>Ideally you could use Bing as your default engine, then fall back to Google whenever there's a search that doesn't yield good results.

This is what I do with DDG. Unless I have a pretty specific search, or want Google's really nice live sports scores widgets, DDG is usually pretty good.

DDG doesn't track what you click on, so unfortunately you are not contributing much to improving its and Bing's search quality by using it instead of Bing.

Thanks for posting, it's nice to hear from someone on the inside!

Downvoters: why on earth are you downvoting Analemma's post? It's constructive, on point and overall the kind of post which makes HN comments valuable.

What I'm about to say will probably get downvoted as well, but I've noticed HN getting more polarized in terms of votes in the last maybe 2 years.

Because of the way that downvoting works on HN, the few times I use it, my thought is "nobody should see this comment". Because enough downvotes delegitemize and hide a comment, it's effectively telling other people that it's not worth being read.

I don't know if that's actually going through people's minds when they downvote, but that's what I think which is exactly why I rarely use it. But too many people now are simply using it on posts they don't like, even if a post deserves to be there. For some reason, it's an unpopular opinion to believe that an idea or opinion you think is wrong deserves to be seen and rebutted rather than silenced.

Downvotes for dislike is supported by HN/pg. Their choice of fading out comments, when combined with downvotes-for-opposition is really terrible for a high quality discussion site. Just goes to show the community around a site makes all the difference.

I still don’t understand the hive mind downvote patterns.

> I used to work at Bing. If you really want Bing to improve, the best thing you can do is just use it: clicks on search results, plus backs and dwell times, are vital training data.

A million flies are attracted to shit. I don't believe this sort of training data will ever become useful if it's in the same pool with the rest of the world. See also: voting with your wallet against the tyranny of the majority of uninformed consumers who buy whatever is most marketed. Those pennies don't matter.

In fact, I believe this sort of training and optimization for the mainstream plays a role in allowing bad results to proliferate.

This is before we even consider the fact that clicking on many results can indicate that they're bad (I click another result because the previous wasn't good), or because they're good (I'm browsing choices). Dwelling long can be bad (crappy & slow site, it takes me long to find the information I want or turn away) or it can be good (I found good stuff and I'm spending a while on it). Whatever conclusion your training system draws might be completely wrong. And probably prone to being gamed.

Clicks are noisy and have their problems (especially clickbait), but as far as ranking goes they still blow out of water everything else you can use. Analemma_ is right on the money. We need more Lemon Pledge^W^W clicks!

Speaking as someone who worked for another big web search engine in the past (not G, not B)

And counting good/bad clicks for a given (query, url) pair is just the tip of the iceberg. There's a lot of other interesting stuff you can do with them _if_ you have the data. Deep learning/NLP with clicks as training signal is probably the most exciting area to me, to name one. Unfortunately almost all of the data currently gets nabbed by google, other search engines are just getting by with the scraps. And it's very hard to bootstrap a competitor from scratch - for example Cliqz had to resort to some shady deals with Mozilla just to get any data to start with.

The above heuristic arguments seem inherently weaker than the direct experience of someone who used to be on the team that improves the results. They are well-funded and should be able to back out the effects you mention.

How are they ever going to know whether they improved results for me or not? You might as well train AI to play a game without ever checking their score. Oh, it's spending 20% more time in each room now and firing fewer bullets than before. Surely it is a better AI now.

Sorry, appeal to authority is no argument. Direct experience is valuable if there's an argument or some real scenario we can dissect, otherwise it's nothing more than a baseless claim. Without concrete examples, it's not even an anecdote. And there are plenty of anecdotes about search results for in-depth content becoming harder and harder to find.

The person I was responding to posted that they're using this training data to order the top 10 results or so. That's already an indication that it's not very helpful for me. I don't get frustrated if the top few results are in suboptimal order, I get frustrated when I get pages and pages of garbage and irrelevant results and can't seem to get anything useful out of it.

A lot of this does, in fact, get solved through machine learning. If one search query is conducted a hundred times by different people with otherwise similar profiles, the machine learning can constantly modify the results until the hundredth person sees the best results first. For example, if the first 25 people all click on the 2nd, 4th, 5th, and 8th result and then stop, it can modify the order of the results for the next 25 people, and randomize the order, so that some see 8-4-2-5 first, and others see 2-5-8-4, etc. The it re-assesses the results and gets smarter.

The gaming part is where there's a lot of importance, since it forces the machine to learn which users are acting nefariously.

