Datawocky

Is Search Advertising a Giffen Good?

The Giffen good is a strange beast from economic theory.  For most goods, demand decreases as price increases. A Giffen good defies this normal market behavior -- the demand for it increases even as its price increases.

Giffen goods have a very interesting history. They were postulated originally by Alfred Marshall in his 1895 book The Principles of Economics. The classic example is staple foods such as rice, wheat, and potatoes. As their price goes up, poor people on a tight budget actually consume more of them, because they are forced to cut back on luxuries such as meat, but still need the same number of calories to survive. Until recently, Giffen goods remained a theoretical beast, with no real documented examples -- until 2007, when two Harvard economists demonstrated that rice and noodles behave as Giffen goods in certain poor parts of China.

Google's recent results raise the possibility that search advertising might be a Giffen good. Here's a simple model. Company X spends marketing dollars on two channels: search advertising and brand advertising (on the web or on TV and magazines). Search advertising drives customers directly to their site, resulting in immediate sales. Brand advertising drives organic traffic, albeit in a more unmeasurable way.

In an economic downturn,  companies get more cautious with their marketing budgets, moving more dollars into measurable and direct channels such as search advertising while cutting back on less-measurable brand advertising. Thus, there is more competition for the clicks, driving up the price (cost-per-click, or CPC) of search ads.

Company X, therefore, finds all their increased spend on search marketing actually drives the same or even fewer visitors to their site. At the same time, since they have cut back on brand advertising, organic traffic is decreasing. But wait -- we need to make this quarter's numbers! The easiest way to do that is cut back even more on brand advertising and channel even more dollars into search, which can drive immediate clicks towards the end of the quarter. Brand marketing's ROI is longer-term, while this quarter's revenue is a more pressing concern.

Witness the result: company X spends more on search marketing, driving more search ad clicks to its site, at a higher price point.  The definition of a Giffen good! Interestingly, unlike the rice-and-noodles example, the increased consumption directly leads to the increased price, because of the auction pricing model.

Google's recent results seem to confirm this hypothesis:  paid clicks increased by 20% from Q1 2007, while ad revenues increased by 40%, implying a CPC increase of 16%.  Of course, there's a limit to this phenomenon: companies cannot pay for more their ad clicks than their profit margins allow. Until that time, the sucking sound you hear is everyone's profit margins going into Google. We're going to see a lot of low-margin revenue increases at online retailers and other companies that rely on paid search.

Network effects propel cloud computing

Erik Schonfeld over at TechCrunch has a very interesting post  revealing that big enterprises are adopting Amazon's EC2/S3 services at a far faster pace than previously imagined:

A high-ranking Amazon executive told me there are 60,000 different customers across the various Amazon Web Services, and most of them are not the startups that are normally associated with on-demand computing. Rather the biggest customers in both number and amount of computing resources consumed are divisions of banks, pharmaceuticals companies and other large corporations who try AWS once for a temporary project, and then get hooked.

This is epochal stuff -- banks and pharma are notorious late adopters of early-stage technology, so to see them in the vanguard of cloud computing (or perhaps I should say utility computing, but everyone says cloud) is astonishing. But it illustrates a very important detail that's been overlooked: that there are significant network effects to the cloud computing business.

There are two basic underlying forces behind the network effects:

1. Code that works with large amounts of data needs to be close to the data (in the network topology sense).
2. Any processing that consumes data generates data.

So, once one enterprising group within a company decides to place some data in S3 and the code to process it on EC2, it becomes a whole lot easier for someone else within the company who needs to run some other code on the data, to move the processing to EC2. And since all this processing generates even more data, we have a virtuous cycle building up. The stable state is for all of a company's data processing tasks to move into the same utility computing cloud, to take advantage of the co-location and minimize data transfer latency and costs.

The network effect extends across companies as well. Often data created by company A is consumed by company B. When this "data interface" is voluminous, it makes economic sense for company B to move into the same utility cloud as company A. There are some ecosystems where utility computing players are already exploiting this trend; for example, AppNexus is creating a utility cloud optimized for the use of ad networks and their associated ecosystem: analytics for publishers and advertisers. There is so much data being shared here (on ad campaigns and their performance) that there is significant advantage to being in the same cloud.

