Towards a hyperlinked society: A critical review of link studies

Studying the link

The present review emphasizes the study of web links, but it should first be underlined that the web is not the only system offering possibilities of hypertextuality. Other hypertext media exist, such as those examined by De Rose (1989) in the context of Bible or Classic studies. Investigations of citations or footnotes (Grafton, 1997) also emerge as direct ancestors and inspiration of web link analysis.

When it comes to web links, two axes of research can be distinguished: on the one hand, a stream of works broadly belonging to the network sciences and, on the other hand, a set of studies more closely related to social sciences. The former aims at describing hyperlink networks, at discovering their properties, while the latter interpret links as indicators of other phenomena.

Studies pertaining to the ‘new science of networks’ (Barabasi, 2003; Watts, 2004a, 2004b) throw light on properties of hyperlink networks at a general level. They are interested in understanding the web’s underlying structure. Such investigations into global hyperlinking patterns have ‘uncovered principles that help to understand networks of all type’ (Tremayne, 2004: 234). ‘Peculiar and fascinating properties’ (Ghitalla, 2009: 8) were identified. For example, researchers have determined that the web is a ‘scale-free network, dominated by hubs and nodes with a very large number of links’ (Barabasi, 2003: 165), and have revealed ‘strong regularities, among which the existence of a ‘universal power law’ (Adamic and Huberman, 2001: 131), that is, many aspects of the web comply with a distribution for which there exists an inverse relationship between the observed magnitude of an event and its frequency. For their part, Albert et al. (1999) established the diameter of the web and its small-world nature: despite the vast total number of existing websites, there are, on average, 19 degrees of separation between two pages (Barabasi, 2003: 165), that is, one webpage is on average 19 clicks away from another. Researchers also demonstrated that finding a path between two pages is not obvious: Broder et al. (2000) proposed a fragmented model of the web corresponding to a bow-tie structure where the direction of the links strongly matters (Barabasi, 2003: 166). According to their analysis, the web’s macroscopic structure naturally breaks into four parts: there is a central core, coined the ‘giant strongly connected component’, all of whose pages can reach one another; an IN-component that can reach the core; an OUT-component that can be reached from it; and tendrils that can neither reach the core nor be reached from it (Broder et al., 2000). This set of findings, belonging to the science of networks, informs us about the structure of hyperlink networks: they form a collection of general observations that one needs to be aware of when dealing with hyperlinked material.

The social significance of links

All the above are inquiries into the structure of the networks formed by hyperlinks. However, other investigations revolve around the nodes and the interpretation that can be made about them: hyperlinks are then used as substitutes for something else, as indicators of other phenomena. Those studies perpetuate the idea that linking behaviour is not random (Rogers and Marres, 2000: 6), and that links are ‘socially significant in some way’ (Hsu and Park, 2011: 355). In this perspective, links have an ‘information side-effect’, they can be used to understand other facts even though they were not individually designed to do so: ‘information side-effects are by-products of data intended for one use which can be mined in order to understand some tangential, and possibly larger scale, phenomena’ (Adamic and Adar, 2003: 211). In a word, there is a ‘widespread belief that hyperlinks between web pages can yield useful information’ (Thelwall, 2006: 1). Two underlying assumptions sustain this claim, at individual and global levels. On the one hand, it means that every link has been created for a precise purpose, that webmasters link selectively (as opposed to capriciously) in accordance with a ‘politics of association’ (Rogers and Ben-David, 2008: 43): ‘Making a link to another site, not making a link, or removing a link, may be viewed as acts of association, non-association or disassociation, respectively’ (Rogers, 2010: 117). On the other hand, links make sense at a global, aggregated level: taken together, they reflect social and cultural structures (Halavais, 2008: 39). In other words, ‘a hyperlink is not only a link but has certain sociological meanings’ (Hsu and Park, 2011: 364).

