In the second part of the Search Off The Record podcast released this morning Gary Illyes from Google explained how Google does ranking. In short, he said Google first cuts down the pages that can rank for a specific query to a short list of let's say about 1,000 pages. Then the magic happens where Google will apply ranking signals or as he calls it "magic signals" to that cut down list. It is on that level where the true Google ranking magic happens, Gary said.
Initially, to create the cut down list, Google looks at topicality and relevancy based on the query and content of the page. Then Gary explained they use a method to create a reverse order list, you then cut what you cannot list and you come up with the the 1,000 list (for example - maybe an arbitrary number Gary used).
This list is created but they are not done with ranking there, with the cut down list of 1,000. He said in this cut down list, that is where "the magic" is, the "magic signals that we still apply on the result set to make them better for the user's query." One example is RankBrain, as a magic signals, but he also mentioned the HTTPS boost as another.
So in short, Google cuts down the max number of pages it can return for a query and then the ranking algorithm and magic algorithms work within that cut down set.
Then Gary gets into how it ranks those cut down list of pages, he said "now ranking is number-based. Basically for each result, we will assign a number and we calculate that number using the signals that we collected during indexing plus the other signals. And then essentially, what you see in the results is a reverse order based on those numbers that we assigned." He added "the magic signals or magic algorithms that we use like RankBrain, what they do is multiply those numbers that we assign to each result by a number. Like for example, if they want to promote a result because it was determined that there would be a better result for lemon coconut cookie, then let's say that it would multiply the result score by 2. Basically doubling its score, which means that it would jump up in the result set."
Gary added "if we wanted to remove a result from the set for whatever reason, we could multiply its score with 0. Because then, that will turn the score to 0 and then like a score with 0, why would you present that?"
Gary then added that it is highly unlikely a page would have the same score. He added that is where the HTTPS boost, which is a magic signal, but it is just a tie breaker and it is "a tiny, tiny boost." The tie breakers would not rearrange the results unless there is a tie but only if there is a tie, otherwise it would not.
This is all pretty interesting to listen to, so I embedded the audio at the start time below at 28:52:
Forum discussion at Twitter.
