Hacker News Reader

Comment by ryandrake | Hacker News Reader

Even if you search for the exact item, including the brand name and the size, you get pages and pages of choices, many of them just wrong.

I just did a search for "sprayway glass cleaner 19 oz 1 pack". This contains all the information you need to uniquely identify a single product. Brand, Product, Size, Quantity. Yet Amazon returns 3 pages of crap, including Wrong Brand, Wrong Product, Wrong Size, and Wrong Quantity. I can't make the query any more specific, Amazon, what the hell is your problem? This query should return one and only one result.

View on HN

Their search result page is just a ranking of whatever they predict will sell best based on your query (based on units, not dollars). It’s stupidly simlistic, actually: if we sell more for a given search term because we’re advertising hard on that search term, our organic ranking for that term will rise almost immediately. So in your case, what’s going on is that basically other people who search with the same term end up buying other products. Thats bad for you, but perhaps its actually good for them?

But think about the corollary where places like McMasterCarr have exact products in known quantities, types and breakdowns because they used to be a catalogue. I will go to McMaster 9/10 times because it's just easy to find exactly what I want and so they get my money.

Amazon seems hostile to this very idea, that I know what I want, will spend 15% more to have it tomorrow and will not by the random stuff it smashes in the search results because I know what I want.

Oh I get it. I actually think there’s a huge opportunity in online retail to gate off categories and brands that are already in, say, Walmart. Then do the Amazon algorithmic model for everything else.

I think you’re missing how crucial Amazon’s algorithmic model has been as a way to connect shoppers with 2 day delivery for an unimaginably long tail of products. We sell a technical product that would be at home on McMaster. On the 2nd largest retailer in the US. It’s incomprehensible that our products would ever be in Target or Walmart stores. Yet we have access to the same customer base at the same place where they can buy Tide. Amazon’s purely algorithmic approach is what allows companies like us to do product/market discovery for them (and their customers).

Did you know that Amazon has a section that drives to compete with McMastercar? It’s somewhere in Amazon business they have parts and stuff and it’s just a horrible shit show.

that's your guess and what they tell investors.

my guess is sheer incompetency hidden by sales volume.

at some point they had the idea of translating the search to categories and showing alternatives. it got both unmaintained and other products teams added their own self interest on top of the algo and now you have the worst search on the planet which nobody have the political power to fix because it touches everyone revenue, even if for worse

It's not stupidly simplistic. It's ads. Random no names bid for keywords to be placed on those pages, the higher the better. Good luck discerning actual ads from the real listings.

Payola.

Pretty sure that's Amazon selling your query results, not a failure in search algorithm.

For the sponsored returns to the search query, sure. But for the rest, isn't that the fuzzy search crapping the bed in the face of our expectations?

When I specify what I want I want the search to return what I searched for, not a lot of kinda similar junk. I've tried to explain the concept of fuzzy search to my girlfriend a few times and she still gets annoyed when a search for "24"x24" shadow box" returns a whole bevy of standard picture frames at an utterly random spread of sizes, for example.

Or is fuzzy search a convenient excuse for "we consider sellers who advertise with Amazon and subscribe to more of our seller services a valuable factor in search rankings"?

There is that

Fundamentally, Amazon seems to have a problem with not understanding priority. It's especially terrible at associating numbers with units. While I have no access to how Amazon actually works I think it's just counting matches and thus "sprayway glass cleaner 19 oz 1 pack" and "sprayway glass cleaner 1 oz 19 pack" would score equally and "bogus glass cleaner 19 oz 1 pack" would score higher than "sprayway glass cleaner 19 ounce". Remember, most quantity 1 items will not say so in the description, "1 pack" is asking for trouble.

I think the only real fix here would be for Amazon to separate out brand, product, size (or sizes in a few cases) and quantity. And the parser needs to attach numbers to the following word if that word is a unit word.

While I've never done it at any real scale I have done a routine that predicts what's related to what. It required giving it a fair understanding of the data (for example, Xabcde is probably related to X despite the distance between them) to get it to quit spewing out whoppers. It's hard even in my case where I could afford to run the test against all reasonable candidates. Amazon simply can't do that. (My routine runs in length * length * table size. You simply do not implement anything that even resembles O(n^3) at scale.)

Trying your search, I'm not seeing a pile of crap.

First hit differs from your search term only in 2 vs 1. I believe Amazon doesn't consider numbers as as much of a mismatch as other things.

The third hit is obviously what you want, but it doesn't have the commas and the unit is "1 pack (packaging may vary)". Apparently that makes it score slightly less. Going down the page I find it in 1, 2, 3, 4, and 6 packs, fresh scent vs unscented and there's also "blue". There's one clear typo of 15 oz, and you get a bunch of things where it's a multi-pack bundled with some other cleaning item.

Farther on down I start seeing different sizes and related products. Is there any harm to it displaying lesser matches farther on down??

Reply on news.ycombinator.com