The geographies and complexities of online networks in the off-street sex market

This research used a data set consisting of 213,693 unique online escort adverts, nested in 15,016 online networks (what is meant by ‘networks’ will be described in more detail below). The data set was created as part of a larger mixed-methods project (see Kjellgren, 2023) and a web scraper (written in Python 3.8.5, using the scrapy package¹) was used to collect adverts on a fortnightly basis, between January 2021 and April 2022. The adverts were collected from one of the UK’s most popular ASW (henceforth referred to as ASW 1). The rationale for selecting this particular website was informed by previous research in the United Kingdom (including Kjellgren, 2019). In addition, some ASWs require more rigorous verification procedures for posting adult services adverts, such as providing passport details (Sanders et al., 2018), and it is conceivable that criminal networks and OCGs would favour ASWs where they can provide as little information as possible, and ASW 1 is one such site. The poster is required to pay a fee to post an advert, which remains online for a limited time. The site in question includes several subcategories of adult services. Since this research is focussed on the exploitation of adult women, the category relating to escort services provided by women was used as the basis for data collection. It should be noted that there are subcategories of different genders on ASW 1 and it is conceivable that transwomen may also choose to post in this category; this analysis simply focussed on adverts posted under the category of ‘female’.

Since the data can be considered ‘found data’ (Connelly et al., 2016) – data occurring as a result of social processes and not created for the purpose of research – some further explanations are necessary. Escort adverts are posted by independent sex workers and criminal networks alike to connect with sex buyers (Crocker et al., 2017) and doing so generates a digital record of the online marketing strategies used (e.g. the textual descriptions, characteristics and services offered). This is a crucial point: an advert represents particular elements of the marketing strategies used by individuals or networks, and they do not necessarily represent a unique individual, nor are the details provided necessarily truthful (Holt et al., 2021; Kjellgren, 2022, 2023). As an example, the advertised nationality can be considered part of a repertoire of marketing strategies, since it can be advantageous to conceal or advertise as a different nationality to appeal to particular demands in the market, or to avoid stigma.

Both independent sex workers and criminal networks are likely to advertise sexual services in different places throughout time, since mobility is key to appealing to new markets and reaching a greater number of clients (Crocker et al., 2017; Scoular et al., 2019). As such, these activities generate digital traces that can be used to analyse how individuals and networks try to penetrate local sex markets throughout the country. In terms of geographical variables, location data are self-reported. To post an advert on the ASW, the poster is required to specify their postcode. The full postcode is not accessible to the public, instead, the geographical information contained in adverts consists of the postcode district and coordinates. For this research, the coordinates were linked to the geographically closest full postcodes and the local authority district (LAD) to which they were related.

In the context of policing organised crime in the sex market, it is crucial to understand where vulnerability and risks of exploitation may be most prominent, including the structure of networks operating therein. The escort adverts collected for this research were used to identify online networks within the sex market. An algorithm was created to form ties between adverts based on shared phone numbers and ASW 1 user accounts (see Kjellgren, 2023, for more details). An online network in this context, therefore, relates specifically to how adverts are connected by these two attributes. Different approaches were also tested in the early stages of this research, including measuring the textual similarity of adverts to form ties. However, based on input from human trafficking investigators as part of other research (Kjellgren, 2023), phone numbers and user accounts were deemed the most reliable and valid measures to establish links between adverts.

There is obviously limited information on the offline structure of these networks, since only their digital footprints can be observed. As such, the term ‘network’ is used to refer to how adverts are structured, rather than information pertaining to the social relations between known individuals. Therefore, the footprints of an independent sex worker can be perceived as a ‘network’, because of how the posted adverts are linked together. The use of the term ‘network’ to refer to a single node can appear inappropriate and contradictory to the normal terminology used in social network analysis. In this research context, it must be recognised that a single advert (i.e. an isolated node), may be representative of multiple individuals part of the offline network. In other words, we can never, with escort adverts as a source of data, accurately assess how many individuals are actually involved in the network. As such, using the term ‘network’ when referring to isolated nodes may seem counterintuitive, but the terminology reflects the uncertainty associated with escort adverts, and allows for the language used throughout the article to be more consistent and concise.

These networks emerge organically throughout time, as independent sex workers and other actors in the sex market advertise their services and move around throughout the country. This also means that a snapshot in time may give an inaccurate picture of the underlying complexity of a network; the first advert posted may not be particularly informative, since it will be difficult to elucidate relational patterns. Throughout time, what might have first appeared to be an isolated advert can potentially grow into a network, simply because the individual(s) in the offline network posts more adverts from the same phone numbers or user accounts.

