There are different practical methods for creating those kinds of stable points of reference. In a generalized way, they can be grouped into two main categories:
Single channel sensors will, at best, refer their internal energy signal transmission towards a reference point external to the sensor itself, like a known calibration gas mixture or the natural fresh air baseline.
Dual channel sensors, in addition to the single channel, have an internal reference within the sensor itself. This second internal reference (or channel) can be an extra light source, a second detector, or multiple spectral filters on the same detector.
Both categories have advantages and disadvantages.
A dual channel sensor is made by two separate emitters, where one main emitter is doing the continuous measurements and the reference emitter will measure at lesser time intervals (e.g. once/day). The idea is that the wear and deterioration in the reference emitter will be negligible compared to the main one so that the sensor for the reference emitter can assume a fixed reference state from its factory calibration. The main emitter could light up about 40,000 more times than the reference emitter over the NDIR sensor lifespan. Receiving energy from the reference emitter, and comparing it to the main one, can then be used to predict and offset the resulting deterioration in the signal strength of the main emitter. However, this does not work very well in practice, since there are many more causes to drift and ageing than only from emitter signal deterioration. Additionally, this setup comes with a higher price and makes for more cramped electronics or larger total sensor size.
A dual channel made by two separate detectors, where one detector will see the spectral absorption of the measured gas and the second, reference, detector will always, and only, measure with a spectral filter where no gases are absorbing anything. This way, the sensor can relate the effects of a deteriorating emitter (such as dimmed optics or other reasons that could be resulting in changes in energy reaching both detectors) with each measurement. The flaw in this setup is that it assumes that the ratio in received energy between the two detectors can only be fixed (i.e. will always remain the same), which is not true. Over time, the Signal to Noise-ratio worsens due to the finite energy output, which needs to be split between two detectors.
The single channel is much simpler in its setup. There is no internal referencing between channels that can go wrong. The only references these sensors will use are external ones, which will then include all possible states of decay and drift from all the individual components internally. Either they use known calibration gases directly or they keep a reference of the strongest irradiated energy received by the detector during a set time period in memory. This memory reference will be assumed to correlate with the fresh air baseline (i.e. the lowest concentration that the absorbing gas can naturally reach). When exposure to this baseline is likely to repeat sporadically, the sensor will repeatedly have this near-fixed reference point to compensate all its internal components throughout its entire lifespan. The single channel proves a simpler, more minimalistic, and more efficient solution for quality measurements over time.
So, when are dual channel sensors more useful? Well, over time, they tend not to be, but during the time closest to their factory calibration, before their internal referencing starts acting up, they do have the always available (not yet harmful) internal reference to make immediate self-adjustments and calibration. This can give them great measurement accuracy out-of-the-box. A single channel sensor can have poor out-of-the-box performance, as it will need to sample the sporadic fresh air baseline and become complete enough through such cycles to correct whatever drifts it endured during transportation and final product assembly or installation.