The Noisy Channel

 

Personalizing Advertising based on a User’s Click-Through Rate

January 4th, 2009 · 12 Comments · General

According to Wikipedia:

Click-through rate or CTR is a way of measuring the success of an online advertising campaign. A CTR is obtained by dividing the number of users who clicked on an ad on a web page by the number of times the ad was delivered (impressions). For example, if a banner ad was delivered 100 times (impressions delivered) and one person clicked on it (clicks recorded), then the resulting CTR would be 1 percent.

The click-through rate measure is the key enabler for the pay-per-click (PPC) advertising model, where advertisers only pay when a user actually clicks on an advertisement to visit the advertisers’ website. A testament to the success of the PPC model is that it accounts for the overwhelming majority of Google’s $20B+ annual revenue.

Most of the attention to CTR has been ad-centric. The quality of an ad–or, rather, of how well an ad is targeted–is largely measured based on its click-through rate, average over all of the users to whom it is presented.

Google, in particular, requires advertisers with a low CTR to place a higher bid per click. The relative ranking of ads reflects a product of the bid and the CTR. This product can be interpreted in one of two ways: either combination of the advertister’s and users’ interest, or as the expected revenue that the ad will generate for Google.

But there is a different way to look at CTR. Instead of looking at the aggregate behavior for an ad across all users, why not look at the aggregate behavior for an user across all ads?

For example, consider a user who never clicks on ads. Perhaps the user is using an ad blocker that the search engine cannot detect, or the user may simply be ignoring the ads. At the other extreme, there are users who click on ads at higher than average rates (though some may be bots committing click fraud).

Of course, all user behavior is averaged  in calculating the CTR for an ad. But a user-centric view suggests a a couple of advertising personalization strategies:

  1.  Don’t bother showing ads to a user who never clicks on them, since there is no value in doing so. If ad display alone is valuable in influencing users, then there should be a cost-per-impression component, though that would require a reliable way to determine that the user actually sees the ad.
     
  2. Calibrate the threshold for ad quality (i.e., the minimum CTR across all users) to a user’s propensity to click on ads. Doing so could reduce the annoyance of people with high thresholds while increasing the ad revenue from people with low ones.

It’s clear that enough people click on ads to keep search engines in business, and that the easy availability of ad blockers (AdBlock for display ads, CustomizeGoogle for Google’s PPC ads) has not made a dent in this revenue stream. Nonetheless, showing ads to users who don’t click on them degrades user experience without generating revenue for anyone–a lose/lose.

Personalizing advertising based on a user’s demonstrated inclination to click on ads feels like a no-brainer. Has anyone tried it?

12 responses so far ↓

  • 1 Jonathan Mendez // Jan 5, 2009 at 9:28 am

    It’s a great concept Daniel.

    I’m not sure though that we search enough to get to the confidence thresholds necessary on a personal level. Especially when you consider that certain search segmentation would also be required (academic queries vs. commercial, recovery vs. discovery, etc.) to make this practice most effective.

    Let’s say the average person searches 1,200 times per year. Could this work? It would probably be more effective with display where we know there are a large number of people who never click – though there is clearly some value in the impression as display can drive up search volume. What about the search impression? I maintain there is unlocked value there as well.

    Then again, maybe some people have just never seen an ad that interests them yet…?

  • 2 anon // Jan 5, 2009 at 11:26 am

    1. Advertisers are not supposed to correlate user information on a historical basis to identify individuals over time. Some do anyway, but of course the biggest networks periodically purge their data, and so there is no real way to tell if a user has the propensity to click on ads anyway.

    2. In any event, users who are most likely to run adblock+ and friends are also likely not to accept cookies, and thus cannot be tracked or identified in the first place.

    3. If it was generally known that people who don’t accept cookies wouldn’t see ads, hell, everyone would turn off cookies. No one really wants to see ads, they’re noisy and useless.

    4. It costs virtually nothing to display an ad if there is no CPM component, so what the heck, why not display to everyone. At least the ad may influence some kind of shopping behavior in the future even if there is no immediately click.

  • 3 Daniel Tunkelang // Jan 5, 2009 at 12:03 pm

    Anon, I’m not suggesting that the individual advertisers would see this data, but rather the search engines and ad networks. Perhaps that’s what you meant. In any case, compared to the sort of information that Google and others learn and store about users, the user’s CTR seems like a drop in the bucket.

    As for people who block or don’t accept cookies, they represent a small minority of web users. Even among sophisticated users, many accept cookies in order to access services, especially Google’s. In any case, it is easy for informed people (e.g., anyone who reads this post) to turn off ads, and yet most people don’t. So either they are uninformed or they actually don’t mind the ads. Clearly some people even click on them.

    Finally, it does cost something for a search engine to display an annoying ad. Annoyed users can become ex-users and turn to other search engines. Or they can stop paying as much attention to ads and in some cases even block them.

    Why do you think Google has worked so hard on ad targeting? Sure, they want to maximize the revenue per click, but they also care about overall user experience. I may not agree with everything Google does, but I give them credit for having this part right.

  • 4 Daniel Tunkelang // Jan 5, 2009 at 12:14 pm

    Jonathan, you may be right that personal data is too sparse and that users would have to be segmented. Though I’m not so sure. I’ve read that average CTR is about 1%. If that’s the case, and if we assume a normal distribution, then someone who doesn’t click on an ad when given 1,000 opportunities will show up as statistically significant. Of course, it’s easier to reason about larger populations, but I think it’s not impossible to imagine tuning this at a personal level. Especially if we include ads on ad-supported web sites in the calculations.

