Personalized Google Ads—Relevant?

Target Group

I have faced a lot with the personalized ad system of Google in the past few months. During the summer I was searching for something for a lot of time, and I recognized, Google started to display personalized Google ads on the pages visited—mostly after I have bought that things abd was therefore no longer interested in purchasing. Later the issues repeated: a couple of month ago I looked for a tool assisting in SharePoint migration and not that long ago I evaluated U0 development su0tes. Though the migration of SharePoint has been finished and the UI component has also been selected, everywhere I go I’m presented ads about SharePoint migration tools and UI components suites.

I do not think I know the search or ad business better than guys at Google, but I really do not understand why rendering completely irrelevant [relevance-lost] ads instead of trying to find out what is relevant for me… But how?

A Simple Tip for Procedurally Predicting Upcoming Relevance

Here I’d like to share a simple idea on predicting what may be relevant for a user at a certain point in time to enhance relevancy of personalized content delivered, just for the sake of curiosity.

Let’s say a topic is an element of a taxonomy to which ads can be assigned, and it may consist of one or more phrases or search terms. An example for the topic is SharePoint migration, which may consist of searches like SharePoint migration, Share Point migration and SharePoint move etc. Say our observed user u₀ searched a lot about a topic t₀—where searching a lot may be defined as the number of searches in the topic per day or other observation period, or the ratio of searching for a given topic within the set of all of the topics she searched for during the observation period.

If we recognize the user stops searching for t0 (or, being more predictive, loses interest in t₀ by running less and less searches), presenting t₀ ads becomes superfluous, since either she has purchased a product or service associated with t₀, or she canceled such intentions of her. Instead we should try to find out what she will be interested in next.

A straightforward idea is to analyze our data, look for people searching for t₁ in the past, and checking what they began to search for after t₀. So get a sample of topics from users u₁, …, u_n which were in their focus after loosing interest in t0. Say the search history of u₁ is t₀, t_1,1, t_1,2 and t_1,3 (h₁ in short, topic history of u₁). That of u₂ is t₀, t_2,1, t_2,2 and t_2,3 (h₂); and so on for the other users. If we have a sufficiently large sample, which can be assumed once we try to sell via personalized ads, chances are that we will be able to find correlations between users’ post-t0 search habits. Generating the intersection H of h₁, …, h_n will yield in the common interests. If we add weighting to the topics, for example adding bigger weight for topics coming into focus sooner after t0), we can also predict an approximate order of topic shift. Say t_x is an element of H, and is of the largest weight among other elements of H, and we have found a potential topic which likely is or will soon be relevant for u₁.

The Result

I think Google’s analysis is heavily based on statistical approaches rather than such procedural concepts described above, but if Google were employing such an idea, nowadays I would presented with ads about SharePoint 2013 Web Parts and code refactoring tools or data access components, which were more relevant for me than ads about something I no longer tend to purchase. This is based on the assumption that if other guys looking for SharePoint migration and UI components later searched for Web Parts or other types of developer components, since they wanted to enrich their migrated SharePoint site or get rapid tools for other areas of software development, than I may also be interested in this topics.

_{(Cover artwork by Svilen Milev)}