I like Unreal Tournament ('99) running the UTBENCH.DEM demo. That way it's a test of realistic gameplay, not the intro cinematic.
Make sure the console is closed before the demo starts playing. It affects the framerate if it's still open.
Once I started testing that, I decided Unreal was probably redundant.
In Ultima 9 I picked a spot, made a save file, and watched the built-in fps counter (it can be enabled in the game's config).
This game likes to move the crosshair when you load the file so I had to try to define exactly where the crosshair should be pointed.
I'd look away from the screen, then look back and record the first number I saw. Then look away and repeat a bunch more times and take an average.
I think that test is still subject to some inconsistencies based on time of day or nearby animals or whatever, but it seemed decent enough.
It's not perfect and nobody cares about Ultima 9 anyway, but I did so I bothered with it. 😀
Gabriel Knight 3 - I did something similar as with Ultima 9, just picked a spot somewhere near the start of the game and used FRAPS to get a framerate.
Driver - You can get a repeatable scene in this for a benchmark, but it requires leaving the game running for several minutes each time waiting for it to happen. The resulting scene is lower than typical framerate because you're waiting until the game has maxed out the number of cars that are around you.
I don't remember all the details of how I did this game but it's in a spreadsheet somewhere.
I've taken a similar FRAPS-based approach with some other games that aren't as relevant to Win98. I prefer to benchmark games I might want to play instead of the ones that have convenient benchmarks built-in.