Shuffling the top 10 results is not fixing the big problem in search for me. I don't care too much whether the company I search for is in first or fourth place. But I'm really irritated when I get pages upon pages of useless results and nothing remotely relevant.

See the replies elsewhere. The metrics go beyond simple, naive "click counting."

Specifically they also measure metrics like "regret" - if you went back to the search results and chose something else this indicates it wasn't a useful search result for the query and this counts against it.

In theory this is quite a nice defense against clickbaity, content free results.

> Specifically they also measure metrics like "regret"

I explicitly addressed this:

> clicking on many results can indicate that they're bad (I click another result because the previous wasn't good), or because they're good (I'm browsing choices).

You don't know the reason why I clicked back. Maybe I got what I needed at a glance, and I'm browsing to the next thing to see if there's more of what I need.

You don't know that the Average Joe (and a million other Average Joes) wasted 15 minutes on clickbaity content free crap despite not getting anything useful out of it. You know, those same SEO spam sites are trying to find ways to make the user stay as long as possible; they are in the game. And I'm sure there are plenty of gullible users who will stay. Who knows, maybe SEO spam is working and that's precisely why we see SEO spam sites in results?

Again, I think training like this is pretending the problem is much simpler than it is. And I think that there is a real danger that training like this just optimizes for 1) average users 2) sites that managed to game the system with their SEO. Neither is optimizing for quality results.

That theory has so many holes and pitfalls.

I just think you're assuming these search engines are using way fewer factors than they actually use. It's probably exceptionally easy to tie your search behavior to someone else who's much more like you than "the average", unless you're using a search engine like DDG, Qwant, etc.

Maybe the best bet is for a search engine to try to include a simple, non-intrusive method of getting user input to determine the usefulness of a search result. Like, if it determines you came back as a "regret", maybe a ThumbsUp/ThumbsDown icon next to the last link you clicked to get you to say whether that link was useful in any way, or maybe a "block" option as well to say that you never want sites like that again.

> I just think you're assuming these search engines are using way fewer factors than they actually use. It's probably exceptionally easy to tie your search behavior to someone else who's much more like you than "the average", unless you're using a search engine like DDG, Qwant, etc.

I should, then, expect dramatically different results when I use someone else's PC or a public computer. I have not witnessed dramatic differences.

It's just you pretending search engines are simpler then they are. Noone's just ranking most clicked sites first - that's too naive and encourages clickbait. Clicks are an ingredient in the overall ranking system, which is generally also trained on relevancy labels by human annotators/raters/assessors. These human labels are far more reliable and can penalize all the bad clickbaity stuff which clicks can't, but unfortunately also very expensive to obtain and thus low volume.

We may not know exactly why you clicked back, or why Joe wasted 15 minutes on a site, but from all clicks in aggregate and from the human-labelled data, we can tell how that correlates with the clicked site's relevancy and quality and utilize your clicks accordingly to improve the ranking.

Finding such correlations is still by far and large what all search engines do, rather than trying to truly understand the query. It's only now starting to change with recent advances in deep learning and NLP.

> clicks on search results, plus backs and dwell times, are vital training data

Interesting. Can you explain more on what is tracked after a user clicks a result and how this data is used?

(Disclaimer: as I said, I don't work at Bing anymore. Things may have changed in the intervening period and I could be mistaken or plain wrong about things)

Obviously ranking search results is immensely complicated with lots of accumulated insider wisdom, but the short answer is that "static" page features (PageRank, how well the query matches text on the page, etc.) are good for pruning down zillions of pages to generate a candidate list of 10 or so results to show the user, but not great at ordering those results, and serving the results in the best order is really important for user satisfaction. For generating the final top-10 ordering, "dynamic" features like how many people clicked that link, how long did they spend on the page, etc. are the most useful.

Regarding what is tracked, obviously Bing can only track what you're doing on Bing itself, so clicks on blue links and backs are tracked, but they can't see what you do on the destination page. This data is used to train models/NNs/etc. but the raw click data isn't kept for very long: it's too large and it has to be removed for GDPR compliance.

> For generating the final top-10 ordering, "dynamic" features like how many people clicked that link, how long did they spend on the page, etc. are the most useful.

Do Google do this as well?

I've never worked there and can't say for sure, but I have to assume that the broad strokes of their search engine are the same as Bing's. Bear in mind my response above was a very very high-level description of a really complex system with tens of thousands of person-years of accumulated knowledge and experience inside it.

Yandex does that. At the bottom of search results page there are links to other search engines. So if you aren't satisfied with results, you're one click away from Google. I believe they track these clicks and know which results are bad.