This is how I've heard it explained by Grindcore:

Assuming a game-theory toy game where all Hero's value hands beat all Villain's bluffcatchers which beat all Hero's bluffs.

Assuming pot-sized bets.

We work backwards from the river. On the river we have 2:1 value to bluffs since we offer Villain 33% odds to call. Say 8:4 combos (because it makes earlier street maths easier).

On the turn, note that the 4 bluff combos we bet on the river can be thought of as value bets on the turn if the Villain is going to fold on the river. (And folding is the same as calling ie EV=0 since GTO frequency so we can assume folding on river and we can hence assume our river bluffs are turn value bets). So pot-sized bet on turn lays same odds as river pot-sized bet. That is Hero can bet 12:6 value to bluff combos on turn.

Ditto flop, so 18:9 value to bluff combos on flop.

Now only 8 flop value combos are river value when called by Villain's bluffcatchers on river so of the 27 combos bet-sizing allows Hero to bet on the flop only 8 are river value leaving 19 bluffs (bit less than 1:2). And on the turn, 8 river value and 10 bluff (about 1:1).

You can do the same calculations for, say, a half pot bet. Each street 3:1 value to bluffs and work backwards from river. Then work out what river value to bluff ratios are on each street.
Half-pot bet-size:
River: 27: 9 (36 total) value to bluff combos or 3:1
Turn: 36:12 (48) implies 27:21 river value to bluff or 9:7
Flop: 48:16 (64) implies 27:37 river value to bluff or ~3:4

Caveats:
This a toy game not real poker. Assumes Value > bluffcatchers > bluffs always. In real poker we have draws that change in value as hand proceeds. Flops vary greatly in terms of hand strength and how dynamically they may change by river. It is non-trivial to apply the results of this toy game to poker: you can't assume the same toy ratios simply approximate to real poker especially on earlier streets invariant across flops. Rather the toy game is an interesting starting point for static flops where the assumptions are closer to true but are of doubtful value on dynamic flops.

Iamindifferent does a good job of explaining how to do it and addressing some of the limitations. I've created a spreadsheet for you to easily plug in the numbers and it also allows for you to play with how bluffing frequency will change depending on how often your bluffs improve and value hands can't continue.

https://www.dropbox.com/s/30r2v16784n5qwk/BluffToValue_Ratio.xlsx?dl=0

Dropbox 404 for me, too.

For pot sized bets :
pot odds offered OTR : 1:2 -> 33% bluffs

River : 66.67 % value combos, 33.33 % bluff combos
Turn : 66.67 % X 66.67 %, or (66.67 %)^2 = 44.44 % value combos, 55.56 % bluff combos
Flop : 66.67 % X 66.67 % X 66.67 % = or (66.67 %)^3 = 29.63 % flop value combos, 70.37 % bluff combos

For 1/2 pot sized bets :
pot odds offered OTR: 1:3 -> 25 % bluffs

River : 75 % value combos, 25 % bluff combos
Turn : 75 % X 75 %, or (75 %)^2 = 56.25 % value combos, 43.75 % bluff combos
Flop : 75 % X 75 % X 75 % = or (75 %)^3 = 42.19 % flop value combos, 57.81 % bluff combos

here you go:
spreadsheet for determining value to bluff ratio

A couple other considerations beyond the initial model assuming that ranges are polarized. With my spreadsheet it will show you what the flop value to bluff ratio should be depending how strong your value and bluff hands are... specifically how often your value hands can continue to value-bet through the river & how often your bluffs will become the nuts and you can then value-bet them.

However, what it will continue to assume is that the callers bluff catchers will never improve to be ahead of the Hero's value range. In addition, it doesn't take into consideration where your bluffs improve to be stronger than the opponents bluff catchers, but not strong enough to turn into value bets. For example, let's assume that you have over-cards and a flush draw. Now if the flush draw comes in, you could value bet your bluff hands; however, if the over-cards hit, you'll be ahead of your opponents bluff catchers but not strong enough to value bet.

I do disagree with IamIindifferent in that due to the limitations of these models, they're not very accurate on dynamic boards. You do need to make some adjustments given the limitations outlined above, but the frequencies will not be far off and I've confirmed this with some indepth CREV trees.

Awesome Santaur, tyvm for your wonderful spreadsheet!

Thanks guys!