Hacker News Reader

Comment by caconym_ | Hacker News Reader

What's the source for it being easy? It seems easy, from the perspective of a human looking at results, but I'm not sure how much that's worth given the scale and complexity of the problem.

View on HN

Just block the domain. At first, you can block manually, but we know Google doesn't like doing things that way. Fortunately, they have a lot of heuristics to find sites like that; usually the content is just copied from another source. And since they scrape the web all the time, they should know which content has appeared first where.

But the issue isn't that they can't; the issue is that they don't want that. Why the sites with copied content exists? To earn money through ads. What earns Google money? Ads!

Any simple heuristic has false positives, meaning they'll end up taking down legitimate sites that had repeated content for a good reason. Say, for example two sites quoting text from the us constitution. The second one to be crawled would be considered to be spam copying the first one and removed from web results. Then you'll get comments on hacker news complaining that Google is censoring it for political reasons.

And any simple heuristic is quickly reverse engineered by SEOs, who will find a way to mask it as legitimate.

tl;dr it's a hard problem.

They could use the heuristics to build a list of domains to block and then have someone review it. After doing it for a long time, they could build a neural model on top of that, and automate it.

As I have said, the reason they don't do it is not because they don't have the skills and know-how.

Captchas are way harder to solve than it is to detect these sorts of poor results. Google should have absolutely no problem building a classifier that could scale to solve the problem. But as you said, it's not worth it to them, but for the reasons of losing revenue rather than scale and complexity.

Kagi filters all of that out. If an upstart can do it, Google with its unlimited resources can do it too

Kagi seems to be in closed beta, so I wasn't able to test this claim even on an anecdotal level.

Reply on news.ycombinator.com