I agree that the figures can vary for many reasons and we shouldn't expect them to be exactly the same (some things we don't end up controlling for).
At the same time, if we take the experiment of the soup for example:
They measured 9,227 sales of it so the 21.1% increase is quite robust and I'd expect the error margin to be much lower than 5% either way - so in some ways the precision is warranted.
I also feel that if I were to round a 21.4% to 20% I'd be miscommunicating the findings of the research :)
First experiment (science conference talks from YouTube):
"We selected two conference talks (in physics and engineering) from YouTube and altered their acoustic features using iMovie software. The good audio quality version of each talk was created with an audio filter called “small room,” which reduces the echo and increases the clarity of the speaker; the poor audio quality version was created with an audio filter called “Large Room,” which does the opposite, increasing the echo and decreasing the clarity of the speaker."
The second experiment (NPR interviews) was again using iMovie and they describe it like this:
"The good audio quality version of each talk was created with no audio filters so that participants heard the interview as it was originally recorded. The poor audio quality version was created with audio filters that made it sound as if the researcher had called in on a bad phone line."
True, I love the way Duolingo does it (I guess we could somewhat fit this under 'gamification' when it comes to digital products).
But it's such a missed opportunity in so many other domains, like encouraging people to eat healthier (although the UK Government does a good job with the "Five a day" slogan)