Do it with actual numbers and not percentages, because you're getting tripped up by the fact that your total population changes.

Take 1000 families, with two children. Assuming a perfect distribution, every possible combination of boy/girl for both children results is represented 250 times. Note that the actual order of the children is largely irrelevant, but if you ignore order you still have 250 families with two boys, 250 with two girls and 500 with one of each.

Now you take the next step. You reveal that at least one child must be a boy. This means you eliminate 250 families from your population, namely those with 2 girls. You're left with a population of 750, 500 of which have a girl as well and 250 who have another boy.

Now, pick one family at random from those 750. The odds of getting one with two boys is 250/750 or 1/3.

Yes, 1/4 of (all families with 2 children) has 2 boys, but 1/3 of (all families with 2 children AND at least one boy) has two boys.