The structure of the data used can best be described as longitudinal, relational and hierarchical. It is longitudinal because data were collected over an extended period and it is also operationalised as a relational matrix: the escort adverts are linked together, in this case by phone numbers and user accounts, to form empirical networks. Finally, it can also be considered hierarchical, since escort adverts (n = 213,693) are nested in these empirical networks (n = 15,016). The PCA was thus conducted on the network level, with variables also measured at this level (e.g. network characteristics, as opposed to advert level information).

To inductively explore the data in terms of network complexity, PCA was deemed an appropriate method for dimensionality reduction. The purpose of this was to reduce several variables that each measured different dimensions of complexity (scale, structure and geographical dispersion) into one variable, which could then be used as a basis for exploring the diversity of networks in the sample. PCA is useful when there is a set of variables measuring different aspects of the same phenomenon, which, when combined, ideally preserves the maximum amount of variability. Through linear combinations of the input variables, a smaller set of uncorrelated principal components (PCs) is derived, aimed at maximising the amount of explained variances in the latent dimensions, or PCs (Härdle and Simar, 2019). The results from the PCA suggested the first component to explain a reasonable amount of variance, while also representing a plausible way of ranking networks in terms of the relative complexity of their digital footprint. There were several iterations of the PCA, including different network-level variables operationalised in different ways. The final iteration, which included the variables shown in Table 1, was selected because of its theoretical and empirical value. This combined measurement – or PC – which will be referred to as the ‘network complexity score’ (NCS) throughout the article, is the result of the linear combinations of the values of the input variables specific to the data set used (Härdle and Simar, 2019). In other words, if the same method was applied, using the same input variables, but on a different and unrelated data set, the derived values would be different. In terms of transferability, whereas the different dimensions of complexity (network scale, geography and structure) are theoretically valid and transferable, the empirical contributions of each of the input variables for measuring complexity will be variable. As such, this work should be perceived as exploratory and an initial step to constructing a more robust, reliable and replicable scale in future research.

The combination of the variables listed in Table 1 taps into three different dimensions of network complexity: (1) network scale, (2) geographical dispersion and (3) network structure. The number of advertised ethnicities is arguably the only variable that largely represents marketing strategies, and it is also plausible that it will more frequently be used (as a marketing strategy or a reflection of reality) in larger networks. The relational measures used are, to some extent, a function of the number of phone numbers and user accounts, but more importantly, also the consistency in how they are used. In other words, networks will become more dispersed if a network frequently changes its phone numbers and user accounts.

A more qualitative and descriptive approach was deemed necessary to better illustrate how networks of varying complexity differ. Following the PCA, four networks with varying complexity scores were purposively selected for a more in-depth analysis. Because network size, to some extent, is influenced by the length of time a network is operating, it was important to choose networks that had been around for long enough to establish a presence. All networks selected had been present on the website for a year or longer. The comparison was focussed on the geographical distribution of adverts and the textual description of the adverts. This comparison also involved an ad hoc examination of digital traces associated with the networks outwith ASW 1. Some of the more prominent phone numbers were used to search for online sex buyer reviews, and other characteristics – such as the name of escort agencies – were also searched for to better understand the context of the posted adverts.

Researching the sex market is complicated and doing so through the analysis of online data poses unique challenges. First, unlike a survey, in which informed consent by research participants would be gained in advance, this was not possible with regard to the online escort adverts. However, whereas a survey would collect data on individuals, the escort adverts in this research did not pertain to individuals per se, but rather, the professional lives of sex workers and other actors in the sex market. In other words, the data collected represented digital traces associated with marketing strategies, which were posted online on publicly accessible platforms, and there is, as such, not necessarily a perception or expectation of privacy with regard to the data posted. While the data collected are not personal data per se, they were nevertheless treated as such, to protect the integrity and anonymity of the individuals posting the adverts. For the geographical figures and qualitative descriptions on networks with few cases, the details regarding locations have been modified, to protect the anonymity of the poster(s). Even so, this does not mean that the analysis of online escort adverts is unproblematic.

Indeed, the analysis of online escort adverts is inevitably a form of surveillance. UK police forces are already either monitoring or using adverts as part of their investigations into exploitation, and as established in previous sections, research using escort adverts has been conducted for several years. The problem, however, both with regards to policing and research, is that the uncritical collection and analysis of online adverts pose serious risks of causing harm to both individual sex workers and, more widely, sex worker communities. Individual sex workers risk having their adverts targeted by law enforcement as being indicative of potential trafficking. The consequences can be varied, from unwanted attention to detention and deportation. As for the wider sex worker communities, and as argued elsewhere (e.g. Kjellgren, 2022), poor research involving online data on the sex market, often motivated by particular anti-sex work ideologies, risks producing seriously misleading estimates or drawing simplistic conclusions and committing a particular form of digital fallacy: conflating online information and behaviours as reflecting an independent offline reality. Caution should always be used when analysing online escort adverts and this research attempts to both improve upon how the police utilise OSINT from adverts and, furthermore, to critically examine the value of escort adverts for social inquiries into the sex market. This research, as such, has the potential to provide a more nuanced understanding of the sex market, in which any evidence of organisation is not misinterpreted as evidence of trafficking or exploitation.