Indicators of authority

For what other phenomena do hyperlinks stand? Which sociological meaning hides behind them? The work of Brin and Page (1998), in which they explained the logics of their ‘Page Rank’ algorithm, initiated the trend of using links to calculate authority or relevance (Rogers, 2010). In this perspective, every link is considered as an implicit vote about the quality or relevance of a piece of content (Finkelstein, 2008): the more a piece of content is linked, the more authoritative it is. Such assumption rests on a parallel between hyperlinks and citations in the academic context: when often cited by their peers, scientists’ contributions are considered important and valuable. Similarly, Brunn and Dodge (2001) argue that we can exploit hyperlinks to assess trust in the context of international online trade. In concrete terms, various methods and different indexes have been proposed to systematically measure this notion of authority or trust on the web, as illustrated by the famous Page Rank (Brin and Page, 1998) used by Google or the ‘authorities/hubs’ model (Borodin et al., 2001). Other measurements directly draw on the study of social networks (Scott, 2000), where it is common to assess a node’s centrality (its importance in terms of connections) through various indicators; even if the notion of authority is less pregnant, the centrality of a node denotes its prominence in the network.

Monitoring academic performance

Research also expands on the parallel between links and citations, and broadly uses links to monitor some aspects of academic performance. Information sciences have extensively explored link analysis and applied its methods, tools and techniques to academic links (Thelwall, 2004). In this respect, Thelwall (2001) built a link-based ‘web impact factor’ to evaluate the research activity of universities in Britain. Harries et al. (2004) mapped the bounds between academic disciplines, with the underlying goal of evaluating universities and identifying patterns in the conduct of research. As for Park (2010), he investigated the ‘semantic variation, disciplinary scope and institutional structure of e-science programs in South Korea’ (Park, 2010: 211), using links as proxies for relationships between institutions and stakeholders (Park, 2010: 223). Those authors, however, remain wary of such ‘webometrics’ approaches. As Thelwall (2001) underlined, web link metrics can yield results that highly correlate with non-electronic phenomena (such as traditional impact factors in academe) but as with traditional citation analysis, it will never fully replace traditional measurements: those who only rely on link counts to judge academic performance risk serious errors (Thelwall, 2001: 18).

Reflecting political affiliations

Link studies have also gained importance in the field of political science, where researchers have investigated the ‘political web’. Papers by Park and other colleagues continuously explored South-Korean politicians’ web presence. They mined links to assess the ‘new ecology of political communication’ (Park et al., 2005). Their main idea is that links signal political affiliations, whether they investigate links stemming from politicians’ homepages to understand their political agenda, the communication network formed by assembly members and political parties (Park et al., 2004) or the relations between citizens’ and politicians’ blogs (Park and Jankowski, 2008). For instance, ‘links embedded in political blogs can be taken as technical indicator of the ideological landscape of a blog sphere being studied’ (Park and Jankowski, 2008: 62). Still in the political field, Williams et al. (2005) examined the hyperlinking strategies of candidates taking part in the 2004 US presidential race, and compared the different strategies used in the candidates’ blogs or homepages. They suggest that different genres of websites go with different linking styles: links on the candidates’ homepages more often point to internal fundraising or revenue-generating pages, whereas blog links are less likely to solicit donations and propose more external links. In this case, the authors suggest, links are loaded with strategic campaign interests.

