Symmetry is central to differential pairs

Differential pairs work by making the received signal the difference between two complementary parts that are referenced to each other, so the effects of their electrically noisy surroundings are minimized. By contrast, single-ended signals work by making the signal the difference between a received signal and power or ground, so noise that is on the signal or power system does not get canceled out. This is why differential signaling is so effective for high-speed signals and why it is used in fast serialized buses and double-data-rate memory.

In a differential pair, the positive and negative sides must be transmitted through identical surroundings all the way along the transmission path. And the two sides must stay together so that the positive and negative signals couple to each other via electric and magnetic fields at the corresponding points on those signals. Differential pairs are symmetrical, so their surroundings must be symmetrical too. Perfect symmetry is, of course, unattainable, not least because of dimensional tolerances. But designers can come close to that ideal to obtain the best differential-signaling results by following a few fundamental rules.

Make sure the signals are at the same point on each line at the same time. Equalize trace sections, indicated by the same letter in the figure. If your pair has series-terminating resistors or a common-mode filter, the connections to those should be the same distance from the differential driver pin on the positive and negative sides.
Preferably route point-to-point, but in any case keep any stubs or diversions (shown as C) to within 0.6Tr inch, where Tr is the driver output rise time. Use the same length restriction for A and E in the diagram wherever possible.
Use a field solver to design the trace separation so you know the even-mode and odd-mode impedance. You need these values for designing terminations. Having a 50-ohm board doesn't mean the even-mode, odd-mode or differential characteristic impedance will be 50 ohms.
Consider the odd-mode impedance when terminating from either line to ground or a reference voltage to terminate your intended differential signal.
Also consider terminating even mode or common mode (half the even-mode value) to terminate unintended noise.
If you terminate between the two lines, consider the differential-mode impedance (double the odd-mode impedance).
Remember that radiated noise picked up from the same source is rejected only if you closely couple the pair, because only when the lines are close together are the surrounding electric and magnetic fields likely to be nearly identical.
Compensate by lengthening near the driver for any skew between the complementary output signals sourced by the driving device.
Lengthen only in a differential pattern wherever possible, remembering to balance the number and style of right-hand and left-hand bends.

When free of external interference and routed correctly, differential pairs exhibit a characteristic differential impedance and propagation delay.

Don't

Terminate considering the single-ended characteristic impedance instead of odd-mode and even-mode impedance: Closely coupled differential pairs are designed to operate in a special way for complementary signals.
Just equalize overall trace lengths instead of equalizing each section.
Route differential pairs over gaps in power or ground planes.
Forget to define differential pairs if you are using an autorouter-otherwise you will get single-ended routing.
Let test engineers add test lands in different positions on each side of the pair. Test lands behave like high-impedance device inputs and unbalance a differential pair just as easily.
Route other signals too closely parallel to a differential pair, or directly below or above on another layer-crosstalk can throw the differential signal out of balance.
Route differential pairs above or below unrelated power or ground areas (for example, a separate analog power plane).
Forget to consider where off-board connections go. Connectors, cables and differential topology on other boards in the system should all be modeled when checking your proposed circuit using simulation.
Be hoodwinked by probe or test equipment parasitic inductance and capacitance. If you put a probe on one side of a differential pair it will probably be thrown out of balance; your measurement can easily be misleading and your equipment can develop spurious faults under test conditions.