Every post on HN with this question receives the same response: “it’s because Amazon knows you’re actually more likely to buy another one if you’ve just bought one, even if it’s the kind of thing you only need one of (e.g vacuum)”

I find that answer unconvincing and there is no published data to back it up, but it is repeated constantly here.

What they're actually tracking tends to be the fact you've searched for $productcategory (whether you bought or not)

In theory since they also have purchase data their 'purchasing intent' categories could exclude people who have subsequently purchased from $productcategory for products which are substitutes (but definitely not products which are complements/consumables/collectibles etc, because someone who has just bought guitar/fishing accessories isn't just likely to buy more guitar/fishing accessories again in future but a lot more likely to than the average person), and yes they could even make guesses about whether somebody is likely to want more than one office chair based on whether they're purchasing as a business or an individual. But in practice it's not easy when they've got a product inventory that's so enormous, complex and dubiously-labelled even the basic product search doesn't work that well, and most of the time the vendors are paying them to show the ads anyway. And still wouldn't be perfect (ironically I seriously considered buying two different vacuum cleaners for different use cases this week!)

It's surprising how bad basic targeting is, like when Facebook allowed ad targeting by sexual orientation but even dating websites spent vast sums without bothering, so its unsurprising subtle distinctions about which buyer types only need one of which office accessory are often missed.

Here’s a different take: when you set up an ad campaign for retargeting, you likely select users who “viewed or added to cart in the last 30 days”; you should also add an exclude for “users who purchased within the last 30 days” but it is very easy to forget the exclusion.

Also sometimes there may be communication lags between retailers and ad networks. Some retailers may only upload purchase data daily, sometimes there are gaps in uploads etc.

I think most advertisers would prefer to not advertise the same item that people just bought, but it takes perfect execution both technically and from the marketers creating campaigns to pull it off cleanly. When you’re managing dozens of campaigns across half a dozen ad platforms, it’s easy to miss some of the details that matter. And while the wasted ad impressions may annoy the consumers on HN, it is likely a rounding error for advertisers and not a huge cost driver.

The most common case of this (as expressed in HN comments over the years) is Amazon recommending products based on Amazon purchase history. In that case it's all first-party data.

Most consumers, in their individual experience, probably avoid foolish purchases like; buying a second widget immediately after buying a first widget.

But it's probably true that there are a lot of individuals out there who have done this (for whatever reason), and it's the statistical likelihood of people buying a second widget that drives this.

An individual who doesn't do this, can't perceive this "fact" of statistics, which is human behavior in-aggregate form. Individuals only perceive individual behavior.

Also, the same-product-again recommendation doesn't have to be a great one; it only has to beat the alternative it displaces. So I could find it plausible that taking a swing on you buying another of a product you already bought beats out whatever the 5th-best recommendation happens to be.

Given how widely this policy is criticized, Amazon must be very aware of it and they must continue to do it intentionally. I would trust, given the amount of money on the line, that they see evidence that tells them it's a reasonable choice.

Caveat: I don't work on advertising systems, but I do work in ML, and I've studied recommendation systems. One way to think about these large systems is that they contain a condensed representation of you that is comprised of all the things you buy/consume on their platform (the raw data) and some derived descriptors (possibly demographic information/information from other platforms/"engineered features" like what "style" of clothing you like, etc. This representation of you (a vector) can be compared to other members of the platform (inner product), and they can give you recommendations based on your similarity.

The problem is that in this case, they've encoded into your condensed representation something about office chairs. A better system would be knowledgeable about the kind of purchase and infer if it's the kind of thing you would buy once, or multiple times.

But then you have the problem of different individuals having the propensity to buy different products at varying intervals. If I like shoes, I might buy many pairs. You might not be into shoes, and so you're not going to buy more than a couple pairs, if that. In this case, how do we train a system to know the difference between the two of us, with our different tastes? The more specific we get with the vector descriptor, the harder it is to compare us to other people (because in high dimension, random vectors are basically orthogonal).

And going through all products and trying to determine if they are the kind of thing that you would buy more or less of, given that you just bought it, is expensive, and likely difficult. The system ultimately falls on a positive correlation between history and future purchases.

Another way to think about this is with Bayesian reasoning, where your prior likelihood to buy office chairs affects the posterior. The system's prior is high (you've bought chairs in the past), which scales the posterior (that you'll buy a chair now).

It might be simply driven by statistics, or as a human-directed strategy.

While it seems counter-intuitive to the person who is targeted; (examples abound) - if you think about it from the perspective of someone selling widgets, the (set of people who are in the market for widgets) will necessarily include people who just bought one. Maybe they are not satisfied with their purchase? Maybe they want to buy another? Maybe they bought one, and it arrived damaged?

Personally, I would think that this kind of "trying to think FOR me" is rather offensive to a lot of people, and probably turns-off a fair amount of people who are subject to it. On the other hand, it is probably true that people exposed to an ad for a widget, immediately after buying one, are statistically more likely to buy another. So as long as that is true (if it is) - then that's going to be a criteria for ad targeting.

Amazon presumably doesn't share purchase history with advertisers; you get lumped into a cohort with certain interests/keywords.

An experiment to try; if you only go shopping/searching for vacuum cleaners but don't actually buy one you are likely to see the same phenomenon.

Why do advertisers pay for this? Because conversion statistics show that showing ads to people shopping for vacuum cleaners do actually result in vacuum cleaner sales. A/B tests back up the causality of the results, so vacuum sellers pay for the ads.

Why doesn't Amazon remove you from the cohort of vacuum cleaner shoppers after you make a purchase? 1) it would reveal that you made a purchase. 2) Amazon would lose out on additional ad revenue.

I have no inside information on Amazon but it matches what I've seen on other large ad networks.

I can't speak to amazon specifically, but my understanding of this phenomena in general, especially with ads that are bid for, is that the systems which track and gauge your likelihood to purchase aren't connected to the systems that know you did make a purchase, at least at the point in the system where it's making the decision of what ad to show. It just sees that you've done things that a person who is likely to purchase that thing has done, which, of course you do, because you DID purchase that thing. It's like the ad is chosen based on the strongest additive signal available, and negative signal isn't considered.

The likelihood is that the chair you got, you'll probably find out it is just bad and after a while you'll return it and look for something else. In the last few months almost 50% of my purchases turned out to be some junk pretending to be a quality product. I still have a few boxes of things that I need to post. There is even no point reading reviews, because most of them are fake. I think I am going to stick to trusted brands and probably drop Amazon altogether, but I like the next or same day delivery a lot. When using other stores you never know if the product arrives tomorrow, next week or never.

I've been puzzled by the same question (as a reaction to the same ads) and it occurred to me that it may be simple math:

If the likelihood of buying by someone who already bought the item is 3% and the likelihood of buying by someone else is 1% (simply because there are many more people in the second category), which group would you focus on with your ads?

Personally, if I like an item, I'll recommend (and gift) it to my family and friends. I can think of at least 3 items that I bought for myself, liked and bought for someone else that are normally more of a long-serving items (like a tea kettle or a vacuum).

That means the recommendation system isn't a neural net trained on the buying habits of users. It's a simple logistics system that recommends items similar to the ones you've bought.