Signs of political homophily

When it comes to assessing political issues, link studies are additionally concerned with the notions of political homophily and cyberbalkanization. Proposed by Sunstein (2001), the notion of cyberbalkanization suggests that with the possibilities of personalization proposed by the web and its ever-growing offer in contents, people tend to generally consult pages harbouring opinions similar to theirs. Such isolation is presented as a danger for democracy, as ‘people will abandon the reading of dissenting political opinions in favour of material that is closely aligned with their own ideological position’ (Hargittai et al., 2008). Hargittai et al. (2008) tested the cyberbalkanization assumption in the 2004 US political blogosphere, and found that widely read political bloggers link more often to those who share their views. But this trend, they added, does not increase over time. Also in the context of the 2004 US Presidential campaign, Adamic and Glance (2005) measured the interconnectedness of the two parties, as embodied by the ‘degree of interaction between liberal and conservative blogs’ (Adamic and Glance, 2005: 1). They witnessed a ‘divided blogosphere: liberals and conservatives linking primarily within their separate communities’ (Adamic and Glance, 2005: 14), with only 10% of the links bridging the two groups. Moreover, they detected that conservative blogs link to each other more frequently and in denser patterns. Ackland and Shorish (2009), further exploring Adamic and Glance’s dataset, claimed that manifestations of homophily – or ‘assortative mixing’ as they call it – must not be attributed to the intrinsic nature of bloggers to favour their own kind, but that this linking behaviour is explained by a model of network formation in the blogosphere and, more precisely, by the dynamics between a political majority and minority. In other words, homophily is merely an effect of the forces of the network, and ‘it is possible to view the creation of links by and between political bloggers as a reflection of the relative popularity of one ideology over another in the underlying population from which bloggers are drawn’ (Ackland and Shorish, 2009: 397). The authors admitted, however, that while focusing on link creation as a result of the relative popularity of one ideology over another, their model abstracts away from other influences likely to dictate link formation (Ackland and Shorish, 2009: 397). The evolution of political linking patterns over time has also been studied, as exemplified by Hsu and Park who scrutinized the differences between different generations of links. Their ‘sociology of hyperlinks’ explored links produced by South-Korean politicians, and they highlighted a significant evolution between web 1.0 (personal websites), web 2.0 (blogs) and Twitter links: the network of links appeared denser over time, and there is evidence that it became increasingly less centralized (Hsu and Park, 2011: 363).

To sum up the state of link studies in the political field, we see that researchers use links as indicators of other phenomena that vary in their levels of abstraction and generalization. The most down-to-earth approach interprets links as mere signs of communication networks between involved actors, or strategic tools for campaigning in one particular case study. One step further, we find the notions of links as signs of ideological affiliations, or symptoms of ‘cyberbalkanization’.

Tracing public debates

Besides politicians’ online behaviours, wider political issues are also on the agenda. Here, scanning links involves looking at how public debate is shaped on the web, and how ‘discursive affinities’ or ‘issue networks’ are structured. Rogers and Ben-David (2008) examined the issue network around the Israeli security fence debate, and showed how the various actors were situated toward each other. By mapping the links between NGOs’ websites, they discovered that Israeli left-wing organizations are in ‘virtual isolation’ in the overall issue space (Rogers and Ben-David, 2008: 63). Similarly, Rogers and Marres (2000) attempted a ‘debate-scaping’ of the climate-change discussion. They encounter different ‘linking styles’ typical of the .com, .gov and .org domains, and noticed, for example, that ‘governmental institutions tend to form one massive, autonomously operating body’ (Rogers and Marres, 2000: 9). Consequently, they argue that the online debate about global climate change is shaped by the presence of a central governmental authority (Rogers and Marres, 2000: 13). Marres (2004), in turn, concentrated on the argument around a controversial project led by the World Bank. Elaborating on actor-network theory, she used links to describe how the issue was evolving and shaped by different actors. All those studies use the links as visible trails of public issues being debated, as tools to clearly map the positions of involved actors. The authors implicitly suggest that such debate-scaping is useful to better understand what is at stake in our democratic societies.

