Here is some insight to our auction dynamics. Of course, this may be subject to change!
Ads from the same advertiser can only show once per hour. Say if a user jumps on Twitter and sees an ad, that user won’t see another ad from the same advertiser until an hour has passed, and a new ad request is called for that user (say a load more tweets, or reopening the app).
The same campaign can only be shown to a user up to 5 times a day.
If a user dismissed the ad, that advertiser’s ads won’t be displayed to that users for 30 days after the dismissal.
You compete with yourself in a sense where only 1 ad can be served per request, but you don’t ever increase your own bids by having a lot of overlap between ads. Having a lot of overlap doesn’t really hurt you, it just might spread out your data/delivery, and make it slower to get statistically significant results. The system will always choose 1 winner per request, as long as the advertiser/campaign hasn’t hit the frequency cap.
Ad set structure - Multi-ads in ad set or one?
This also doesn’t matter much, as long as you are under the limit number of line items. The system will always choose that 1 ad (using prediction models, and if there is little data, it is probably more random) to promote per ad request.
The prediction models do take into account global creative performance, so for example, an ad with an image/video would usually get served over a plain text ad if given the same targeting, bid/budget, if they don’t have historical data.
How does the bidding system work?
It is a generalized 2nd auction. Say you open up Twitter, and a bunch of tweets load. In between those tweets, there are ad slots. Tweets between ads in timeline =8, when tweets between ads in search is 15. If you are the first place winner in the auction, you get the first place slot, if you get 2nd place, you get the 2nd slot from the top, etc.
Mobile App install ads also have a carousel slot - these are usually after the initial timeline slots are taken up. So say Ad winners 1-5 are first slots in the timeline (including the first ad in the carousel), while 6-10 are then the carousel scroll.
As for your point in #3, it could be something about that ad that makes the system think it would perform better than the other.