They have changed their methodology over the years.
They originally did it solely on DNA samples and used key profiles who had clear genetic profiles. These are what they call 'reference panels' then compared people to these points and how much DNA they shared.
This recent update is a little more complex. They've used reference panels AND they've used the data from peoples profile in a more complex way - they've used networks of people based on how they are genetically linked as well as using some data from users trees to connect these networks to specific places.
They have a white paper explaining it all:
https://www.ancestrycdn.com/support/us/2025/10/2025ancestralregionswhitepaper.pdf
If you read their methodology, it's a bit like picking a tree out from a particular wood and labelling it as the correct species without actually looking at it, on the basis of probability. You know which species of trees have been identified in this particular wood and would expect to find in that wood. The difficulty is getting the exact correct species rather than a very closely related species on the basis of probability. Your tree will be surrounded by lots of other trees of a particular species but your tree could be a closely related species and thus difficult to pick out as an individual.
(See figure 2.1 from the white paper for an illustration of my tree analogy).
My trees are probably pretty useful to Ancestry in that I've used pro-tools and then actively linked other users backed on DNA relationships and paperwork to my trees. (The link to tree feature rather just using predicted common ancestor feature). Each of my DNA trees has about 200-300 linked individuals with full paperwork trails now.
My profiles are probably pretty accurate precisely because of the extensive linking in I've done. It clarifies whether an ancestry prediction of expected relationship to a more accurate relationship. For example it's the difference between predicting it's a potential half 1st Cousin 1x removed, 1st cousin 2x removed, half great grandniece, 2nd cousin or half great grandaunt and KNOWING that it's a 2nd cousin and how.
If enough people within the same network do the same and reinforce this same connection, then ancestry starts to give greater weight to the probability of that information being accurate and how all the individuals network together. Even if my profiles aren't on the reference panels, my matches might be, so my work is useful for them. (Ancestry should also be able to pick up on errors to a certain extent within this). You then have both the DNA and the paperwork showing the same thing and again crucially how rather than estimates.
I suspect that if you happen to be closely related to a profile picked as one for their reference panel you are also more likely to have more accurate results - and you'll see some disparity in accuracy of results between those close to reference panel profiles and those furthest away from those reference points because you have to do more estimation.
It is true that the more people's data you have the better the results you get generally. You have more reference points and more network connections that can be linked up. Again the problem is where you have people with intercrossing trees creating noise. Sometimes it demonstrates a double connection with that area (so you know that intermarriage in that community and a narrow gene pool is a feature of that community) other times it just confuses the issue because a match is related to you in multiple ways from different lines making it harder for Ancestry to distinguish between the two. My Dad appears to be picking up matches that he's related to in Utah BUT also people related to them but not my Dad - I think it's something to do with this problem.
You can hit slightly incorrect data or odd artifacts of DNA from how they appear in sequence - you get one small sector representing A next to one which represents B from two separate ancestors but in sequence A next to B looks like C which is a DNA signature that is often found in an unrelated population or makes you potentially look related to someone you aren't.
This is particularly true within communities which come from smaller gene pools (not necessarily marriages of biological cousins but equally a small number of families marrying into a small number of families over a course of several generations creating multiple networks of how people are connected which are more difficult to unpick).
Ancestry themselves have said using customer trees alone is problematic hence why they've combined it with looking at DNA but as I say this also isn't without issue either.
There's still the problem of NEP events several generations ago - that are within closed communities from a close relative which would be hard to spot but would affect the data you get in terms of a profile for an ethnicity. If your reference points are slightly off, it has an impact on all your predictions. It comes back to whether they've got the right samples in their reference panel or not.
Last update they definitely got things wrong connected to the Shetlands, Isle of Wight and Italy I believe. They are much better this update. The last update did however improve issues with identifying German / Scandinavian DNA which were off from the 2023 update. What you see is sometimes they take a step backwards on one ethnicity for an update, but that does seem to usually get corrected at the next one.
The DNA profiles I managed are pretty spot on for location with the exception of the one which lacks Isle of Man. Instead it's showing as Northern Irish/Central Scotland which I don't think is wildly off. It's still in the right ballpark of closely linked ethnicities, perhaps with a deeper shared history. I expect to see this improve in future.
You also need to keep in mind that previous updates were really heavily caveted. A couple of the Germanic and Scandinavian ones basically said 'also found it England' meaning that you weren't part Swedish at all - you had a DNA signature that was most often found in Sweden but was also very common in modern English populations. It's a English signature. Lots of people really don't understand this. They get upset that they've 'lost' the Swedish (viking) they really never had and it undermines their trust in what Ancestry says.
There also a certain amount of snobbery and preference for 'interesting' results. Getting a result saying you are likely from the East Midlands really doesn't sound as cool as Denmark - yet it might still represent migration from the Vikings as the East Midlands is identifiably different from areas that didn't have viking migration.
My mum is now down to 2% Danish which I suspect in future updates will eventually disappear as they identify it as a distinct East Midlands signature. I think it was about 7% at one point. The same goes for my Dad where 2% Norway and 1% Sweden with turn out to be bits of his Highland and Yorkshire ancestries.
It's not made up shit. There is logic to it all, but I definitely think the range of accuracy varies a lot from one person to another for understandable and legitimate reasons.