Introduction ------------ From September 2009 to August 2012, I used as my everyday web search engine an app that merges results from several search engines. This page shows some stats extracted from my tracked click data, in order to help answer the question that kind of sparked the whole project, "which engine gives the best results?" For some background see here: https://saintamh.org/code/search.shtml.en Select a period to analyze -------------------------- Displaying only clicks recorded since Mar 01 2020. Select a different time period: * last year * last 3 months * last month * last week * everything Clickthrough rates ------------------ With engines hidden and fair blurb selection: Each line shows, for each engine, the ratio of the number of times they were invoked to the number of times one of their results was clicked. If an engine always returned off-topic results, I would never click on their results, and so their clickthrough rate would be 0. If an engine always returned results that seemed so relevant that every result I ever clicked on had been returned by them (along with possibly other engines), then that engine's clickthrough rate would be 100%. The number in square brackets is the average rank, within the engine's result page, of the results that I clicked on. So smaller values mean I tend to click on results that are higher in that engine's results, higher values mean I tend to dig deeper in their results. As the headline suggests, these numbers exclude the times that I opted to see which engine had contributed which results, as my opinion about engines might skew my choices (I rarely select to see the engine names, except when debugging this app itself). Some engines were invoked fewer times because after a while I stopped using them. User votes ---------- Since 2011-05-20, every search result has two links next to it that allow the user (that is, me) to flag results that stand out as signficantly better or worse than the rest of the lot. Here are the tallies. The numbers next to each engine indicate the number of upvotes ("+") and downvotes ("-") the engine has received. The percentages in brackets indicate how many votes that engine receives on average every time it is invoked: An alternative way of scoring these votes is to give more weight to results that are ranked higher by the engines. This seems intuitive: if for instance I downvote a link that was returned by two engines, and the downvoted link occupied the top position in engine A's results, but only the 30th position in engine B's results, then it sounds reasonable that more blame should go to engine A than to engine B: it's not so bad to return irrelevant links in 30th position as it is in the top position. In the table that follows, each value is the sum, for all votes on links returned by that engine, of the inverse of the rank that the engine gave to the link. So for instance in the previous example, the downvote would add a full negative point to the tally for engine A, but only 1/30th of a point for engine B. Engine correlation matrix ------------------------- The value in the cell at the intersection of row R and column C answers the question "among all search results returned by search engine R that were ever clicked, what percentage had also been returned by engine C?". This means that if a cell has value 100%, the results returned by the engine in that row are a strict subset of the results returned by the engine in that column (assuming that this is independent of whether I've clicked or not on the results). Conclusion ---------- The results have been pretty consistent over time: Google emerges as the clear winner no matter how you measure it. Their results are more relevant and complete than any other engine's. The engines complement each other quite well, however -- Bing often has relevant results that Google doesn't have, for instance. And 1 out of 4 results that I click didn't appear in Google at all. Another interesting finding is that the quality of Yandex and Google results is slowly but surely improving all the time, while other engines seem more stable. Last recorded click: Fri Feb 1 11:30:39 2013 UTC