Angela J. Hill

Last week, I gave a presentation on Improving Marketing ROI Through Web Analytics for the Silicon Valley chapter of the American Marketing Association. It was a great group of people and a lively discussion. My presentation centered mainly on building a story around web analytics, and specifically bringing the story of the customer lifecycle to life through the integration of analytics throughout the cycle.

One of the things I got to do, since this was more of a personal viewpoint presentation, was use this slide:

Tibetan has different symbols for numbers, which came as a surprise to me since every other language I’ve encountered uses the same number symbols, including Japanese and Mandarin. Perhaps they *have* their own, but on signs and in printed materials, I’ve seen the ones I’m used to. A few of the Tibetan symbols are similar to ours, but that can be misleading: the one that looks like a 4 is actually a 5, for example.

Anyway, somewhere along the line of presenting scorecards to clients, I realized that when we numbers people look at spreadsheets, we immediately see patterns, key changes, etc. jumping out at us–but it’s just not that way for most people. For the uninitiated, it looks like the above–a jumble of crap and some words that are easy to gloss over. Anyway, I used this slide to demonstrate that, then switched back to the usual numbers and text and went on to talk about monetization. I think it made the point, and got people to engage a bit more with the following slide (with legible data) than they might have otherwise.

…now come in gold as well as silver. I heart the stamp. I don’t heart cutting out the card stock and doing the stamping, but I do heart the results. With tonight’s cutting/stamping extravaganza, the card drought has officially ended.

Let’s talk about web analytics. After all, I am a Math Detective. An International Math Detective.

One mysterious thing that comes up again and again is Direct Traffic. Technically, it’s anything that doesn’t have any information in the referrer field of the request. In theory it accounts only for people who type your URL into their browser, or arrive via a bookmark. In practice, there are a LOT of other ways for the referrer to get stripped from a request. I keep looking around for these lists, so I’m compiling my own here. I reserve the right to update this when I find more.

1. URL typed directly into the browser
2. Visit came from a bookmark
3. Came via link in an email–though if it’s webmail, the webmail domain will show up as a referrer
4. Came from a link in a document
5. The origin page is secure (https) and your page is not (http)
6. The link to your site was via javascript or Flash, and the viewer was using IE
7. The link came from a page behind a proxy or firewall that strips referrers (like an intranet)
8. The visitor set their browser to strip out referrer information
9. Another site is calling your content via an iFrame
10. Some of your site’s pages aren’t tagged, and the visit came from one of those (referrer traffic from your own domain may show up as direct traffic)

So, the question becomes, how do you figure out which of these is applicable? Here are some ideas:

1. In Google, search for “link:www.yourdomain.com” to get a list of links pointing to your site (as indexed by Google. Research those to see if any of these options apply.
2. If there was a sudden increase/decrease in direct traffic, look at your referrers report for a corresponding sudden decrease/increase. Site technology on a major referrrer may have changed.
3. Review your New vs. Returning visitors percentage. In theory, people who have bookmarked you should always look like returning visitors (unless they’re deleting cookies). If your domain isn’t immediately guessable by someone that’s never visited, then theoretically most direct traffic should be returning. If the numbers don’t match up, there may be other forces at play.

Taking things in a new direction, I started modulating pitch based on daily visits. (Well, technically, average hourly visits by day.) When I showed it to Gabe, his response was, “Are you talking to aliens or something?”

This one modulates the pitch based *changes* in visits per day. When you hear a blip of a high note, it means that day had a big jump in visits.

This one modulates the pitch based on the raw number of visits per day. This data spans a few years, so as the site gets more and more visitors over time, the pitch goes steadily up, with some seasonal regressions.

LOTS more to be done, but I am SO into the sound.

I have a client with a sort of complex setup. We host most of their site, so we can manage URLs and page titles. However, their press releases are hosted offsite, where we have very limited control. Unfortunately those are files on which they need very specific reporting. They need to see aggregate page views to these pages as well as be able to look up views to specific press releases. The page URLs are unintelligible, and the names have nothing in common, but all of these pages are in the same folder on this offsite host.

I can get aggregate numbers in the Content Drilldown report, since they’re all in the same folder. However, in the Content by Title report, the client gets frustrated at having to search for the desired page name, because there are so many other page names on the site–and since the press release page names don’t have anything in common, we can’t do a search for the common term to reveal all of the press releases.

