Full explanation
If a Le Zeus bonus buy looks like it is returning around 94% rather than the published ~96%, there are two likely explanations, and one of them is in your control.
The big one is the version. Just as the base game ships at 96.26%, 94.18%, 92.37% or 86.27%, the operator's chosen version sets the baseline for the whole game, buys included. So a buy that pays out around the 94% mark is a strong hint you are on a reduced-RTP version of the slot. Check the info screen, and if that is the case, the same buy will return closer to 96% on a site running the standard version.
The second is plain variance. RTP is a long-run average over a huge number of buys. Across your last ten or fifty buys, the realised return can sit well below or above the theoretical figure, and 94% over a short run means nothing on its own. You would need thousands of buys before the realised number settles near the published one.
So the practical move is the same as always: confirm the posted RTP version before you buy, and do not read a short losing run of buys as proof the game changed. It did not. Either you are on a lower version, or you are looking at small-sample noise.
Two checks before you blame the game
First, the version. If buys are returning around the 94% mark, open the info screen and confirm which RTP setting the casino runs. A reduced version sets the baseline for the whole game, buys included, and that alone explains the lower figure.
Second, sample size. RTP is a long-run average over thousands of buys. Across your last handful, the realised return can swing wildly below or above the theoretical number, and a short losing run of buys is statistically meaningless. Confirm the version, ignore small samples, and you will usually find the game is behaving exactly as designed.
Two things are happening at once here, and it helps to separate them. The buy carries its own RTP that is usually a touch below the base game, and short-run results scatter widely around that figure. A run of poor buys is almost always variance on top of a slightly lower setting, not evidence that the feature itself is broken.