This research received ethical approval from the University of Stirling’s General University Ethics Panel on 8 October 2020 (reference number: GUEP/20/21/1007).

The results from the PCA are shown in Table 2 and they indicate that PC1 explains a reasonable proportion of the variance in the underlying dimension (0.63), which in this instance, may relate to an unmeasured (latent) dimension of complexity. The other components (e.g. PC2–6) represent other (uncorrelated) dimensions of these networks, and while not within the scope of the current analysis, may nevertheless be interesting. PC2, for instance, is strongly and positively correlated with geographical dispersion, and modestly and negatively correlated with the number of advertised ethnicities and user accounts; this could perhaps indicate highly mobile independent sex workers. With regards to PC1, however, it can be observed that all the variables are positively correlated with PC1, except for density; this is a result of how density is measured. A density of 1.0 indicates that all theoretically possible combinations of ties are formed in the network and a lower density thus indicates the absence of possible ties, which renders its network structure more fragmented.

Without explicitly measuring the network size (as could be defined by the number of adverts), the benefit of using PCA to explore complexity is that it can capture important dimensions that are not solely related to size. It is, nevertheless, quite strongly correlated with the total number of adverts posted by networks (r(15,014) = .61, p < .001) and the average number of adverts posted each month by networks (r(15,014) = .51, p < .001). PC1 can be perceived as representing one dimension of complexity, which is sensitive to the fact that smaller networks can be quite complex, yet it also captures the importance of network size in understanding processes in the online sex market. It is conceivable, though not possible to demonstrate with this data alone, that the more complex an online network appears, the more complex the organisation of the offline network may be. For instance, the use of multiple user accounts and a geographically dispersed pattern may suggest that a higher level of coordination and logistical capabilities are involved.

After extracting the first component, it was rescaled to range between 1 and 10 to aid interpretability. This measurement can be perceived as representing a ‘NCS’, in which each of the networks in the sample is ranked from the least to most complex, as defined by the combination of input variables. The distribution of NCSs is shown in Figure 1. The distribution is heavily skewed (skewness = 3.8) and can be considered leptokurtic (kurtosis = 35.6). The average NCS is 1.3, with a standard deviation of 0.4. The skewness aligns with what we would expect: most networks tend to be fairly simple, likely indicating independent sex workers or smaller collectives of sex workers, with comparatively few highly complex networks. The simplest form of network is simply an unconnected, single advert. The most complex network observed was assigned a score of 10; this is a truly large and complex network responsible for posting over 3000 adverts in the study period. As can be seen, it is considerably more complex than the other observed networks and this is one of the case study networks (network 124) presented below.

Figure 2 shows the distribution of networks by LADs. The table to the left shows the 20 LADs with the highest rate of networks per 100,000 population. First, it can be observed that all but one of the LADs (Isle of Scilly) had one or more networks in the research period. The average number of networks per local authority is 105.8 (SD = 117.3) and the median is 59.5. This, of course, does not tell us if these networks in fact were present in these local authorities; it does, however, suggest that the observed networks are marketing their services across the whole of the United Kingdom. Birmingham has the highest number of networks (1126) and some of the London boroughs also have a substantial number of networks. The cities of Manchester, Leeds, Glasgow, Leicester, Liverpool and Edinburgh also appear to be targeted by a significant number of networks.

In Figure 3, the 50 most complex networks are shown by LADs. These 50 networks can be observed in all but 45 LADs (shaded in grey). The average number of complex networks in LADs is 4.4 (SD = 3.8), with a median of 3. Some of the most popular locations of these 50 networks include Manchester, Reading, Bristol, Glasgow and Slough. In addition to these, some slightly smaller cities and locations also appear to be quite prominent, including Oxford, Derby, Dundee and Maidstone. Even though the graph only shows 50 networks, the geographical footprints are quite widespread and it is clear that some quite remote and sparsely populated areas (such as the Scottish Highlands) may also be lucrative to complex networks.

Table 3 shows a comparison of the 50 most complex networks to the other observed networks in the sample (n = 14,966). The most complex networks can be characterised by, on average, operating a substantial number of user accounts and phone numbers, posting a higher number of adverts each month, advertising multiple nationalities and ethnicities, and having a greater variability in the advertised age (perhaps indicating the presence of several individuals). In terms of geography, complex networks are, on average, more likely to have posted across a greater number of LADs and a higher average geographical distance between the posted averts. In terms of their structural composition, the most complex networks are, on average, signified by a lower network density and being more dispersed.