8086, 80286, 680[01]0 all had a 16-bit word size and could only access RAM at 16-bit granularity (I/O is a different story). Of course they had the ability to extract just a byte as well.

68000 generates a bus error if you try unaligned 16/32-bit access.

But you could have had a system with 64k 16-bit words that gives 128k bytes.

16 bit microcontrollers are a $25 billion market.

Those are usually supporting byte accesses, not limited to word accesses.

Doesn't the PDP-11 and competing microcomputers of that era count? Being the PDP-11, it's lavishly flexible (i.e. it can do byte-addressed as well), but ultimately 16-bit words were conventional.

I worked with a 16-bit minicomputer as late as 1995.. word addressable, not byte addressable.

I worked on embedded code in the early 2000s using 16-bit DSPs. Texas Instruments C54x if I remember right.

Not being byte-addressable was a real pain. There was a C compiler, but off-the-shelf C code hardly ever worked properly because everyone assumes CHAR_BIT is 8.

The one I worked with had a C compiler as well, though it was K&R and not ANSI. The instruction set supported operations on 8-bit values by op codes which selected the lower or higher 8 bits of a word, so the compiler itself worked transparently as you would expect (the architecture was big endian so that had to be taken into account sometimes, but that's true of byte-addressable systems too).