To get around this, I needed to create a profile and filter that preserves the folder structure of the press releases, but replaces the file name in the URI with the page title. This will let us use the Content Drilldown report for everything, and when we navigate into that press release folder, we can see the page names instead of the unintelligible URLs. Bam. Best of both worlds. Continue Reading…

Bayesian search theory uses the work of mathematician Thomas Bayes to find lost objects–particularly objects lost at sea. For example, submarines.

What’s great about this method is that it works with hunches. In the search for the USS Scorpion, John Craven used Bayesian search theory, along with Vegas-style rounds of betting by a group of experienced submariners, to construct a theory about where the Scorpion could be found. The key elements of Bayesian search theory are:

1. Creating a variety of hypotheses, and probabilities, about where the object might be
2. Determining the likelihood of finding the object in each of those places, assuming it’s there (i.e., if the water’s deeper, it’s harder to find)
3. Multiplying these two together to make a probability map
4. Continuously revising the map as the search is conducted (i.e. when it’s not found in the first location, it’s more likely to be found in the second location, and so forth)

What does this have to do with anything, you ask? (But…you may ask that about nearly any post on this blog…) First, this technique is used in prediction markets and spam filtering online today. But second, I think there’s no reason that the concept can’t be used in just about any type of research where we’re looking for something within a finite amount of area (physical or conceptual). In that case, the keys would be:

1. Defining the area in which you’re searching
2. Developing good hunches
3. Quantifying the strength of those hunches
4. Knowing your subject matter well enough to determine the feasibility of the hunches
5. Doing the math

I’m going to try this the next time I’m looking for predictive analytics relationships!

Last night I fell asleep reading Chaos: Making a New Science, which I’ve posted about before. I woke up with that didn’t-I-have-a-really-cool-dream,-oh-crap-what-was-it-about?-I’m-losing-it!-No,-I-remember! feeling. I opened up the book–I’d been reading about circadian rhythms and had had a brainstorm about applying that science to web analytics.

The basics: circadian rhythms are built into plants and animals and keep us going through a 24-hour-ish period of activities (for a while, at least) even when external cues (i.e. light, temperature) are removed. Everyone has a slightly different internal clock, so if light and temperature were held steady for a long long time, we’d drift out of phase with each other. However, the cues (particularly light) sort of re-set us every day so that we’re all on the same cycle. In the chaos book, they were discussing some research done with some kind of animal (it was late…I’ll look it up later) where they isolated them from those regular exogenous cues, then showed flashes of light at specific points in the circadian cycle, just to see what would happen. What happened?
Continue Reading…

I posted earlier about my desire to listen to analytics data. So far I have barely scratched the surface of this, but just getting the system up and running is pretty inspiring. Now it’s time to dig into the data processing and analysis. By the way, to get this to an audible frequency I have to interpolate some values, so every time I do any processing, I go back to the source, process, and then interpolate. Anyway. So far what we have is this:

First, hourly data for a website stretching back to 2003. What I like about this is that it clearly shows some seasonalities, through repeating rhythms, in the data–this site gets big boosts around Thanksgiving and Christmas. Interesting, if you ask me.

Continue Reading…

My brother Adam recommended this book: Chaos: Making a New Science by James Gleick. It was the first book to introduce the principles of chaos theory to the general public, and it’s FASCINATING. A really readable description of some very advanced mathematics.

I find myself drawn to population dynamics, which is really not the focus of this book at all, but does get referenced several times. The more I think about this, the more likely it seems that techniques used to predict populations could extend almost seamlessly to web analytics. We already see that people online behave as populations (and we can segment visitors into all sorts of populations of our choosing). And many of the graphs of population groups look like they could be web analytics data if the axes were re-labeled. Case in point:

Continue Reading…

Ok. It works like this.

I majored in Music Technology (technically composition with that emphasis, but whatever). I’ve studied a lot of music synthesis techniques, sometimes from a music class and sometimes from a physics class. I also deeply, deeply love numbers, and would have gone for a math minor had I not graduated early.

All tonal music is composed of periodic waveforms. The frequency of the waveform determines the sound’s pitch, and the shape determines the timbre. I’ll do a basic intro to the physics of sound at a later date; for now there’s a pretty good description at numbera.com from whom I’ve also borrowed this image:

Which, it seems to me, looks kinda like this:

Which got me to thinking: what would that analytics data sound like, if I could listen to it?

Continue Reading…