Here’s Google’s search liaison clarifying things 2 hours ago on Twitter. Localization shouldn’t be confused with personalization. Disclaimer: I work at Google, but not in search
https://twitter.com/searchliaison/status/1070027261376491520...
https://twitter.com/searchliaison/status/1070027261376491520...
I suppose that Google wants to optimize for people typing "football" in their search engine, at the expense of those trying to type "football results $countryname", or something more precise. But this seems ultra-annoying for anything more substantial than random trivia searches.
As for language, well, years ago I set my default search engine in Chrome to explicit "https://www.google.com/search?hl=en&q=%s", because language optimization made 80+% of my searches return wrong answers. I admit I might be in a minority here, insofar as English-speaking people with work to do are a minority.
This is a constant frustration - I fight with this feature every time I search for info about some place I'm not, for instance to plan travel. Searching with a country/state/zip alongside the query helps, but it's a kludge that's much worse than actual location-specificity, especially if one is going to an ambiguous place like Portland or Washington. I'd very much prefer it if Google offered the same sort of mechanics as weather sites: defaulting to my location, then offering a little box saying "search as if I'm located in:". Instead, the only option is "use precise location", which is no more helpful than basing location on my IP and history.
In this sense, it's the same standard-uses-only paternalism that plagues Android. Maps on desktop has a great feature of "leave at" or "arrive by", which was removed from mobile in favor of a 'helpful' integration that can notify you about when to leave. It's terrible for anyone who wants to check a route for odd travel hours, or even do something as simple as decide when to set an alarm based on likely morning traffic. And so, of course, people have been requesting the more flexible option for years while consistently being told that the existing version is "more helpful".
Increasingly, this stuff doesn't just bother me on a privacy or information-neutrality level. It's actively damaging, to the point where DuckDuckGo is actually more usable for certain types of search.
Yup. It's a great example of what I complain about when I say that software is being dumbed down and loses its usefulness. As another recent thread[0] pointed out, the handling of street labels is ridiculously broken. It makes Google Maps useless as a map - i.e. as a tool I could use to look around the place and find my bearings, or plan a route in advance, or scout the traffic across many roads without putting in a destination.
I too am bothered by this more than by privacy issues, to be honest. The latter are maybe more important, but the former are causing me frustration daily.
-
I use this feature all the time and it's still there on the mobile versions, as far as I know. Ask for directions to a place, you'll see a list of routes, above the routes is a dark blue bar saying "depart at HH:MM". Touch that and it will let you select departure time, arrival time, or the last train.
Maybe this feature has moved around but it’s definitely there on mobile. If you touch an individual route it will take you to a different screen which doesn’t have this button, so go back to the list of routes and it will be there.
At any rate, I just tried to enter a destination from the main screen, hit directions, and provide a source. I also tried hitting 'Go' and then entering the destination, and tried each with "my location" and a specific address. In every case, I get several routes. The shortest is already highlighted, they all list times if I leave now, and there's a top bar asking if I want to drive, walk, etc. Touching 'back' returns me to the destination, with no departure source, so there's no list of routes which doesn't have one preselected.
From there, all I can do is pick a route and start, which begins navigation. None of available menus before or after starting let me set a "depart at" time.
If you're willing, could you tell me where our flows differed? I'm always curious about this sort of fragmentation, or if this is available I'd love to know where.
(Poking around, I did just discover "remember monthly driving stats", which is new since I last opened that menu and on by default. So thank you for letting me turn that off - and another way in which I'm irritated by Maps.)
In particular, DuckDuckGo attempted to check for the effects of fresh browser windows, changes from localization, and A/B testing or incomplete rollouts.
- The study results found that the 'distance' between two people's incognito results was 2.8 times larger than the distance between one person's normal and incognito results. If this is correct, the tweet's claim that you can use Incognito to see the impact of personalization for yourself is seriously misleading.
- The study attempted to control for localization by tokenizing all local results, so that two result sets which each had a local story in the same spot would be treated as identical.
- The study attempted to control for rollout and testing effects by running all searches at the same time and assuming that those forces would lead to most users seeing similar results, with a few differing. Instead, they found substantial variance across all users.
I can certainly come up with explanations for how each of these changes could be non-personalized. Incognito variance could be a consequence of region and device effects which are user-independent. Tests might be scaled up near 50/50, and rollouts may not hit all datacenters at once, invalidating the "changes for a few users" assumption. And most significantly, DuckDuckGo's localization control looks completely inadequate to me. It would completely fail the given "football" example, and might fail for the given search terms: 'vaccine' produced highly-localized results they adjusted for, but 'gun control' could produce pseudo-local results like a preference for national stories with the user's state as a keyword, and that wouldn't have been identified.
A better version of this study would presumably try to alter its variables separately, for instance by using multiple devices in one location and one device in multiple locations via VPN. As is, the controls seem seriously lacking.
But I'm just making up reasons, and frankly they don't seem likely to explain the sheer size of the differences. When I run those searches, I don't get "football" and "Paris" level customization, I get a bunch of national-scope results that have no obvious reason to vary between incognito users. I wish the tweets here had touched on any of that. As is, they explain general search mechanics while frustratingly bypassing the most significant claims.
Or it is 20 bits of entropy when you go to a website only a thousand other people in the world visit.
That's literally what is being demonstrated here.
DuckDuckGo's control definitely wasn't adequate, but in return Google's examples ('football' and 'Paris') have unambiguous localization value that's not present for two different Americans searching "immigration".
Quite a lot of the variance was link reordering and result changes that didn't have any clear regional aspects, whether at the domain or story level. If localization is the active factor, it's still pretty interesting to know that where I live determines whether to show a Wikipedia link and how to order HuffPo against the Tribune.
I have no clue how DDG works, but I am familiar with the Google search infrastructure as of several years ago and if you told me that the same query should always return the same identical results, I would tell you that you just don't know how Google works, even if you ignore personal results. That made monitoring and catching regressions not fun. It's better than running searches days or weeks apart, but running queries at the same time as they did does not guarantee anything, sorry.
Keep also in mind that the average query goes through thousands of machines, according to public statements. Behavior at the long tail can differ for obvious reasons. From a Jeff Dean talk, https://static.googleusercontent.com/media/research.google.c...
| Some issues:
| –Variance: query touches 1000s of machines, not dozens
| • e.g. randomized cron jobs caused us trouble for a while
| -Availability: 1 or few replicas of each doc’s index data
| • Availability of index data when machine failed (esp for important docs): replicate important docs
(see also the slides about Universal search, as well as patterns such as the elastic systems and the different tiers)