The second key area of search development was opened up by Google only after the first algorithms for reading the WWW proved successful. This second area was that of personalisation, which only became possible once Google became big enough to start collecting significant datasets on those using its search engine. Exploring these datasets enabled the targeting of search results, with different users receiving different search results. This is particularly the case if the searcher uses other Google services, such as Gmail, and has a Google account. Personalisation appears to many to be the process whereby Google judges whether a searcher who uses a term like ‘surf’ is interested in surfing on water, musical channels, or the Web and so on. It also seems to identify users individually, each having a certain age, location, gender, race and so on, bringing users the results that are judged appropriate to their demographics. However, reading personalisation in this way is to read it from the point of view of the user’s practices rather than Google’s. For the latter, the key is not so much each individual but the correlations between many individuals; it is the inter-relations that are key to producing a useful result for an individual, not the other way around. This is because the inference has to be constantly made that if many individuals of a certain type favour a particular search result then this can be delivered to individuals who fit that type. It is these kinds of mass correlations that allow for the targeting of particular groups of people – assuming, for example, that men of a particular age group might prefer the Burt Reynolds version of the film The Longest Yard, whereas those from a younger age group might be looking for the Adam Sandler remake with the same name, and those of a different nationality may be interested in the Vinnie Jones-led soccer version called The Mean Machine (Feuz et al. 2011; Hillis et al. 2012).
Personalisation achieved by building correlations between categories, or profiling as it is sometimes known, is a second way to mine social relations to create Google search (Elmer 2004). The results delivered to individuals are partially based on correlations which are meant to mathematically capture what social and cultural life means. This is not a totalising analysis which posits one set of internally consistent social dynamics, but a tracing or mapping of whatever social relations can be found from analysing the data Google collects. In this way, Google’s practices of delivering search results and generating data on which ads can be based include various ways algorithms can read the relations between people.
Starting with the social relations that can be read from the structure of the WWW, Google search then progresses through various means of manipulating and extending that reading. Once enough data has been collected, it can progress to reading the kinds of correlations that measure social relations, which may then be used to personalise search results. Google search practices intertwine different kinds of people, algorithms, datasets and constant updating and storing processes to deliver an answer to a question. These algorithmic logics, that are interweaving different kinds of actors in people, software, data, hardware and so on, must continue to deliver a successful search engine, but they must also conform to the corporate logics Google has embraced as a for-profit company.
For example, one of Google’s initial problems in profiting financially from its search engine was how to generate trust in its very different way of selling advertisements (Auletta 2011: 3–6). As already mentioned, Google runs automated auctions to allocate search words. From the advertiser’s point of view, this is about connecting their ad with the best search query or term on the best site, while for Google it is about balancing income with advertiser trust and ease of use. These interests are resolved in the specific auction practice in which the winner does not pay what they bid but only a small amount more than the second-placed bid. While this is usually less than Google might have earned, the process has the advantage of building long-term trust with advertisers. At the same time, the process removes the advertiser’s interest in bidding lower than they might otherwise due to worries about over-bidding (Levy 2011: 89–91). Here is a specific kind of practice, again automated through algorithms and networks, that mediates the search result into an advertising program that generates revenue for Google and, possibly, for the advertiser. But search had to come first: the value of Google had to be established by the practices of creating a functioning, free and attractive search engine, so that practices could then be generated connecting the value of search to the value of money.
The corporate logics of revenue and profit have to be implemented after the practices of search as value, but these logics also find that the practices of creating search can themselves be translated into and reused as practices of revenue and profit. In particular, the processes of personalisation can develop closely with those of profits derived from advertising, because in both the issue is one of using datasets to generate ways of grouping users together. The same clues that allow personalisation can be moulded to deliver targeted advertising. These Google practices consist of translating both advertiser trust and dollars through the prism of profiling users, similar to personalisation. This is not to say that personalisation and targeted advertising are exactly the same processes, only that they draw on the same idea that correlations and profiling can be generated from Google’s vast datasets of its users’ behaviour (Hillis et al. 2012). In this way, Google’s algorithmic and corporate logics intertwine.
Essential to Google’s practices is a set of algorithms that distributes ads, according to words won at auction, onto its own or other websites. It should be particularly noted that Google has access to recursions within its data, recursion being the process of creating infinite information by returning the results of an information process to itself: when the output serves as input to create a different output then information may increase exponentially and infinitely (Jordan 2015: 29–44). Here we catch sight of a corporate reason for the mass information storage and processing, because a core recursion involves using the information collected to refine searching and so to refine the delivery of advertisements. In this sense, Google’s advertising practices are essentially a recursion of some of their search practices, particularly personalisation, but with the information being delivered in the form of advertisements rather than as answers to search queries. While we know this, the nature and specificity of these practices are obfuscated as a trade secret, leading to the ‘Search Engine Optimisation’ industry, which seeks to find ways to improve clients’ ranking in search engine results by analysing and manipulating their secret algorithms (Havalais 2009).
The obfuscation of algorithms and networks will be a repeated issue when analysing digital economic practices, but it should not be over-emphasised. As demonstrated above, while we cannot follow the details, we can follow the nature of the practices that create search, partly because we know what users, in this case searchers or advertisers, are doing, and because we know the nature of what the company is doing. Closer work would have value, but in defining digital economic practices generally, or Google’s practices specifically, the level of detail that is available is more than adequate.
With this analysis of Google’s economic practices, we have thoroughly examined a leading example of digital economic practice in which the searcher, the advertiser, the advertising site and Google itself each have different practices that intersect to create the overall practice. At its core, we see that while the money comes from advertising, this revenue is dependent on prior search processes. Advertising is then second both temporally – the search engine has to first be established