The thing you have to remember is that the areas of sensitivity for each focus point are larger than the representation of those points in the viewfinder. This is especially true when using zone focus. The camera will focus on the area of highest contrast within the entire area of sensitivity. This will not necessarily be the area directly behind the little square you see in the viewfinder.

This is a map of the 7D focus system. The 70D is almost identical. To see a full explanation of the data on this chart, please see https://photo.stackexchange.com/a/41179/15871

---

From the comments:

It should be noted that the 7D AF system is known to have precision problems. I have used the 7D for a couple of years now, and because of the actual design of it's AF sensor, small differences in the actual focal plane for a selected point or zone (i.e. 4.8ft vs. 5ft) are pretty much to be expected. You also get a slight amount of jitter when focusing with the same AF point at the same spot on the same subject over and over. AF won't remain "locked"...it shifts forwards and back by a very slight amount each frame. Using AI Servo, I tend to take bursts of 3-5 shots every time to make absolutely certain I get at least >one frame in sharp focus. Ironically, while the AF system is one of the 7D's (and now 70D's) biggest draws, it is also it's achilles heel...it just isn't precise or consistent. The only way to fix that problem is to move up to a better camera...the 5D III or 1D X, which have very high precision and consistent accuracy. – jrista♦

I've found that if you manually run the focus past infinity before each shot you get more consistent results with shot to shot variation because the focus is always moving in the same direction.

If you have a larger phase differential, then yes, it is usually more accurate. However, when you are tracking a subject with AI Server AF and shooting at higher speed, that is simply not an option. I've tracked birds and wildlife, and had sequences that range from five frames to thirty frames long. You would expect, so long as the subject remains underneath the selected AF zone or expansion point, that it would remain in sharp focus for each frame. Sad fact is that the 7D (and apparently 70D) is simply not capable of that kind of consistency or precision...it jitters. Because of this issue, I have seriously considered getting a 1D X for its high frame rate and phenomenal AF system. I have held off so far, in hopes that the 7D II will have an improved and more accurate AF system, maybe something like a 41pt AF sensor akin in design to the 61pt system found in the 5D III and 1D X. For anyone who needs consistent accuracy and precision for action work, Canon's 19pt AF system really doesn't cut it in the long run. (It seems AMAZING when you move up from a 9pt system, but for critical work, it just has that annoying flaw that really ruins the whole camera.)– jrista

Yes, the 7D shot to shot standard deviation is a little over twice that of the 1D X. But part of some folks trouble is that they don't realize that the camera is going to focus on the area of highest contrast of all areas in the frame that are active. For any given focus point and especially when using Zone AF or AF Point Expansion the areas of sensitivity are far larger than the squares for each point in the viewfinder. See Andre's blog

I find that the more I use it the better I get at "seeing" when an area of higher contrast may be lying just inside the zone of sensitivity (but well outside the viewfinder square) of a given focus point. The 1D X and 5D III, while they do have much lower shot to shot variation when using newer, more accurate lenses, they also have this 'overlap' where adjacent focus points share real estate on the focus sensor array. With older lenses (the watershed seems to be the EF 70-200mm f/2.8 L IS II released in early 2010), the 1D X and 5D III have about the same shot to shot variation as the 7D/5D2.

Update: The 7D Mark II does have a more consistent AF system than the 7D, but it still isn't as consistent as the 5D Mark III (I use both often). The narrower baseline required by the smaller mirror seems to be the limiting factor.