You can tell a surprising amount about the genome from a subset of SNPs. Most germline variants occur in groups, so you only need to measure one position to know the values for that group. The variants that are chosen for the array based tests are carefully chosen to give you the most amount of information possible. If you’re interested in this, look up “genome imputation”.

For most of the tests that 23andme would want to run on a population, like risk assessment, you don’t get much more information from whole genome vs arrays. And the extra data (and noise) you get from whole genome can make association studies significantly more difficult.

Also with whole genomes, the extra cost in sequencing, storage, and processing time (all of which are significantly higher) wouldn’t be worth it from a cost/benefit point of view. Remember, most of the data is redundant or common between people, so full sequencing is really only beneficial in specific scenarios.