The network effects argument leads to the interesting possibility that cloud computing becomes a winner-take-all game, like auctions; we might end up with one winner (maybe Amazon?) A more likely outcome is, we might end up with a couple of big general-purpose clouds (Amazon and Google, perhaps?) and a few niche clouds optimized for different ecosystems (such as ad networks and social networks).

More data beats better algorithm at predicting Google earnings

Readers of this blog will be familiar with my belief that more data usually beats better algorithms. Here's another proof point.

Google announced earnings today, and it was a shocker -- for most of Wall Street, which was in a tizzy based on ComScore's report that paid clicks grew by a mere 1.8% year-over-year. In the event, paid clicks grew by a healthy 20% from last year and revenue grew by 30%.

In comparison, SEM optimizer Efficient Frontier released their Search Performance Report on their blog a few hours ahead of Google's earnings call. EF manages the SEM campaigns of some of the largest direct marketers, handling more SEM spend than anyone in the world outside of the search engines themselves. Their huge volumes of data give them more insight into Google's marketplace than anyone outside of Google.

EF reported a 19.2% increase in paid clicks and 11.2% increase in CPCs at Google Y-O-Y. Do the math (1.192*1.112 = 1.325), that's a 32.5% Y-O-Y revenue increase. That's the closest anyone got to the real numbers!  And this quarter is not a flash in the pan: in January, EF reported a 29% Y-O-Y increase in SEM spend, with 97% of the increased spend going to Google: that is, about a 28% Y-O-Y revenue increase for Google. That compares very favorably with the actual reported increase of 30%.

As Paul Kedrosky points out, this is a huge indictment of ComScore's methodology (ComScore's shares are trading down 8% after-hours post the Google earnings call). ComScore sets a lot of store on their "panel-based" approach, which collects data from a panel of users, similar to Nielsen's method of collecting data on TV viewing using data from a few households that have their set-top boxes installed. ComScore has been in this business longer than anyone else, and has arguably the best methodology (i.e., algorithm) in town to analyze the data. They're just not looking at the right data, or enough of it.  Some simple math using the mountain of data from EF handily beats the analysis methodology developed over several years using data from a not-so-large panel.

To my mind, this also puts in doubt the validity of ComScore's traffic measurement numbers. For websites where I personally know the numbers (based on server logs), both Quantcast and Hitwise come far closer to reality than ComScore. The latter two don't rely as heavily on a small panel. ComScore's value today is largely driven by the fact that advertisers and ad agencies trust their numbers more than the upstarts. Advertiser inertia will carry them for a while; but a few more high-profile misses could change that quickly.

Disclosure: Cambrian Ventures is an investor in EF. However, I don't have access to any information beyond that published in their public report.

Can SMS be a publishing medium?

When we think of SMS (Short Message Service), we think of short  text messages sent between friends, or to small groups via services like Twitter. This is how the Internet started too -- primarily as a communication medium. But soon the World-Wide Web made the internet a publishing medium as well.

The question is, can SMS messages to mobile phones become a free, ad-supported publishing medium? It seems unlikely, given the limitations: text messages must be no longer than 160 characters. In addition, they also cost both the sender and the receiver.

Necessity, however, is the mother of invention. By a coincidence of three factors, India seems to have the perfect conditions for the emergence of SMS publishing:

• Huge cell phone penetration -- 246 million and counting.
• Very low internet penetration -- about 20 million internet connections.
• A cost structure where senders pay for text messages but recipients don't.

I met today with Rajesh Jain, whose company Netcore offers a service called MyToday in India that is is effect the first SMS publisher. Rajesh Jain is an Internet pioneer, having started India's first internet portal, indiaworld.in, back in 1994 -- he sold it for over $115 million in 1999, in what was perhaps the first big Internet deal in India.

