Yesterday I googled “Everything Trump Touches Lies” and it came up on Amazon. Good, no problem.
Then I scrolled down to the “Books you may like” section. The algorithm apparently has decided, based on “Everything Trump Touches Dies”, that I may like the following:
Women Who Run With the Wolves: Myths and Stories of the Wild Woman Archetype
Hold the Line : My Story from the Heart of the Freedom Convoy
The Quran
The Ballad of Songbirds and Snakes (apparently a prequel to the Hunger Games novels)
Safe and Sound: A Renter-friendly Guide to Home Repair
Do they just throw five titles together at random?
I don’t work for Amazon, and have no actual knowledge of how their particular algorithms work, so YMMV…
But… it seems to me that what they try and do is not necessarily match up the words of your query with specific books, but rather with other people who search for/buy similar products.
So it’s likely that for example if you’re looking for 5/32 silicone tubing, you may start seeing stuff related to RC airplanes, because that sort and size of tubing is used in model airplanes and that’s what the other people who searched for it / bought it were interested in.
I’m sometimes mystified about what Amazon throws up, but various articles online claim that 35% of Amazon’s sales are attributable to its recommendation algorithms. I haven’t linked them, because I haven’t been able to track down the source of this statistic or exactly what it means - what’s the “control” here, since there is no “Amazon without recommendations” to compare to?
But apparently its recommendation AI has a strong reputation, at least in relative performance against other algorithms.
They know that people who looked at that book have some reasonable likelihood to buy one of the others based on many millions of samples. It doesn’t have to make sense. A machine is deciding.
They are constantly tweaking and dialing things in. Most people see one thing and several sets of a few thousand people see something else. If one of those do better, that’s what the big group sees next.
Of course. But I was questioning exactly what the widely claimed “35%” really means. I suspect that it may be an exaggeration - the recommendation algorithm may be involved in 35% of purchases, but that does not mean that this is all additional revenue to Amazon. In some cases, the customer buying a recommended thing means that they don’t buy something else that they otherwise intended to buy.
So how are you inferring the specific narrow connection to your Trump search? Why isn’t “books you may like” based on all the data they have on your preferences?
That’s a good point (as was your point on the following post). It also could be something they were going to buy next anyway which a good algorithm should have predicted.
I’ve had an amazon account since nearly the very beginning and made thousands of purchases. I don’t think I’ve made more than five purchases based on recommendations – maybe even less than that. Sometimes if I have the time, I’ll check out one or two, but usually I don’t end up buying. Mostly it’s good for reminding me that an author has a new book in a series I’ve been reading (like Elly Griffiths who has three series I’ve read).
I tried this, and at first it took me to the hardcover, where I didn’t get a “Books You May Like” section.
What I did get included a “Products Related To This Item” section, a “4 Stars and Above” section, a “Related Products With Free Delivery” section, a “Customers who viewed this item also viewed” section, an “Inspired by your browsing history” section, and a “Best Sellers in this category” section.
So I switched to the Kindle version of the book, and then I did see a “Books You May Like” section. Based on what I saw there, I strongly believe that the suggestions there are based on your (i.e. my) past search/items viewed history, not on the specific book that that page is for.
I think you’re just considering the “you might also like” type stuff that spontaneously pops up. It’s not just that - the AI is obviously also involved in what comes back when you proactively search for something. Perhaps it’s finding you what you really wanted, when you didn’t know the best way to search for it, or when there was something better suited to your needs that you did not realize existed.
If I may risk a very slight hijack by extending the OP to other online algorithms, particularly those of Google…
I watch a lot of YouTube. The ads shown during videos should, IMO, show ads that are relevant to me, based on a lifetime of viewing stuff (and, since it’s Google, they also know my search patterns, websites visited, and even gmail). However, they don’t seem to keen on using that data for me, instead relying to a very large extent on the video at hand.
For instance, if I’m watching a video about machine repair, suddenly they’re giving me an ad for welding equipment or a welding certification course. Next video, about furniture repair and so…I get an ad for cooking spices. OK… And if I watch an ad that happens to be in German (could be about classical music or whatever)? I get a German language ad. Same deal for Spanish language ads, but they pop up a heck of a lot more, despite my rarely watching ads in Spanish.
Suffice to say, I think Google is squandering their partner ad dollars by splurging ads all over the place like this, using a pretty lame criterion (you view this video, therefore you get this ad). Google doesn’t pay for these ads, but places them. It’s the advertisers who pay everything, which makes Google rich.
I’ve been involved in a similar deal at work, where we yanked these ads because their placement was so deeply irrelevant. We “wasted” a ton of ad revenue on such ads before we stopped. This was fully a decade ago, or more. Anyone allowing Google to mis-target their ads is just wasting money. But again, it makes Google rich.
I don’t think so, because it doesn’t usually find things I’m having trouble searching for. It’s frustrating. And I meant all the forms of recommendation they provide. Or is a simple search result considered a recommendation?
I don’t know that there’s a bright line here. Any search result is going o contain both “simple” results that take your search terms at face value, and AI-influenced results that reflect what you may really want given the combination of your search term, your own history, and the totality of customer behavior.
Given Amazon’s success and the resources they devote to it, I’m skeptical that this is generally true. We all inevitably have selective memory about this, tending to remember the things that we have trouble finding.
Of course, the fact that their algorithms are going to maximize the total amount of revenue over whether you are subjectively “frustrated” may sometimes create a conflict of interest. If it’s a situation where someone who can’t find something immediately on Amazon goes elsewhere, that’s obviously not desirable. But if it’s hard to find anywhere, then maybe the frustrated person just keeps browsing and ends up making other purchases.
I’m always surprised that it’s often so difficult (for example) when something comes in a range of colors at various prices to search through to find the cheapest colors. But how many times have I walked away from a purchase after I’ve reached that point? So maybe there’s no revenue payoff for Amazon to devote resources to make that any easier.
I Googled this and clicked on the Amazon link. Oddly, I don’t have a “books you may like” section.
I have “customers who viewed this also viewed” which is a bunch of other anti-Trump books, “what do customers buy after viewing this item” which is mostly the same anti-Trump books and “inspired by your browsing history” which has not books at all and totally appropriate*.
*Willie Nelson CDs because of a CD I bought for a Mother’s Day present and cookware because I just upgraded a bunch of kitchen items.
The algorithms definitely like people with common & narrow tastes, rather than with unusual interests or wide-ranging interests.
If you want an anti- or pro-Trump book, or the latest best-seller, they’ve got lots of data and lots of suggestions that are mostly on-point. If you’re looking for something on the history of underwater basket-weaving or on tower crane design they have very little data on those queries. But they still want to supply your response page with 5 “you might also like” suggestions. Something will always bubble to the top of their suggestion logic; they just won’t be very strong suggestions.
It’s a bit like finding the 5 highest points in Saskatchewan closest to the OP. Everything is so flat there that a) they’re hard to determine unambiguously, and b) from the observer’s POV they don’t look much different than anywhere else within 100 miles, and so appear to be random selections.