Connections between blogs

Efforts to map links between blogs also relate to this movement. In this context, hypertextual proximity helps to make sense of vast ‘blogospheres’, that is, sets of interconnected blogs. Blogs are particularly salient objects when it comes to studying links, because they are a link-driven genre (Blood, 2000). Bruns (2007) mixed the issue-centred approach with the emphasis on blogs, as he maps the interactions between Australian bloggers about the David Hicks case. His findings highlighted a ‘tendency for discussion to cluster around a handful of sites which are defined by their political orientation’, a lack of coverage of the case by bloggers outside of Australia and limited interactions between bloggers and the mainstream media (Bruns, 2007). Tremayne et al. (2006) mapped the structure of the Iraq war blogosphere. Their findings are consistent with previous comments on cyberbalkanization, and they highlight a blogosphere structured in two distinct halves, the liberal and the conservative – ‘but also a fairly robust cluster of blogs that serve as conduits between the sides’ (Tremayne et al., 2006: 305). Etling and his colleagues’ approach is wider: they scanned nation-wide blogospheres at a global level (Etling et al., 2009, 2010; Kelly and Etling, 2008) to understand how public debates between bloggers are shaped and how they fit in wider media ecology. In that perspective, links significantly help in detecting ‘attentive clusters’, that is, groups of blogs that show similar linking behaviour and that link to the same resources. Those methods are ‘based on the principle that macro structure arises from the tendency of individuals to link more frequently to things that interest them and to people with whom they share attributes or social relations’ (Etling et al., 2010: 11). Researchers use linking patterns to make sense of large amounts of blogs and to gather them in significant clusters – as similar linking behaviours ‘often arise from intuitively grasped interests, ideologies, preferences and affiliations that exist among bloggers’ (Etling et al., 2010: 14). From then on, hypertextuality constitutes a non-semantic tool to discover semantic patterns (Jacomy and Ghitalla, 2007: 4): if two blogs link to exactly the same resources, it is likely that they will cover similar topics, and perhaps hold similar views on them. In substance, hyperlinks help to detect ‘topical localities’ (Ghitalla, 2009); hypertextual proximity indicates similarities in content (Ghitalla et al., 2005: 8).

International flows of information

Links studies also set out to tackle journalistic issues. In this framework, hypertextuality is associated with positive journalistic values such as interactivity, transparency, credibility or diversity. Adding links to stories allegedly improves online news: ‘Choosing links to include in your story gets to the very essence of what it means to be a journalist’ (Foust, 2009: 161).

Therefore, links stemming from news sites are investigated to assess those sites’ compliance with the shared professional standards of quality online journalism. Nevertheless, empirical findings tend to indicate that news sites do not live up to the expectations, as their use of external linking is scarce (Dimitrova et al., 2003; Dimitrova and Neznanski, 2006; Kenney et al., 2000; Quandt, 2008; Stray, 2010; Tankard and Ban, 1998; Tremayne, 2005; Tsui, 2008).

Links on news sites are also examined to map international information flows. Chang et al. (2009) tested the structure of links originating from news sites in regard to the world-system theory ‘which emphasizes the asymmetry between information-rich and -poor countries’ (Barnett et al., 2011: 370). Contrary to the impression of openness favoured by the internet as a borderless medium, linkages between nations via outgoing hyperlinks to websites of other countries remain relatively closed (Chang et al., 2009: 155). Halavais (2000) presented similar findings concerning the web at large. He argues that most links are restrained within national borders: ‘although geographic borders may be removed from cyberspace, the social structures found in the “real” world are inscribed in online networks’ (Halavais, 2000: 7). The intertwinement of ‘real-world’ geography and international link structures stands at the core of other studies. Barnett and Sung (2005) examined the patterns of communication among nations, through the interconnection of top-level domain names (such as ‘.de’ or ‘.fr’). They show that hyperlinks comply with a centre-to-periphery structure. According to them, economic relations are the primary organizing mechanism of international communication. Links reflect power interactions between countries, and core countries are clearly central to the web. Typically, the United States constitutes the most central node in the hyperlink network (Barnett and Sung, 2005: 226). This phenomenon turned out to be even clearer when Barnett et al. (2011) overcame an inherent bias of previous studies by unveiling what is behind the .com domains: the core countries’ domination becomes stronger. Also mixing virtual and physical geography, Lin et al. (2007) analysed links between blogs as indicators of connections between US cities. They pinpointed ‘super metropolitan clusters’ (Lin et al., 2007: 21), that is, the densest hyperlink networks around cities with cultural and political prominence. Thus, ‘centers of cultural and news production still attract the most attention nationwide’ (Lin et al., 2007: 22), whereas cities of less cultural–political significance mostly connect with places near them.