MyToday "publishes" a number of "MyToday Dailies", on topics such as News, Cricket, Health, Gossip, Local News, and so on. You subscribe to one one of these dailies by sending a text message to a published phone number from your cell phone. Subsequently, you will receive a daily (or sometimes, more frequent) text message "articles" on the topic you indicated interest in. The articles are editorially assembled, and also include an embedded ad. How, you might well ask, can you fit in a story and an ad within the 160-character limit? There are two parts to this answer:

1. You can do a lot within 160 characters, if you try really hard and your expectations are not very high. I remember fitting very playable video games into 16K of RAM on my Sinclair Spectrum when 48K was way too expensive. I looked at some of the MyToday examples, and I'm impressed.
2. It's possible to split stories across multiple text messages; more and more phones have the ability to handle messages larger than 160 characters; the SMS transport breaks them up into 160-character chunks and the phone re-assembles the message.

The real economic key to this whole enterprise, though, is the cost structure of SMS in India. There you don't pay to receive a message, only to send it. So for subscribers the service is entirely free, after the initial message to subscribe. And the numbers prove it. MyToday has over 3 million unique subscribers, each with 3 subscriptions on average. They send more than 10 million text messages a day, making them India's largest sender of text messages.

For MyToday, ad rates today are still below the cost of sending a text message, but text messaging rates -- especially for huge bulk purchases -- are falling dramatically, so it's not hard to see the crossover happening next year. And that could usher in the era of the SMS publishers.

The story behind Google's crawler upgrade

Alon Halevy and Jayant Madhavan have a post on Google's Webmaster blog disclosing that Google is now harvesting data that is hidden behind HTML forms. I'm very satisfied about this, because it's something that I had a hand in making happen.

I've known Alon since 1995, when he was a researcher and I was his summer intern at Bell Labs in New Jersey (those were the days when Bell Labs was still relevant to Computer Science research). Until that summer, Alon and I had been working independently of each other on the topic of integrating information from across different data sources (such as websites) -- Alon at Bell Labs and I at Stanford, for my PhD dissertation. We put together our ideas and created the Information Manifold, an information integration system that introduced a key idea: you could describe each data source (declaratively) by a local data schema; map that schema to a global schema that unified all the sources; and process queries across the global schema.

This task is complicated by the fact that each data source exposes different query processing capabilities, by virtue of the different HTML forms it uses. We came up with some simple solutions to this problem, but several difficulties remained. In the meantime, Alon and I went our ways with our careers; Alon joined the CS faculty the University of Washington, while I started a company, Junglee, that applied information integration ideas to build the first comparison shopping engine and online job bank. Junglee was acquired by Amazon.com in 1998; in 2000, I started a Venture Capital firm called Cambrian Ventures and in late 2004 a startup, Kosmix.

Alon and I kept in touch through all this, and in 2004, he and his student Jayant Madhavan had made great progress on "schema matching", a research area that has application to the HTML forms problem we had first encountered in the Information Manifold. I was very happy to egg them on to commercialize their breakthrough, and to provide the funding for the resulting company, Transformic, through Cambrian Ventures.

Between 1995 and 2005, Web search had become the dominant mechanism for finding information. Search engines, however, had a blind spot: the data behind HTML forms. Called the Invisible Web, this data is often estimated to be even larger in size and usefulness than the "visible web" that web crawlers usually index. The key problems in indexing the Invisible Web are:

1. Determining which web forms are worth penetrating.
2. If we decide to crawl behind a form, how do we fill in values in the form to get at the data behind it? In the case of fields with checkboxes, radiobuttons, and drop-down menus, the solution is fairly straightforward. In the case of free-text inputs, the problem is quite challenging -- we need to understand the semantics of the input box to guess possible valid inputs.

Transformic's technology addressed both problems (1) and (2).  It was always clear to us that Google would be a great home for Transformic, and in 2005 Google acquired Transformic -- a nice return for Cambrian, but also a great place for the Transformic team to make a real difference with their ideas. The Transformic team have been been working hard for the past two years perfecting the technology and integrating it into the Google crawler. I'm very happy to have played a small role in their success story.

An aside on the world of academic research. Alon and I had some difficulty getting our Information Manifold research published: it was rejected the first time we submitted it to a leading academic conference; we had to address a lot of criticism and skepticism from the establishment, and it was finally published at the VLDB conference in 1996. Remarkably, this paper has since become one of the most cited and influential papers in its field (see this survey, page 52). In 2006, the paper received the 10-year Best Paper Award at VLDB, given retrospectively to the publication from 10 years ago that made the most impact. The moral of the story is, sometimes it pays to swim against the current in academic research; what is fashionable today is rarely what will ultimately make the most impact.

