Another alternative is that perhaps people are clicking on ads for things they already own. I mean, let's say there's a 1/5000 chance that a random person is going to buy a SDXC card. You go ahead and buy one. After the purchase, statistically does your likelihood of buying another SDXC go up or down? For some people, they've fulfilled their SDXC quota, so now their likelihood is zero. For other people, they may be in the market for SD cards generally, so they will be more likely to make a second purchase than a random person. Or, after having bought one, they'll want to buy another one as a gift. Or they're looking to return the original and buy a cheaper one at a sale price. Just the fact that they have equipment that takes SDXC cards means they're more likely to buy another one than a random person.
Same thing with a pool filter even. If you live in an apartment, or if you're a minor, or you don't own a pool, your chances of buying a pool filter are extremely low. So the fact that you just bought a pool filter puts you in the narrow group of people who would potentially have need for another pool filter. Even though you just bought one, you're still more likely to be a customer than a random person. Hence, generally speaking, sending you ads for the same thing you just bought may be a rational choice for the advertiser.
An intelligent advertising engine would know the typical interval between two purchases of the same product, and start advertising similar products just before that interval has elapsed since the last purchase.
Examples:
pool filter: 12 months
reptile UV lamp: 6 months
gift-wrapping paper: 12 months
oil filter, recycling box, 5 qt oil: 4 months
8-ct paper towel rolls: 2 months
television: 4 years
automobile: 10 years
magazine subscriptions: 1 year
tacos: every Tuesday
Indeed, but this engine doesn't exist, and in 15+ years of weaponised adtech, noone's been sufficiently economically incentivised to make it. Telling, no?
You're underestimating how hard it would be to build such an engine. Google and Facebook have sunk $billions into their knowledge graphs so far, and not made much of a dent.
Maybe. The current heuristic is "he has bought X so show him an ad for X", but would it be so difficult to flip that to "he has bought X so show him the next thing on the list of interests we know he has" or even just "anything but X". I strongly suspect noone is doing this because there is no point, they get paid the same regardless.
That's not the current heuristic; it's not nearly that simple.
Think about how you know if someone is interested in something. What does it mean to be interested? How does that translate into a sale?
A person just bought a vacuum cleaner. Does that mean they are interested in vacuum cleaners? Are vacuum cleaners things that people get interested in? Do people who are interested in vacuum cleaners buy more than people who are not interested in vacuum cleaners?
To us these answers are so obvious that it's humorous. To a computer, not so much.
There are many millions of interests, opinions, and things to buy--so count the combinations. We understand it all intuitively because we are adult humans with decades of experience in this society.
This is why search and ad companies hire so many AI researchers.
EDIT: to understand the financial incentives, see my other reply to you. In a PPC ecosystem it is better to match too greedily than to risk missing a customer ready to buy.
That's the intuition, but I wonder if there isn't a spike in near-term repurchase behavior that the algorithms are correctly exploiting?
Ex: I bought a vacuum, but I'm unhappy with it. I bought a pool or oil filter, but it was the wrong one. I bought an X, but it was defective. I bought an 8-ct of paper towels because I run a cleaning service or property management company and use a lot of paper towels. I bought a pair of jeans and they're the wrong size or I bought a pair of jeans and they're exactly the right size and now I want to order 3 more pair so I can have the ones I like.
Years ago, we built a conditional offer engine and the humans operating it kept complaining that it was "broken" because it was suggesting non-sensical combinations. In split-run, those non-sensical combinations were clear winners, even if the humans couldn't get comfortable with them.
In the cases you provide, though, it seems to me (and I am certainly no expert) that you wouldn't need to advertise the same item again. If I bought a vacuum that I'm unhappy with, showing me another ad for it (which is what Amazon does) is the wrong thing to do. Not only will I not buy the same machine again, but I will google and research the products- I will ignore ads unless they are offering me a discount on the one I decide on. Likewise for wrong-sized jeans or buying three more pair.
Acxiom has done this in a more invasive way for decades: figure out women's menstrual cycle and then mail them ads when they will be most likely to make a purchase according to prior research.
You've got to think beyond the computer. Most big companies pay per ad view, not per click. Per click is for the low-end advertisers. Per action is for bottom-feeders.
Think about billboards, TV, radio, and magazines. Can't click on any of them, yet hundreds of billions of dollars are spent on them every year because repeating the brand, logo, visual, or message works. There's a reason McDonald's and Coca-Cola advertise heavily. It works.
I strongly disagree about your CPM/PPC/CPA categorization.
I work on conversion focused ad campaigns, so YMMV. For conversions, obviously CPA and PPC are superior. I do agree that CPA can be bottom feeders, but PPC is probably the bread and butter of most people I know.
I would suggest that most big companies are paying their agencies, who like pay per view because they can spend more without direct accountability like Return on Ad Spend (ROAS). Their job is easier without having to prove their efforts work.
Grain of salt and all. I'm also a guy who has yet to see proof that a branding campaign has value. (It may lift sales, but is it the most effective use of those dollars? Doubtful.)
Why is CPA for bottom feeders. What do you mean by that? Seems like the perfect way to optimize the funnel if you know exactly what you're after (an "acquisition").
The problem isn't 100% negative matching (ads shown only to people who already own the thing), the problem is greedy pattern matching (ads shown to anyone who might possibly be interested). That ends up matching against people who want the thing, and people who just got the thing.
And that's not really that bad for advertisers because they only pay for the clicks. People who don't want it (perhaps because they just bought it) just won't click.
It is only a "problem" for the ad platform, in terms of the efficient use of inventory. In theory a poorly matched ad is displacing a better targeted ad that would perform better and make the platform more money.
But practically speaking it is only a competitive disadvantage if your targeting is worse than competing platforms. Even if the targeting leaves a lot be desired, if it is even a little better than everyone else, you'll still land contracts and make money.