  • 5 anon // Jan 5, 2009 at 12:30 pm

    > Anon, I’m not suggesting that the individual
    > advertisers would see this data, but rather the
    > search engines and ad networks.

    Yah, I meant the networks. These guys purge all their data periodically — Yahoo! recently switched to purging after just 90 days, so user CTR propensity information will be lost frequently even when the users accept cookies.

    > I’ve read that average CTR is about 1%.

    As an aside, I wish I could find some good data on CTR. I’ve heard that it averages 1%, and also that 1% is “very good”, and other random remarks, so I suspect this figure is just a random number that people have pulled out of thin air. Obviously it various greatly per site and ad type and style.

    > Annoyed users can become ex-users and
    > turn to other search engines.

    True, this is why Google text ads are so well tolerated, because they are as minimally annoying as they can possibly be. Adblock+ doesn’t even block ads on Google itself, by the way, though I suppose it easily could, as they are so easy to identify.

  • 6 Daniel Tunkelang // Jan 5, 2009 at 12:39 pm

    I suspect that if a search engine offered you the prospect of better ads in exchange for tracking your click-through rate, people wouldn’t feel all that violated. I think that even the governments setting the framework for what constitutes personally identifying data could live with this.

    As for CTR numbers, I’d love to see good data too. But I’m not holding my breath. My point is that, if 1% is in the ballpark, then it’s possible to see statistically significant deviation from the mean in somewhere on the order of 1,000 click-through opportunities.

    Finally, Adblock+ may not block Google ads, but CustomizeGoogle does.

    http://www.customizegoogle.com/

  • 7 Panos Ipeirotis // Jan 5, 2009 at 8:49 pm

    Ha, very interesting!

    It also seems to have connections with the problem of identifying the correct “review” for a paper given the (noisy) reviews of many reviewers.

    Step 0: We do not know the “true CTR” of an ad, neither the propensity of users to click. So we start by assuming the same “estimated propensity” for all users, and same “estimated CTR” for all ads.

    Step 1: Given the estimated propensity of users, revise the estimates of CTR for the ads.

    Step 2: Given the revised CTR estimates, revise the estimated propensities.

    Step 3: Go to Step 1 until convergence.

    There are a million other things that can be done in this space (user propensities’s varying by ad category, time varying CTRs, optimal matching of CTRs and propensities) but I will keep the comment short 🙂

    Thanks for the inspiring post!

  • 8 Daniel Tunkelang // Jan 5, 2009 at 9:35 pm

    Panos, my pleasure! And you’re right that there a more general space of problems where data points are people-object pairs, but somehow all of the focus is object-centered rather than people-centered. The paper-reviewing problem is quite analogous, and it would be nice to see a reviewer’s input valued based on his or her historic track record.

    In any case, it certainly seems clear that we can do better by exploring the joint click-through distribution of people and ads than by always taking a marginal distribution over the ads and projecting out the people.

  • 9 Josh Young // Feb 13, 2009 at 7:09 pm

    Here’s a quick thought about the costs of showing ads to people who don’t want them. I certainly agree that showing ads to users who hate them has its costs. This may boil down closer to bickering about improbable edge cases than I’d like, but note that there’s a potentially problematic group of people: those who click on ads now but who wouldn’t if they thought that forbearing would make the ads go away altogether. If this group were sufficiently large, the advertiser would be better off forcing ads on everyone, even those who don’t and wouldn’t click on them under any circumstances, than it would be showing ads only to people who really want them. The intuition here is that we may be unfairly smuggling in our intuitions about how many people actually click on ads as would click on them in substantially different possible worlds. CTRs might actually drop, in other words. Also note that if this were actually a dominant strategy, it would be tautologically true that one search engine wouldn’t have to worry about losing a user to a competitor on these grounds because the competitor would follow the same facts and logic to the same strategy of showing ads to all users.

  • 10 Daniel Tunkelang // Feb 13, 2009 at 9:00 pm

    What ever happened to blog posts having a 24-hour lifespan? 🙂

    To your point, I doubt that there are many people who would normally click on ads but who nonetheless dislike them enough to avoid clicking on them long enough to make them go away. I think of my mom, who ignores my advice to never donate money to any charity that relentlessly sends her junk mail or calls her (and to tell them that), despite her telling me that such invasive tactics annoy her. I think she’s representative of a sizable demographic.

  • 11 Josh Young // Feb 13, 2009 at 9:14 pm

    There’s sort of a natural law that blog posts can’t live for more than twenty-four hours. I don’t see why we need to compound that kind of forgetfulness with an explicit law too! 🙂

    But wait, you don’t think an indulgent mother might be able to control herself and her charity for a just few months and beg off the intrusive phone calls if she thought that doing so would mean freedom from such guilt-reliant tactics forever thereafter (or at least a long time thereafter)?

  • 12 Daniel Tunkelang // Feb 13, 2009 at 10:52 pm

    That’s my point–I’ve explained to her that a bit of such discipline would do exactly that, but to no avail. Perhaps she’s exceptional. But I doubt it. It’s not a big secret that telemarketers target those people who have a history of responding to telemarketers. And yet most of those same people complain about the intrusive tactics of telemarketers. Nonetheless, they don’t take the necessary action to make themselves less desirable targets.

Clicky Web Analytics