Just how accurate are advertising predictions produced by machine learning systems?

2-Color-Design-Hi-Res-100px-widthThanks to Mikio Braun, who on Thursday, January 2015 published an article on the InfoQ website. Braun’s article includes mention of a Google acknowledgement about the role played by machine learning (also known, at least in part, as data processing by algorithms) as a predictive tool in its ad placement technology for its click ad business. Readers interested in this topic should read Braun’s article, which is titled Google on the Technical Debt of Machine Learning.

I have written about the inaccuracy of click advertising in earlier posts to this blog. To quickly summarize my opinion on this topic, I found the systemic tendency towards poor ad placement to be especially difficult to overcome when the items to be promoted provide subjective, intangible benefit. So gaining a perspective on just how much of the ad placement technology behind Google Adwords and, in all likelihood, its direct competitors (principally Microsoft’s Bing advertising system), as Braun points out in this short article is very helpful.

What is also very helpful in Braun’s article is the manner in which the Google researchers (Braun’s article is really a news report on a presentation at a recent conference event held in Montreal, the Software Engineering for Machine Learning workshop, part of the annual Neural Information Processing Systems, NIPS, conference held in Montreal) shed light on the precariousness of proper performance for machine learning systems, in this (online advertising) context, given the effect they have on other related computer processes. These researchers make clear how the basic assumptions powering Neural Networks can actually adversely affect these siblings, and, thereby, produce erroneous results along with very little value to people depending on them. Readers should note this conclusion is my own, and not a conclusion expressed anywhere in Braun’s article.

From Braun’s article, and the technical précis of a research paper on the algorithmic process behind machine learning, which was also published by Google researchers, online advertisers should be careful to set realistic expectations about paid placements. Perhaps it will make sense to horizontally structure these campaigns, with a panoramic reach wherever possible, if they are to produce any meaningful results.

Ira Michael Blonder

© IMB Enterprises, Inc. & Ira Michael Blonder, 2015 All Rights Reserved


Just What Costs are Included in Google’s Traffic Acquisition Costs?

Google’s Q3 2013 earnings report includes an important section titled “Traffic Acquisition Costs”. Google has offered a definition of these costs. The Google definition of Traffic Acquisition Costs is very specific to the “network” component of their revenue. Google’s network affiliates receive payments based upon operational terms governed by contract.

So these costs should directly correlate to the “network” segment of any quarterly earnings report for this business. For Q3, 2013, the “network” revenue component posted a decline in revenue by approximately 1%, but the “Traffic Acquisition Costs” declined by approximately 1.3%. Is it safe to say the network affiliates are under some pressure? How else to account for a reduction in these costs exceeding the commensurate drop in revenue production for this component?

If this is the case, then analysts may want to dig a lot deeper into these numbers. If Google benefits more from its search revenue component, which operates wholly on its own websites, and, further, Google appears to be making very serious efforts to accelerate the efficiency of its network business, then perhaps the value of its pay-per-click and pay-per-thousands of impressions products has deflated more than otherwise appears to be the case. Personally I think this view, that the standard Google ad products are not delivering on their expectation, is the case.

I would add to the above information some observations on the amount of effort Google is expending to ensure the success of SMB advertiser campaigns. Via its “Engage” agency program, Google has been maintaining human support resources, accessible to any advertiser, via telephone contact, for several quarters. This program, alone, likely adds a cost component of some magnitude to their advertising products and, therefore, diminishes the profit margin.

Finally, it would be good to get some ideas as to the costs of the Analytics product. This product has a hefty price tag for enterprise customers, but the free version, which appears to be the method by which most users avail of this product, requires a lot of infrastructure and periodic systems development.

Ira Michael Blonder

© IMB Enterprises, Inc. & Ira Michael Blonder, 2013 All Rights Reserved


Promoting Tech Products to the Google Adwords Display Network with the Topics Feature

One of our clients sells help software for a popular collaboration tool widely used by enterprise businesses and large groups in the public and not-for-profit sectors. This client is promoting its products online with Google Adwords campaigns. We perform the daily maintenance required for these campaigns. Separate search network and display network campaigns have been set up for the help software product.

These campaigns haven’t produced desired results. The landing page for campaign ads was part of the problem. The editorial content for this page has been corrected. But clicks on ads for the display network campaign still exhibit too high a bounce rate (nearly 100%).

We spoke to Adwords about the problem. They recommended completely removing keywords for the display network campaign. They also said that topic targeting may provide a better way to select useful placement domains for our ads. We followed their recommendations.

They also recommended using image ads. Adwords now has a “Display Ad Builder.” This tool builds image ads for the advertiser at no additional charge. We think the ads are attractive and certainly worth testing live on the public. Each of our text ads was replaced yesterday with an image ad.

We spend most of our time with Google Analytics. Whenever we receive a click that ends up in a visit beyond our landing page, we want to know more about it, including the placement domain producing the click and the geographical location where the click originated. The Adwords team showed us how to build a custom report, which produces most of the information we need, including each page visited. Note: we still can’t identify the geographical location of the visitor for each click. We can easily determine the placement domain by running a standard report where the Primary Dimension is the Campaign Ad Group and the Secondary Dimension is the Placement Domain. This report provides us with the information we need on Pages/Visit, Avg. Visit Duration, and Bounce Rate.

In most cases, once we run our custom report for the Placement Domain identified by the standard report, we can take each of the pages included in the specific visit and run them back through another report, this time on content, by day, to determine the geographical location for the visitor. The visitor identification problem comes up when the visited page is very popular.

On March 7th the Adwords team demonstrated how we can use the “Dimensions” feature for the Display Network to expose the geographical location of the visitor. The “Customize Columns” feature, in Adwords, can be used to display City, Region, Country and Associated Geographical information.

Ira Michael Blonder

© IMB Enterprises, Inc. & Ira Michael Blonder, 2013 All Rights Reserved