Hacker News Reader

Comment by binarymax | Hacker News Reader

Love this article. Worked as a search relevance consultant for years and these are exactly the problems we work on. Most of the time fixing search is a thankless job. People expect it to work and when it does they take it for granted, but when it doesn’t it damages your reputation and brand.

If you want to get started fixing website search relevance I recommend the books “Relevant Search” by Turnbull and Berryman, and “AI Powered Search” by Grainger, Turnbull, and yours truly. Both published by Manning.

View on HN

Yep I work on Search as well -- both of these books are excellent.

Website search is... hard. A lot of the faceting still needs to be done by hand. I think there's probably some opportunity for LLMs to make some sort of autotagging/categorization easier, but there will likely still need to be a human in the loop to verify.

Great recommendation on Relevant Search. I worked on an ecom site trying to improve search relevance for years, and it’s an incredibly difficult and challenging problem. Especially because it’s difficult to measure success, and execs would constantly come up with “hey why does this query not return exactly the results I expected” which led to a lot of resources spent optimizing a handful of queries based on executive input instead of based on query frequency and query result conversion metrics.

Thanks for the recommendations! I also work on a decently large ecom site and search experience is a difficult problem - especially in aftermarket automotive.

I'm not trying to fix search so what's the tldr on why it's so hard to do with such constrained databases? The Nintendo one seems insanely egregious. Is this not just a symptom of corporate laziness rather than search problem complexity?

The searchable content may be constrained, but the queries aren’t ;)

Search is hard because you need to anticipate and model the language of all potential searchers and the content.

There’s also lots of ambiguity because you’ll only get one or two keywords from the searcher without any other context, and you need to take into account trends and content quality and metadata.

Also, in lots of cases search is bad because the product team either doesn’t know or doesn’t care.

But his examples were about hardcoded filters/facets.

The examples were about a keyword or two (“lentils”, “laundry detergent”, etc) and filters.

He means the last examples about Nintendo.com

The reality is that yes there is a long tail of searches people might search for, but for popular searches like “laundry detergent” you can check them manually. You get a lot of coverage just by fixing bugs from the top N queries.

> People expect it to work and when it does they take it for granted, but when it doesn’t it damages your reputation and brand.

That's the definition of a table stakes feature. It's not as if you can not implement it and people won't notice!

Reply on news.ycombinator.com