Finally, links are also used as signs of relationships between people. In Adamic and Adar’s work (2003), for instance, hyperlinks between homepages are used as an indicator of ‘friendship’ between students. The authors argue that using links in this context allows the collection of data to track social networks in a convenient way – whereas traditional social network research has to cope with long interviews and observations.

Barriers and limitations to link interpretation

When considering the scope of research projects outlined above, one may think that interpreting links as signs for social phenomena is a straightforward operation. However, limitations exist and need to be mentioned. The first promise of link studies relies in their apparent technical simplicity. At first glance, links may seem easy to detect: they are well-defined strings of HTML code and there is no obvious barrier to their collection. One only has to browse through the code and to retrieve everything between the <a href> and </a> tags. Doing this at a large scale, however, is time-consuming; therefore, tools have been designed to facilitate such tasks. Some of them are ‘web-crawlers’, designed to follow and index links automatically, while others draw on search engines (Thelwall, 2001: 15). Pointing out all these tools would be impossible, but a handful of freely available tools oriented towards social science research and integrating link analysis in their features exist: SocSciBot, ¹ LexiURL, ² the VOSON project, ³ IssueCrawler, ⁴ Navicrawler, ⁵ etc.

However, the fact that tools exist to tackle the important practical and technical difficulties related to hyperlink analysis should not lead us to the assumption that studying links is easy. Issues in the collection and interpretation of links must be taken into account.

A false feeling of unobtrusiveness

When discussing the strength of link analysis, Park (2003) argues that its main quality is its unobtrusiveness, because ‘hyperlink data can be gathered naturally without intruding in the research context’. However, Halavais stresses that ‘those measuring the system have become a significant part of it’ (Halavais, 2008: 48) and that, therefore, measuring a hyperlink network is not as unobtrusive as it seems. The mere fact that it is widely known that Google’s search engine algorithm draws on links creates a potential bias. Search engine optimization is more and more common: ‘the study of macrolevel hyperlink networks has been complicated by users becoming more aware of hyperlink structures’ (Halavais, 2008: 53). Ghitalla also remarks that his hyperlink map of the European web, when published online, logically appears at the centre of the network pictured – as it contains links to every other concerned website: ‘the map, once published on the network, modifies the territory it represents’ (Ghitalla, 2008: 10). But there are other ways in which the observer, who thought he was invisible, influences what he observes. In Marres’ case (2004), the crawl for links did not go unnoticed and resulted in an email from the owner of the website she was interested in. This reaction, the author argued, ‘could lead us to pose all sorts of questions, for example about the inevitable involvement of Internet researchers in the phenomenon studied (the anthropologist’s problem of “meshing with the natives”)’ (Marres, 2004: 125).

Context matters

Link interpretation often implies something about the intentions of the people that have created them. Even when considering the simplest dichotomy it is obvious that a link can mean praise as well as criticism. There is no guarantee that indication of the author’s intentions can be found in the link itself, or in the linked resource. Zimmer (2009) compares links with renvois, an 18th century ‘system of cross-references’ featured prominently in Diderot and d’Alembert’s Encyclopédie (Zimmer, 2009: 96). He states that these antecedents of hyperlink text are not as candid as they seem: the writers of the Encyclopédie employed renvois to avoid censorship when discussing controversial topics, relying on ‘irony, innuendo and indirection’ (Zimmer, 2009: 103). Similar practices exist in the context of linking. For example, how to make sense of the ironic linking practice known as ‘Rickrolling’? This famous internet prank happens when a person provides a hyperlink that they claim is relevant to the topic at hand, but the link actually takes the user to a music video for Rick Astley’s song Never Gonna Give You Up (Wikipedia, n.d.). The nature of such link is equivocal and we can only understand it if we are aware of the Rickrolling meme.

