It seems I've had the most changes in fare price or availability when identifying flights on United specifically, more so than any other airline, in my experience. That being said, it's still pretty infrequent given the volume of airfare we've booked through it for our people over the last 2 years or so.
Another thing I would add is that in connecting flight situations, it will occasionally present an itinerary that's technically possible to make, but not always entirely practical or ideal. (e.g. the window between flights is really narrow, not leaving much margin of error for flight delays, getting to another gate/terminal, etc., if not near the one you arrive at.)
All that being said, it really does work pretty great and we prefer using it to any of the other travel sites.
I would love to know why there are occasional fare discrepancies though... as far as I know google treats the click-through to the carriers' sites like an ad click, so I don't think Google is "marking it up," rather, the carriers are in some cases pricing based on the referer link or something. It can also be a demand pricing issue too, as the flight fills up, I suppose. My understanding is that Google is basically just the informational conduit between the user and the airline so it's weird that there are occasional price discrepancies.
Short answer re price discrepancies: it's complicated. It's absolutely in our interest and the airlines' interest to get pricing 100% right, and we do work really hard to eval and improve on this.
One situation that can happen is that an airline has, say, 3 seats remaining at one fare and 1 seat remaining at a cheaper fare; when you selected the flights or immediately before, that 1 seat was still available, but by the time you click through, it's not, so you get the pricing for the fare with 3 seats left.
One of the things that's amazed me in my flights education after joining the team is just how super-dynamic pricing is.
No, to my knowledge it's not based on your referrer or cookies or anything like that ;). I'm just talking -- person-independent -- pricing from hour to hour.
On a related note, I used to think that airline pricing was pretty capricious and random; now I realize that the airlines hire some of the smartest folks to develop pricing strategies, and nothing is left to chance ;).
Also, do you book on Tuesdays?
And no, don't bother trying to time your bookings. It's kinda like the stock market; trying to beat the market, er, optimize your flights purchase time is likely to lead to sorrow and regret. From the research our Google Flights team did recently, we found that if you wait a week, on average you're gonna pay more for the same flight.
And no, we didn't see any universal discount on Tuesdays, either ;)
I haven't experimented extensively with things, but usually setting my exit server to a random country on the other side of the planet (e.g. Malaysia) has yielded consistently good results.
And yes, I do tend to book on Tuesdays.
Booked all flights incorrectly :)
* Google's data seems to be not as fresh as Kayak or Hipmunk's is. There have been multiple times when I've tried to book a flight via GF only to be told that there were no seats left when I arrive on the airline's page. I suspect part of the greatly improved speed comes at the cost of aggressive caching that runs a risk of things being out of date.
* Google's data is all over the map on pricing. Very frequently the airline's price will be much cheaper than GF. And more than once (especially on United or American Airlines) I've noticed that if I book directly through an airline over a VPN, I save several hundred bucks versus if I click through from Google Flights or Kayak. Conversely, sometimes GF is less expensive and when I land on the airline's page it's more expensive.
I think the best thing to use it for is route discoverability: which kind of planes and which airlines fly to which locations, on which days at which times? It's also great for doing complex open-jaw routes or hidden-city fares -- and since those routes don't change very often, the data is more likely to be accurate.