3 ms·

The idea that ad targeting sucks because ad networks have no incentive to improve it is extremely dubious. Google and Facebook have an incentive to use up adve

The idea that ad targeting sucks because ad networks have no incentive to improve it is extremely dubious. Google and Facebook have an incentive to use up advertisers' budgets. But they also have an incentive to keep advertisers happy by providing good returns so that they will keep spending / increase spending in the long run. Advertisers are not always perfectly rational, but I don't think they're as dumb as the article makes them out to be. The simpler explanation for poor ad targeting is the bidding component. If it was the case that you were shown whichever ad had the highest predicted CTR, I'm pretty sure you'd be seeing some damn relevant ads. But instead you're seeing whichever ad has the highest (bid per click * ctr). Or worse, advertisers can also bid per impression, in which case there might be no relevancy component at all. My guess is the car ads on YouTube mentioned in the article are that last case. Both of those components (YouTube and car companies) are notable for skewing towards "brand" advertising rather than "performance" advertising. They're not trying to get you to buy a car right then. They're trying to get you to feel good about their company in the long run so that five years from now you're more likely to buy their car.

tilolebo 7y ago
+1 with that. Having worked for an ad network: it doesn't matter how good your recommendation algorithms are if all you have in your ad campaign pool is shitty diet and car ads.
- a123b456c 7y ago
  There is an alternate explanation that doesn't rely on advertiser stupidity. Most advertisers are statistically unable to estimate the profitability of their campaigns, because ads generally exert weak effects in very noisy environments. There is some high quality evidence supporting that, e.g. https://academic.oup.com/qje/article/130/4/1941/1914592 https://academic.oup.com/qje/article/130/4/1941/1914592 Nota bene, weak effects does not mean ads are necessarily profitable. Profitability depends on many factors including margin, retention, ad cost, etc.
bogomipz 7y ago
Might you or someone else mind explaining how this "bidding component" works? I'm not familiar with it or with the specific bidding models at work - "bid per click * ctr" etc?
- joefkelley 7y ago
  Advertisers go to an ad network and configure their campaign with: "I want to show X to users who are Y or are on a site about Y, and I'm willing to pay Z for it". X is just the content of the ad and is usually entirely advertiser-defined. Some ad networks will allow you to do some fancy stuff to help determine the best phrasing or arrangement or whatever of your ad. Z can be in dollars per impression, per click, or per conversion. An impression is one person being shown an ad one time. A click is a person clicking an ad. And a conversion is advertiser-defined but is usually someone buying the product being advertised. Bidding per click is the most common. Then when a user loads a page with an ad on it, the ad network finds all campaigns that are eligible - meeting the criteria defined in Y. It then runs an "auction". It has to calculate which eligible advertiser is willing to pay the most for that potential impression. But not all advertisers are bidding per impression, so some prediction comes into play. If an advertiser bid per click, then the expected amount they will pay is the probability the user will click multiplied by the price if they do click. Their probability of clicking is often called CTR which stands for "click-through-rate" so that's where the bid * CTR comes from. Usually an ML model takes what is known about the user, the ad, and the page, and predicts CTR. Similarly for conversion-based bids, it's bid * CTR * CVR, where CVR is "conversion rate". Whichever ad comes out with the highest result of this calculation is shown.
  - bogomipz 7y ago
    Thanks for the clear and thorough explanation. The parent comment makes good sense now. One follow up question - is the "bid per impression" sort of the campaign of last resort then? In other words its the cheapest type of campaign for an advertiser to run as well as the least profitable for the ad network? Cheers
    - 7y ago
summerlight 7y ago
Yeah, this is why the whole ad industry is moving to automated bidding which gives a bidder more rooms to improve. Clicks maybe a relevant proxy for the performance but still the most important metric is ROI which cannot be accurately predicted at the moment of bidding. Sadly, there's a huge conflict of interest between user privacy and relevant ads targeting here; we currently cannot accurately measure ads performance even with the current privacy-invasive practices and this will get even worse without them. This is why I think user-tracking needs to be standardized in a more explicit way; more accurate and reliable tracking can significantly improve ads relevance while it can give some level of user control and transparency by default. Of course, I'm 100% sure that this won't come anytime soon.

tilolebo 7y ago

+1 with that. Having worked for an ad network: it doesn't matter how good your recommendation algorithms are if all you have in your ad campaign pool is shitty diet and car ads.

a123b456c 7y ago
There is an alternate explanation that doesn't rely on advertiser stupidity. Most advertisers are statistically unable to estimate the profitability of their campaigns, because ads generally exert weak effects in very noisy environments. There is some high quality evidence supporting that, e.g. https://academic.oup.com/qje/article/130/4/1941/1914592 https://academic.oup.com/qje/article/130/4/1941/1914592 Nota bene, weak effects does not mean ads are necessarily profitable. Profitability depends on many factors including margin, retention, ad cost, etc.

bogomipz 7y ago

Might you or someone else mind explaining how this "bidding component" works? I'm not familiar with it or with the specific bidding models at work - "bid per click * ctr" etc?

joefkelley 7y ago
Advertisers go to an ad network and configure their campaign with: "I want to show X to users who are Y or are on a site about Y, and I'm willing to pay Z for it". X is just the content of the ad and is usually entirely advertiser-defined. Some ad networks will allow you to do some fancy stuff to help determine the best phrasing or arrangement or whatever of your ad. Z can be in dollars per impression, per click, or per conversion. An impression is one person being shown an ad one time. A click is a person clicking an ad. And a conversion is advertiser-defined but is usually someone buying the product being advertised. Bidding per click is the most common. Then when a user loads a page with an ad on it, the ad network finds all campaigns that are eligible - meeting the criteria defined in Y. It then runs an "auction". It has to calculate which eligible advertiser is willing to pay the most for that potential impression. But not all advertisers are bidding per impression, so some prediction comes into play. If an advertiser bid per click, then the expected amount they will pay is the probability the user will click multiplied by the price if they do click. Their probability of clicking is often called CTR which stands for "click-through-rate" so that's where the bid * CTR comes from. Usually an ML model takes what is known about the user, the ad, and the page, and predicts CTR. Similarly for conversion-based bids, it's bid * CTR * CVR, where CVR is "conversion rate". Whichever ad comes out with the highest result of this calculation is shown.
- bogomipz 7y ago
  Thanks for the clear and thorough explanation. The parent comment makes good sense now. One follow up question - is the "bid per impression" sort of the campaign of last resort then? In other words its the cheapest type of campaign for an advertiser to run as well as the least profitable for the ad network? Cheers
  - 7y ago

summerlight 7y ago

Yeah, this is why the whole ad industry is moving to automated bidding which gives a bidder more rooms to improve. Clicks maybe a relevant proxy for the performance but still the most important metric is ROI which cannot be accurately predicted at the moment of bidding. Sadly, there's a huge conflict of interest between user privacy and relevant ads targeting here; we currently cannot accurately measure ads performance even with the current privacy-invasive practices and this will get even worse without them. This is why I think user-tracking needs to be standardized in a more explicit way; more accurate and reliable tracking can significantly improve ads relevance while it can give some level of user control and transparency by default. Of course, I'm 100% sure that this won't come anytime soon.