In a word, ‘the universal nature of hyperlinking makes it a very difficult sort of artefact to understand’ (Halavais, 2008: 43). Most authors thus acknowledge the importance of the context of links. For instance, scholars studying blog links often make a difference between ‘in-post’ links and ‘blogroll’ links, that is, the links present within the text of a blog post or those displayed in a side-column (Adamic, 2008; Lin et al., 2007), and argue that they cannot be amalgamated. Adamic (2008) claims that blogroll links serve as badges that bloggers may display, they indicate ‘a general social awareness on behalf of the author’ (Marlow, 2004: 3), and constitute an indicator of the ideological landscape surrounding the blogs (Park and Jankowski, 2008; Park and Thelwall, 2003). For Lin et al., blogroll links are more indicative of interpersonal affiliation than in-post links (Lin et al., 2007: 17), whereas in-post links are considered as a better indicator of active conversations going on in the blogosphere. They allow ‘authors to have a sort of distributed conversation’ (Marlow, 2004: 3), and they ‘signal active engagement with another blogger’ (Hargittai et al., 2008: 76). As argued by Adamic and Glance (2005: 7), the ‘linking behavior within posts is more indicative of a blogger’s reading activity than are blogroll links’. To sum up, there is no cohesive theory about the potentially different meanings of blogroll and in-post links.

More broadly, differences between web genres also matter. For instance, the framework used to make sense of links on personal homepages is deemed inadequate for hyperlinks in the blogosphere ‘because there are several types that are unique to blogs’ (Park and Thelwall, 2008: 860). Moreover, in their attempt at mapping links between different academic subjects, Harries et al. (2004) notice disciplinary variances in the types of links, and again underline that the context is crucial. The need to understand the motivations behind linking appears as a leitmotiv. The parallel with academic citations is again prominent, with Thelwall (2003) elaborating a typology of the motivations to link on the basis of previous research about motivations to cite. In their study of the evolution of linking behaviour, Hsu and Park noticed that incentives to link vary: a link on Twitter, for instance, is likely to result of a different motivation than a link on a homepage or on a blog. Most authors acknowledge that grasping the motivations behind linking would greatly enhance their analysis. Park (2003: 58) in particular observes:

…several methods need to be employed to examine the reasons developers of websites form a network with other sites via hyperlinks: survey, in-depth interviews, observation, comparative analysis of website contents and other network data would contribute to an understanding of the social relationships among the network’s components.

But even by interviewing link creators, notes Thelwall (2006: 4), it would be almost impossible to systematically identify what someone has in mind when placing a link. Concretely, he adds, finding the author of a webpage might be difficult; the authors may not remember why they included URLs; and in-depth interviews are difficult to carry out on a large scale.

Promises of automation

In this respect, one of the main attractive aspects of link studies should be considered with caution: the possibility to crawl the web at a large scale, to harvest a large number of links and to study them globally. If all links are not created equal, the automatic treatment and interpretation of a large amount of links may therefore become delicate. In their attempt at studying the whole ‘web graph’, Bharat et al. (2001) used the link structure to determine if sites are related. However, something as simple as link frequency to determine relatedness is sometimes misleading, as the authors witnessed ‘strongly connected host pairs that didn’t seem to be otherwise related’ – such as large hosts like www.geocities.com and members.aol.com that are strongly connected only by virtue of their immense size (Bharat et al., 2001: 5). Thelwall summarizes the issue as follows: ‘the reason for high link count can be completely different to the average reason for link creation’ (Thelwall, 2006: 9). More radically, he pinpoints an important issue regarding the use of statistics: ‘fundamental statistical problems can arise if links are not the primary object of study but serve as a vehicle to draw conclusions about processes underlying link creation, which is typical of social sciences link research’ (Thelwall, 2006: 7). Links are not statistically independent observations, ‘because of factors such as imitation between web authors (…), the copying of pages or parts of pages, and automatic creation of web pages by server software’ (Thelwall, 2006: 7). According to Thelwall, this means that inferential statistics, such as confidence intervals, are not valid. As Harries et al. (2004) phrase it, web link creation is a ‘social activity that inspires imitation, the opposite of statistical independence’ (Harries et al., 2004: 439).

