iGo: Receptivity to Contextually Targeted Information (2008)

This was a project that my entire team undertook together as a collaborative effort. It was great fun, very informative, and Ellen Isaacs wrote it up and submitted it to CHI. Unfortunately, some reviewer at CHI had a quibble with the statistics (amazing because we had Nick Yee working on those and he's a known expert at data mining) so it didn't get in. This is unfortunate because it offers some very valuable lessons about contextual targeting of information. Alas, no one had the time to work on getting a paper accepted somewhere after that.

As smartphones acquire ever more sensors and systems get smarter at making inferences, everyone is anticipating the advent of contextual targeting of information and services, based on people’s situation. But a major part of the UX design challenge is to determine whether and when people want to have anything ‘targeted’ at them when they are about their usual business and what information that should be.

We decided to conduct a feasibility study, using push and pull SMS information snippet delivery and experience sampling surveys to simulate sensors; we asked about things like location, noise-level etc. at the time snippets were received.

Information snippets were interesting trivia items (e.g. “There is an ant in Brazil that has a gland that causes the ant to explode like a bomb, spraying a sticky toxic goo on everything nearby”). Users were divided into two groups for a two-week study. One had snippet SMSs pushed at them and they rated their receptivity when they arrived (the push group). The other pull group requested snippets whenever they felt like it.

Sensors on a Smartphone

Percentage of times a cognitive context of a number of different kinds was reported in surveys. For push, this represents a baseline of context frequency when information randomly pushed. Differences in percentage of pulling content for the same context indicates higher or lower preference for that context.

Detailed analysis showed that sensor data of physical context would not predict receptivity very accurately (location was the best predictor). When looking at the cognitive context (what people said they were doing at the time) we used push group data to get a baseline for activity prevalence during random snippet delivery and compared it with pull group reports of when they asked for snippets. People pulled in micro-moments while they were not otherwise occupied. These moments would be hard to differentiate from moments when they were occupied, hence sensors would not be up to the task.

Conclusion: Contextually targeted information should be queued up for users to pull rather than risking interrupting users.