CPM refers to the cost of reaching 1000 target audience persons with an ad in a media vehicle, online or offline. A $20.00 CPM on adults 18-34 means that it will theoretically cost $20.00 to generate 1000 adult 18-34 impressions. By comparing CPMs across media you can theoretically get an idea of which media are most and least cost efficient.
Sellers of online time and space often pitch low CPM as a reason for the advertiser or media buyer to buy. In reality, lower CPM is likely to be untrue. For example, those selling online video as a more cost efficient alternative to over the air or cable TV may get caught in a numbers trap. If the online video proposal is priced at $20-$50 CPM the seller may have a tough time selling against television priced for $5.00 per 1000 to broadcast buyers with far more accurate audience estimates behind the numbers.
The problem is that comparing online and offline media CPMs is kind of like comparing apples and oranges especially if convoluted audience estimates are entered into the equation.
First, the definition of audience is different for different media and the numbers are not comparable even for broadcast TV and internet video. For TV, Nielsen measures audiences in a fairly sophisticated manner using set top boxes and surveys to gather continuous ratings data, even taking DVR behavior into account. Nielsen measures real time average minute viewers on a continuous basis, 24 hours a day. Example: The Superbowl generated a 44 rating among US men, meaning that during the average minute 44% of men were tuned to the Superbowl.
In contrast, the online industry puts up a very questionable metric called "viewability" which purports to measure the number of people who have the ad on their browser for 1-2 seconds. Whatever that translates to vis a vis other media, I'm not sure, but it is definitely not a rating based on a random sample of respondents whose sets were tuned into programs around the clock.
In magazines, the audience concept of average issue audience is utilized. The measurement methodology attempts to estimate how many people saw or picked up the average issue of a magazine in home or out of home. Clearly, magazine audiences can get to be extremely large if the out of home passalong readers are counted as part of the total audience (e.g., People Magazine has 18 readers per copy). But magazine average audience is still nothing like the online definition of display ad audience, where all that is required is a theoretical 1-second view of an ad on a webpage. And by the way, what can be meaningfully communicated in one second?