There are different practical methods for creating such, more-or-less stable, points of reference. In a much generalized way they can be grouped into two main categories;
Single channel sensors that 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 potential addition to this single channel, have an internal reference within the sensor itself. This second internal reference (or channel) can be an extra light source or a second detector or multiple spectral filters on the same detector. There are advantages and disadvantages with both main methods.
A dual channel sensor made by two separate emitters, where one main emitter is doing the continuous measurements and the second reference emitter will measure with a much slower time interval, maybe only once per day. The idea is that the wear and decay in the reference emitter will be negligible compared to the main one, so that the sensor can assume a fixed reference state from its factory calibration for this emitter. The main emitter might light up about 40’000 more times over the NDIR sensor life span. Receiving energy from this second emitter, and comparing it to the main, can then be used to figure out and offset the resulting decay in the signal strength of the main emitter. This doesn’t work so well in practice though since there are many more sources to drift and ageing than from emitter signal decay. Additionally this setup comes with a hit on 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 with each measurement relate effects of decaying emitter, dimmed optics and other reasons why there might be resulting changes in energy reaching both detectors. The flaw in this setup is that it assumes that the ratio in received energy between the two detectors can only be fixed will always remain the same. Which is not true over time, and additionally it worsens the Signal to Noise ratio as the finite energy output needs to be split between two detectors.
The single channel is much simpler in setup. 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 directly use known calibration gases, or they keep in memory a reference to the strongest irradiated energy received by the detector during set time periods. This memory reference will be assumed to correlate with the fresh air baseline, the lowest concentration the absorbing gas can ever reach naturally. 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 by throughout its entire lifetime. The single channel proves a simpler, minimalistic, and more efficient solution for quality measurements over time.
So when are dual channel sensors more useful? Well, over time, they tend to never be. But in the nearest time after their factory calibration, before their internal referencing starts acting up, they do have this always available (not yet harmful) internal reference to make immediate self-adjustment and calibration. This can give them great measurement accuracy out-of-the box. A single channel sensor can have poor out-of-box performance, as it will need to wait long enough to sample this sporadic fresh air baseline and complete enough such cycles to correct whatever drifts it endured during transportation and final product assembly or installation.