Because ad click through rates are so low and the signal is so incredibly sparse. It's all measured against a test set which the researchers rarely look at directly (because the data set is so large). So if something lifts ad vlicks by 1% and that's from 100 in 10k to 101 in 10k, they don't really care if 9k of the ads get appreciably dumber.
Sure, but if you buy a rice cooker, maybe the ads to show you are for: rice, cookbooks (featuring lots of rice?), pressure cookers, cast iron pans, sous vide recirculators, and so on. I mean, you have a rice cooker, you won't want another, but chances are you'd buy other things you don't already have.
Its funny you should use that example. I recently bought a rice cooker, and then bought another a few weeks later.
I've done similar things a lot, in this case it was because the first cooker had too large a minimum cooking amount. I'm currently looking at buying my mother google WiFi - which I bought a few weeks ago. And a few months back I rebought a tool I had just purchased because I lost it.
I currently work for a (responsible) own site adtech (personalization) company. Its surprising what behaviours our clients find profitable for their business.
There's a whole lot about advertising I don't know, but the question that sticks in my mind is, did you need that ad to know which rice cooker to buy the second time? Did it change your decision? Did you see an ad for a different rice cooker and act on that?
It seems like an effective ad would be this: Say you bought a rice cooker on Amazon for $75. Now Walmart shows you an ad for the same rice cooker for $60. That would certainly stick in my head- not for the rice cooker but for the price difference that Walmart can provide.
These are amazon internal adds, they don’t care about the specific rice cooker they show as much as the rice cooker category. Thus simply getting you to think about them while you’re on the website is enough they don’t need to predict your specific purchase.
Yes. Without the ads I wouldn't have bought the second cooker - it was a pain point for my partner not for me, so she complained once then a week or two later I saw the ad and purchased another.
With the Google wifi I had forgotten, and it caused me tk reconsider and write it on my whiteboard.
You said you work for adtech, I'd just like to comment that I personally like targetted ads. Areas of my life have improved from targetted ads, I'm shown things I didn't even know I want until I've seen them, etc. I look at it as a positive.
Its an interesting one. I've made some pretty active choices about jobs for ethical reasons but ended up in a slight geographical bind in my last search. I found my current company through a friend (who works in another department) and was very skeptical. After not finding anything bad I decided to go with it for a few months, but had a resignation letter ready in my car. I'm now over a year in and despite the normal problems businesses have there is not a moral one here - and I've had pretty great insight into the decision making.
It's not as simple as that. There are multiple ways to figure out which ad to show to a user, but I think the most efficient way (development and compute) is to put a user in segments (e.g. lives in area X, searched for Y, age Z, ...) and then do the bidding magic. There are tight tolerances for bidding--you want it to happen within a fraction of a second so the ad quickly loads for the user. Feeding purchase data is likely something that requires a bit of work, and I imagine the ad team at Amazon doesn't see enough value to prioritize it (especially since they make money either way--the advertisers pay for the ad with a higher CPC).
One thing to keep in mind is that the purchase data is per-user, so it's different from the segment approach I was mentioning above. It would require a whole new schema* that is queried whenever bids are processed for an ad event.
* Sorry for using the word schema here. It was used a lot on a team I was on in a similar context, and it bothered me at first but I got used to it. What I mean by schema is a "blob" with purchase data that is periodically processed by the bidding subsystem to add bid exclusions per user.
What is the ratio between people who viewed a rice cooker (the group I'm assuming they target with ads) vs. those who actually bought a rice cooker?
Is it possible the ratio is so large to where while it obviously could be optimized to exclude people who have since purchased said item - the optimization is only a rounding error, and may not turn out to be worth it?
Well, I buy things from Amazon very frequently, and I have a high conversion rate between searched-for-category and bought something to satisfy the search, so Amazon ads are for things I've already taken care of >90% of the time. I'm guessing they're leaving a lot of money on the table from me because of it. But maybe I'm unusual in my shopping patterns.
But couldn't they improve the signal to noise ratio by showing contextually relevant ads e.g. showing rice cooker cookbooks after purchasing a rice cooker? Wouldn't this lift ad click through rate? This is truly baffling expected behavior, who would ever buy a second rice cooker after having searched Amazon for rice cooker and compared a few top models already? I'd love to see some studies or data proving otherwise...
> they don't really care if 9k of the ads get appreciably dumber.
But maybe they should. The recipients of those 9,000 ads are being trained to ignore the recommendations because they’re so useless, and may continue to ignore its suggestions even after it improves. Likewise, a few clueless recommendations can “spoil” good ones shown at the same time.
None of this shows up in the test set, of course, but people tend to turn off their brains when ‘evaluating’ ML stuff.
> The recipients of those 9,000 ads are being trained to ignore the recommendations ... Likewise, a few clueless recommendations can “spoil” good ones shown at the same time.
How can you be sure that "random" untargeted ads don't have just as bad of an effect? Ultimately Amazon can see the numbers and we can't, and they made their decision based on that. It's pointless to argue over theories without data.
Amazon seems to think that I’m a real vacuum aficionado/collector as well.
It’s quite possible that those are still the best items to recommend to us, to mop up any “unhappy with the one they got, so save them with a return/rebuy” sales.
I wouldn't be surprised if the actual reason was they are capturing extra sales from people who bought something, liked it a lot, and then bought another to give to their kids or for their second home. My mom buys me random chotchkies all the time after she bought them and liked them, and if my parents were well off I could easily see her buying me a vacuum or a new mattress after she got herself one.
For 99% of people that probably doesn't happen, but the people with the money to do that probably spend a TON of money, so they are worth targeting with ads.
I suspect adtech industry would shrink down half its size if it could accurately attribute sales to ads. It sometimes smells as if it was standard practice to convince the client (intentionally, or unintentionally by also convincing yourself) that correlation is causation and therefore their success is attributed to ad spending.
I read somewhere (can't remember exactly where) that this happens because people are most likely to buy the same item they just bought due to not being happy with their original purchase.
Reminds me of booking.com that keeps on sending me hotels in small towns I just visited once (for a weeding, driving thru) etc. I don’t why they get smarter than that
But they are hard to reconcile with why Amazon keeps showing me a rice cooker ad, after I buy a rice cooker from them.