Affinity and Herding Determine the Effectiveness of Social Media Advertising

A recent piece in The Economist raises a provocative question: social networking sites such as Facebook, MySpace, and Bebo have grown tremendously in usage, but are they viable businesses? In other words, is it possible to monetize these services in an effective fashion? To answer this question, it helps to take a step back and look at the monetizability of social media as a whole.

Since most social media sites rely on advertising revenues, let us restrict ourselves to advertising as the monetization mechanism. Regardless of the model (CPM, CPC, CPA), advertisers value three key measures: reach, frequency, and targeting. Many social media sites certainly score high on reach and frequency, but how do they fare on targeting? Targeting is key, because it determines the CPM rates advertisers are willing to pay. And CPM rates vary very widely: from $16-20 for TripAdvisor to $0.10 for Facebook and MySpace. See, for example, this media plan.

What drives such a wide divergence in CPM rates among social media sites? Are the low rates at social networking sites a transient aberration, with higher rates around the corner as advertisers get more comfortable with the medium? And is there a simple model to predict the targetability of different forms of social media?

Remarkably, there appears to be a single factor that explains a great deal of the available data. Consider the difference between a Facebook profile and a TripAdvisor travel review. A typical pageview on the former is by someone known very well to the creator of the profile – a close friend or acquaintance. On the other hand, a TripAdvisor travel review is seen by people completely unrelated in any way to the person or persons who wrote the reviews on the page.

We quantify this distinction with a measure called affinity. The “affinity” of a social media service is the average closeness of relationship between a content creator and someone who views that content.  The affinity of Facebook is very high, while the affinity of TripAdvisor is very low.

Here’s the key observation: There is an inverse relationship between the affinity of a social media service and its targetability. Why is this true? The act of viewing a Facebook profile gives us very little information about the viewer, other than the fact that she is friends with the profile creator; when someone views a TripAdvisor travel review, she is definitely interested in traveling to that location.

I estimated the affinity of several forms of social media, and plotted affinity aginst CPM (which I used as a proxy for targetability). The resulting graph (click for a larger image) shows the landscape of Affinity versus Targetability for several forms of social media. Some of these data points are from published data and others are extrapolated. We can see that there is a strong inverse proportionality, with a couple of outliers. We’ll get to the outliers in a moment; for now, note that Social Networks and Photo Sharing sites are even higher affinity (and therefore lower targetability) than email. This is because we often email people we don’t know or know only in passing. Instant messaging has the very highest of affinities: my IM buddy list includes only my very closest friends, who I trust with the ability to interrupt me any time of the day.

What about the outliers? Video sharing sites, such as YouTube, have low affinity, because the majority of people see videos posted by people they don’t know. However, the targetability is lower than we would expect, because of a compensating factor: herding. Most people see videos featured on lists such as “Most Popular”, which reduces the targeting value of such videos. This is also true of social news sites such as Digg.

A couple of caveats:

• This is a broad brush-stroke, and individual services might well differ from the overall category. For example, popular blogs have much lower affinity and therefore much higher CPMs than the typical blog.
• Targetability is not the only factor determining CPM; there are others. For example, certain viewer intents are inherently more valuable than others.

But with these caveats, this simple model is highly instructive. We may conclude that, when all the dust settles, the CPM rates of instant messaging services will not exceed those of social networks, which will not exceed those of email. These are inherently low CPM businesses.

What can social media sites do to increase their CPMs? There appear to be two options:

• Create sections of the network that are more topic-oriented, and less about individuals. For example, band pages and groups on MySpace, and Facebook groups.
• Mine individuals’ profiles, or their off-site behaviors, to target them behaviorally rather than contextually. This approach carries with it dangers of privacy violations, as the Facebook Beacon fiasco demonstrates.

