The other day I saw Paul Haahr, a Google Search engineer vet, retweet something shared by Daphne Keller, former Google Search lawyer, who shared Facebook VP’s Nick Clegg Medium piece named You and the Algorithm: It Takes Two to Tango. With me so far? Anyway, the tweet from Daphne Keller “we will reward this behavior until you adapt your business and get too good at it, and then we will punish it” caught my attention. Paul Haahr explained “The idea here about adversarial/evolutionary behavior is one I’ve thought about a lot over the years. One interesting aspect of it being evolutionary is that intent doesn’t really matter – it happens because certain behaviors are rewarded.”
Here is the tweet, I 100% recommend you click on it, and scroll up and read through the whole summary. When you are done with that, keep reading below.
It’s kind of like platforms saying “we will reward this behavior until you adapt your business and get too good at it, and then we will punish it.” That’s the unavoidable cycle with spam or SEO. For news orgs, it has been awful. 18/
— Daphne Keller (@daphnehk) April 1, 2021
So when I that Paul Haahr tweeted the above tweet, I asked him for his thoughts on this cycle of a search engine coming up with a way to reward behavior and then having later to punish it because SEOs figured it out… Paul took some time to respond but his tweets are super insightful. Here they are:
The idea here about adversarial/evolutionary behavior is one I’ve thought about a lot over the years. One interesting aspect of it being evolutionary is that intent doesn’t really matter – it happens because certain behaviors are rewarded.
— Paul Haahr (@haahr) April 5, 2021
From the (early) Google perspective, links were just about relevance, authoritativeness, prominence, etc. But as soon as it’s known that a search engine uses links as an indicator (a ”signal”), there is an incentive to create links that only exist for search engines.
— Paul Haahr (@haahr) April 5, 2021
This behavior is going to happen, I’d expect, with any signal that becomes understood by an optimization community. But it’s going to lead to worse divergence when there’s greater distance between what the signal is used as an indicator of and how it’s used in optimization. /fin
— Paul Haahr (@haahr) April 5, 2021
I did ask Paul to expand on this last tweet and this is what he said:
I don’t want to cut you off, please do finish & come back to this line “it’s going to lead to worse divergence when there’s greater distance between what the signal is used as an indicator of and how it’s used in optimization.” what do you mean by this? I don’t fully understand
— Barry Schwartz (@rustybrick) April 5, 2021
Here is his response:
And by “greater distance between what the signal is used as an indicator of and how it’s used in optimization,” I was referring to the way a signal gets gamed (e.g., buying links) versus how it was originally seen by a platform (e.g., a signal of good reputation).
— Paul Haahr (@haahr) April 6, 2021
On the other hand, keyword stuffing can make word counts less of a proxy for relevance than naive information retrieval would tell you they should be, so that would increase this ”distance.”
Hope this clarifies.
— Paul Haahr (@haahr) April 6, 2021
I do not want to add any commentary around what was written here because I think I might misstate something or potentially take some of the words out of context. But I found these responses from Paul to be super interesting and thought you’d all like to see this.
Forum discussion at Twitter.