M. T. L., LazyData, T. R. U. E., Rcpp, I., & Rcpp, L. (2017). Package ‘ggraph’.
Pennington, J., Socher, R., & Manning, C. D. (2014, October). Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) (pp. 1532-1543).
R Core Team. (2020). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. [Computer software].
Shieber, J. (2021, January 12). Amazon is removing products promoting the QAnon conspiracy. Techcrunch. https://techcrunch.com/2021/01/11/amazon-is-removing-products-promoting-the-qanon-conspiracy/
Votta, F. (2016). Package ‘rgoodreads’. https://github.com/Famguy/rgoodreads
Selivanov, D., & Wang, Q. (2018). Text2vec: Modern text mining framework for r. Retrieved from https://CRAN.R-project.org/package=text2vec
Straka, M., & Straková, J. (2017, August). Tokenizing, pos tagging, lemmatizing and parsing ud 2.0 with udpipe. In Proceedings of the CoNLL 2017 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies (pp. 88-99).
Timm, J. (2019). Package ‘textvardatr’. https://github.com/jaytimm/lexvarsdatr
van Prooijen, J. W., & Douglas, K. M. (2018). Belief in conspiracy theories: Basic principles of an emerging research domain. European journal of social psychology, 48(7), 897-908. doi:10.1002/ejsp.2530
Wang, Y., Hou, Y., Che, W., & Liu, T. (2020). From static to dynamic word representations: a survey. International Journal of Machine Learning and Cybernetics, 11, (pp. 1611-1630). doi:10.1007/s13042-020-01069-8
Wickham, H. (2017). The tidyverse. R package ver, 1(1), 1.