If social networks are to become a viable business, they need to pursue aggressively one or both of these approaches. Of course, it may be possible for some services to sidestep this question entirely and develop business models that don’t depend on advertising. We haven’t seen such a model emerge yet, but there is so much creativity and ferment in this space that it might just happen.

Update: I received some questions about the affinity versus targetability landscape. Here's a brief description of the methodology. I used published CPM numbers where they were available; e.g., Yahoo Mail ($3-4), TripAdvisor ($16), Facebook ($0.10-0.15). Note that published CPMs are generally to be taken with a pinch of salt, since they may apply only to small portions of the overall publisher inventory and not represent real market-clearing prices e.g., Google's stated goal of $20 CPM for YouTube -- only a very small number of YouTube videos show ads today. I've used Metacafe's $5 net CPM payout to video producers as a more reasonable benchmark -- this likely represents a gross CPM of $10 assuming a 50% rev share. For blogs, the numbers are all over the place: BlogAds ratecards for various blogs vary from $1-$4CPM, Valleywag reports $6.50-$9.75, and Federated Media has ratecards charging $7-$40. I took $10 to be a median for blogs with reasonably high traffic. Some of the other data points are based on guesses and informal conversations, since these sites typically don't publish their CPMs. Please email me if you have additional data on these; I will update the graph accordingly.

More data usually beats better algorithms, Part 2

The post More Data Beats Better Algorithms generated a lot of interest and comments. Since there are too many comments to address individually, I'm addressing some of them in this post.

1. Why should we have to choose between data and algorithms? Why not have more data and better algorithms?

A. There are two parts to the answer. The first is resource limits; if, for example, you are a startup that has to choose where to invest resources -- in gathering or licensing data, or hiring the next PhD -- then I think beyond a certain level you are better off getting more data. If you a startup that doesn't need to make that choice, more power to you.

The second reason is more subtle, and has to do with the structure and computational complexity of algorithms.  To a first approximation, simple data mining algorithms (e.g., page rank; market baskets; clustering)  take time roughly linear in the size of their input data, while more complicated machine learning algorithms take time that is quadratic or even cubic in the size of their input data (e.g., Support Vector Machines; Singular Value Decomposition). As a simple rule of thumb, simple linear algorithms scale to large data sets (hundreds of gigabytes to terabytes), while quadratic and cubic algorithms cannot scale and therefore need some data size reduction as a preprocessing step.

For those familiar with Computer Science basics, scalable algorithms involve only  a fixed number of sequential scans and sorts of data (since large data sets must necessarily reside on disk and not RAM). Most algorithms that require random access to data or take time greater than O(N log N) are not scalable to large data sets. For example, they cannot easily be implemented using methods such as MapReduce. Thus, choosing a more complex algorithm can close the door to using large data sets, at least at reasonable budgets that don't involve terabytes of RAM.

2. Does your argument hold only for the addition of an additional independent data set (as in the IMDB/Netflix case), or does it hold also when we add more of the same kind of data (e.g., more user ratings in the case of Netflix)?

Adding independent data usually makes a huge difference. For example, in the case of web search, Google made a big leap by adding links and anchor text, which are independent data sets from the text of web pages. In the case of AdWords, the CTR  data was an independent data set from the bid data. And Google went even a step further: they became a domain registrar so they could add even more data about domain ownership and transfers into their ranking scheme. Google consistently has believed and bet on more data, while trumpeting the power of their algorithms.

You might think that it won't help much to add more of the same data, because diminishing returns would set in. This is true in some cases; however, in many important cases, adding more of the same data makes a bigger difference than you'd think.  These are  cases where where the application sees the data embedded in a very-high-dimensional space. For example, in the case of Netflix, we can  think of dimensions as users (or movies); and in the case of web search, the dimensions are k-grams of terms.

In such cases, the available data is usually very sparse and has a long-tailed distribution. For example, in the Netflix example, we have ratings available for less than 1% of all movie-user pairs. We can add lots of ratings data before diminishing returns becomes a real issue -- say add 10 times more. In such cases, adding lots more of the same kind of data makes a big difference.

3. Can you provide more specifics about the algorithms used in the student projects?

I don't plan to do so at this time, because I don't have permission from the student teams and don't wish to jeopardize their chances in the Netflix competition.