Typical Lifespan of a Web Image: 2-3 Days

Image Entropy for April 2011

Pixazza’s chief scientist Anton Kast uses the above graph to illustrate his analysis of image entropy, a term he borrows from thermodynamics that, in this case, refers to the distribution of imageviews over time. High entropy on a particular image would mean its total imageviews are evenly distributed over time. Low entropy would mean an image gets all its attention over a short period of time, and after that initial burst is largely ignored. There are more than 3 trillion images online, and 200 million new ones posted each day to Facebook, but how should we track viewership or or engagement with those images?

“What is the real, practical scale of images on the web? Which images are people really interacting with, and how much?

“At Pixazza, we have a window into this question. Pixazza technology is applied to images that are viewed more than 100,000,000 times per day. We need to know which images are transient and which get stable traffic, because it is the basis for many of our statistics and optimizations, such as our ad targeting.”

It turns out — at least among the April 2011 data set — that most images get the bulk of their views in the first 3 days they’re online. But at the same time, taken together, there’s a gianormous amount of views on older images. Say one percent of those 3 trillion images is viewed once each day by one person. That’s 900 billion imageviews a month.

  1. # Jean/NOTCOT said: May 19th, 2011 at 1:49 am

    What about the impact of things like google images, ffffound, tumblr, etc?

  2. # Chas said: May 19th, 2011 at 9:41 am

    I think those services are driving higher entropy for images (if I’m using my physics jargon correctly!). Back in the digital dark ages, when we navigated the web by way of links on portal homepages, an image got all of its views in the day or two it was prominently promoted on a handful of large sites. Sharing sites are extended the shelf-life of photos at least somewhat: I post a picture of my kids to Facebook and a few people view it each day for a month or two. Image search has the potential to increase image lifespans further, if the search engine can associate an image with keyword-based queries. And that’s currently the rub. Unless an image’s publisher supplies detailed and meaningful metadata, or someone else (a Mechanical Turk, a delicious user, a Pixazza expert, etc) tags it with additional information, it remains a mysterious rectangle of pixels to a search algo. I think tagging is going to push more imageviews into the long tail over time.

  3. # Matt Hooker said: May 2nd, 2012 at 1:35 pm

    Chas, Is there a mirror of Anton’s analysis? The link appears to be down.

  4. # Chas said: May 4th, 2012 at 12:39 pm

    Matt–I fixed the link (thanks for telling me!). Anton’s post is here, http://blog.luminate.com/2011/05/image_entropy.html

Leave a Reply