Conclusion: Mixing methods

Facing these difficulties in treating links automatically and at a large scale, most researchers successfully opt for semi-automatic methods, or a mix between large-scale, automatic data retrieval and smaller, manual processing. For example, Etling et al. (2010) mainly drew on automatic methods in order to map the ‘attentive clusters’ of the Russian blogosphere but they combined them with a manual, qualitative assessment of 1200 blogs. Most of the projects presented above rely, at some point, on manual coding, qualitative appreciations and human expertise. Hargittai et al. (2008) qualitatively qualified links between liberal and conservative blogs; Park (2010) manually classified sites by authors; Park and Jankowski (2008) qualitatively analysed blog titles; Thelwall (2003) manually categorized links by motivation. Some authors especially insist on the role played by experts: only people that are specialists in the field investigated can make sense of the links (Ghitalla, 2008; Jacomy and Ghitalla, 2007).

Most of the papers reviewed here do more than pure link counts or link network analyses. They often mix link counts with other methods and other datasets. Content analysis of texts is often brought in: Adamic and Adar (2003) looked at the texts on presidential candidates’ homepages. Etling et al. (2010) use computer-assisted text analysis of blog contents. TACT (Textual Analysis Computing Tools) were also used by Rogers and Marres (2000) in order to locate key phrases in their context. Park and Jankowski (2008) carried out a follow-up qualitative analysis on blog titles. Rogers and Ben-David (2008) chose to pair their link analysis with a term analysis: they scrutinized the words used to frame their topic of interest (i.e. the building of the Israeli ‘security fence’). All those approaches mix link studies with the examination of other web contents. For Thelwall (2001), however, this is not enough: we also need to combine link analysis with ‘non-web based information’, in order to go ‘towards a hybrid calculation combining web information with another source’ (Thelwall, 2001: 1158). For example, in his study of universities’ ‘web impact factor’, Thelwall integrated the number of Faculty members of a particular institution (i.e. a non-web-based information) as a factor in the calculation.

Through this review of research projects involved in link analyses, we have shown that link interpretation is convoluted and sometimes problematic. However, the variety of fields applying link studies, as well as the multiplicity of possible interpretations, should not lead us to conclude that these tools and methods are chaotic or sterile. Several pragmatic conclusions must nevertheless be pointed out: a univocal, unambiguous interpretation of the link does not exist. Similarly, the possibility of automatically handling large amount of data should be limited to certain areas, such as the science of networks – that aims at determining how the network of links is organized. When starting a research project involving link analysis, one should always ponder the following question: when do we stop to purely describe the structure of links as technical objects and when do we start to relevantly exploit links to make sense of a social phenomenon? If we choose the second option, we must ensure validity in order to guarantee that the web-based elements we study significantly connect with the offline social phenomena we are trying to understand. This is an issue transversal to the field of web studies and that needs to be explored in a broader context. In the limited context of hyperlink analysis, however, the juxtaposition of innovative link research shows that mixed methods appropriately allow using hyperlinks as proxies for other phenomena.

Validity may be achieved, as suggested by most of the works mentioned earlier, by combining link analysis with other tools and methods, as well as with a strong focus on human expertise to make sense of the connections. The researchers whose work is examined above implemented such a framework, implicitly or explicitly. Inspecting their work side by side shows us that there is such thing as link studies in the social sciences. What they have in common is not a unique interpretation of the hyperlink, but rather a shared, mixed methodology. Various authors proposed their ad hoc tricks to tackle the various methodological issues detected. Their rapprochement in this literature review pleads for a systematization of their methods in a unified yet adaptive framework that would consist in combining quantitative link counts, qualitative inquiries and valuation of field expertise to support link interpretation.

Since the advent of the hyperlinked society, connections between actors embodied in websites and webpages have been explicitly visible and we should not let such opportunity unstudied. This review hopes to open a path towards a sustainable link analysis agenda.