What is it?
The "Attn. Time" or "Avg Attn. Time" metric in CMP measures how long an average pageview is engaged for. If you're reading a piece of content and scrolling or moving your mouse cursor then you're considered "engaged" and that counts towards engaged time. Avg Attn. Time is calculated by summing all the engaged time of all users and dividing between the number of pageviews. This gives us a rough estimate of how much time on average was spent on the page. The higher average attention time the better because people spent more time reading your content.
What counts as engaged?
When you're reading a piece of content, you're consider "engaged" if your page is in focus and you've scrolled or moved your mouse cursor in the past 5 seconds. After five seconds of not moving your mouse or cursor or not looking at the page (it's in another window or tab) we take current time and subtract the time you started being engaged from it. So if you loaded the page, scrolled/moved mouse for 7 seconds, then stopped and went to get some water, then your total engaged time would be 12 seconds because you were active for 7 and we assume you were active the last 5 seconds. Why is there a 5 second timeout? Chartbeat found that if someone is inactive for more than 5 seconds then that means they really aren't looking at the screen. If you come back and start reading, we start adding to your cumulative engaged time of 12 seconds. If you decide to then close the window we'll immediately send your total cumulative engaged time.
Why is there a spike in the chart after Sep 23rd?
We changed how we aggregate total engaged time. Previously if you managed to be active for a whole 5 seconds we would send an event indicating 5 seconds of engaged time. Instead now we send a cumulative engaged time rather than just the individual interval. This is more accurate because if you were active for 12 seconds and then left the page by closing the browser we would send 12 seconds worth of engaged time instead of having to wait until you get to 15 seconds when the timeout would have occurred. The numbers are much higher now because we collect the chunk of engaged time that was missing before.