How much do advertisers pay to reach you? Panagiotis Papadopoulos - PowerPoint PPT Presentation

*phrase originator: Andrew Lewis, a.k.a. blue_beetle If yo If you ar are no not pa paying ng fo for it it, , yo you ar are the the product ct * : : How much do advertisers pay to reach you? Panagiotis Papadopoulos FORTH-ICS, Greece & University of Crete, Greece panpap@ics.forth.gr Nicolas Kourtellis Telefonica Research, Spain Pablo Rodriguez Rodriguez Telefonica Alpha, Spain Nikos Laoutaris Data Transparency Lab, Spain

Data-driven economy • The user data of an IT company → contribute to its overall market valuation • Companies pursue more and more users personal data • By purchasing them • By providing free services (Google search, Facebook etc.) Internet Measurements Conference [IMC'17] - P. Papadopoulos 3

How is all this data converted to money??? ecosystem of digital advertising ($194.6 billion in 2016) Internet Measurements Conference [IMC'17] - P. Papadopoulos 4

Digital advertising progressively moving towards a programmatic model ads are matched to interests of individuals elaborated user tracking privacy implications Internet Measurements Conference [IMC'17] - P. Papadopoulos 5

How much do advertisers actually pay to reach you? Internet Measurements Conference [IMC'17] - P. Papadopoulos 6

The Motivation… Internet Measurements Conference [IMC'17] - P. Papadopoulos 7

The Background… Internet Measurements Conference [IMC'17] - P. Papadopoulos 8

Programmatic auctions of RTB Real-time Auction Bid Request (+user info) Available ad-slot 0.95 CPM (+impression) Website on the Ad Exchange (ADX) user’s browser Demand Side Platforms (DSPs) Internet Measurements Conference [IMC'17] - P. Papadopoulos 9

RTB price notification channel Real-time Auction Available ad-slot Impression delivery A (i) you won, 0.95 CPM (+impression) D (ii) here’s the charge price, (iii) the impression is Website on the rendered successfully Ad Exchange (ADX) user’s browser Winning Demand Side Platforms notification (DSPs) nURL example: cpp.imp.mpx.mopub.com/imp?ad_domain=amazon.es&ads_creative_id=ID&bidder_id=ID&..&bidder_name=..& charge_price=0.95&country=ESP&currency=EUR&latency=0.116&mopub_id=ID&pub name=.. Internet Measurements Conference [IMC'17] - P. Papadopoulos 10

The Challenge… Internet Measurements Conference [IMC'17] - P. Papadopoulos 11

Encrypted prices on the rise • Charge prices in nURLs tend to be encrypted Ø Encryption is a regular practice in desktop RTB auctions (~68%) Ø Lower but rapidly increasing in mobile RTB auctions (~30%) Encrypted Previous work [Olejnik, 2013] assumes encrypted prices follow the same distribution as cleartext. But is that so? Internet Measurements Conference [IMC'17] - P. Papadopoulos 12

Our approach 1. Leverage Real-Time Bidding (RTB) protocol: 1. 74% of programmatically purchased advertising 2. $8.7 billion in 2016 only in US 2. Methodology to calculate at real time the overall value advertisers pay per individual user based on her leaked information. 3. Year-long dataset (2015) of 1600 real users + 2 real probing ad campaigns Internet Measurements Conference [IMC'17] - P. Papadopoulos 13

Methodology Internet Measurements Conference [IMC'17] - P. Papadopoulos 14

YourAdValue browser extension Monitors RTB nURLs & collects features: (i) auction-specific metadata and (ii) personal data the user leaks while browsing the web Internet Measurements Conference [IMC'17] - P. Papadopoulos 15

Price Modeling Engine (PME) Internet Measurements Conference [IMC'17] - P. Papadopoulos 16

Required input for the PME Internet Measurements Conference [IMC'17] - P. Papadopoulos 17

The Evaluation… Internet Measurements Conference [IMC'17] - P. Papadopoulos 18

Evaluating our approach • Offline year-long (2015) dataset D with Metric D mobile traffic from 1600 real users Time period 12 months Impressions 78,560 IAB category 18 of publishers • Weblog Ads Analyzer: RTB publishers ∼ 5.6k/month Ø filter RTB traffic and Ø extract features (auction’s metadata and user data) from nURLs Internet Measurements Conference [IMC'17] - P. Papadopoulos 19

Feature extraction Internet Measurements Conference [IMC'17] - P. Papadopoulos 20

Features that affect prices (1/3) 100 100% Charge price (CPM) 80% 10 RTB share 60% 1 40% 20% 0.1 0% 1 2 3 4 5 6 7 8 9 10 1112 0.01 Month of the year Android iOS Android Windows Mob Mobile Device OS iOS Other Distribution of charge prices for the 2 most Percentage of RTB traffic for top mobile OSes. popular mobile OSes. Whiskers: 5 th , 10 th , 50 th , 90 th ,95 th percentiles More Android devices, but iOS-based devices draw higher prices Internet Measurements Conference [IMC'17] - P. Papadopoulos 21

Features that affect prices (2/3) 2 Charge price (CPM) 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 320x50 468x60 728x90 120x600 300x250 160x600 300x600 Ad-slot sizes Distribution of the charged prices per ad-slot size (sorted by area size). Whiskers: 5 th , 10 th , 50 th , 90 th ,95 th percentiles Larger ad-slot sizes do not necessarily get charged more Internet Measurements Conference [IMC'17] - P. Papadopoulos 22

Features that affect prices (3/3) • user location affects the median charge prices • during the day median charge prices are of similar range -> early morning hours - noon: more charge prices with increased values • some IABs are more costly than others -> (“Business & Marketing” more expensive than “Science”) See more about features in the paper… Internet Measurements Conference [IMC'17] - P. Papadopoulos 23

Price Modeling Engine: ad campaigns Internet Measurements Conference [IMC'17] - P. Papadopoulos 24

Real probing ad-campaigns • 2 real probing ad-campaigns in 2016 ( A1 , A2 ): various experimental setups Filter name Range of values (type) Cities Madrid, Barcelona, Valencia, Seville Time of day 12am-9am, 9am-6pm, 6pm-12am Day of week Weekday, Weekend Metric D A1 (enc) A2 (clr) Type of device Smartphone, Tablet Type of OS iOS, Android Time period 12 months 13 days 8 days Ad-format (smartphone) 320x50, 300x250, 320x480 or 480x320 Impressions 78,560 632,667 318,964 Ad-format (tablet) 728x90, 300x250, 768x1024 or 1024x768 IAB category 18 16 7 Ad-exchange MoPub, OpenX, Rubicon, DoubleClick, of publishers PulsePoint RTB publishers ∼ 5.6k/month ∼ 0.2k ∼ 0.3k Content category of publisher all IABs possible Internet Measurements Conference [IMC'17] - P. Papadopoulos 25

Cost per IAB in cleartext and encrypted prices Median encrypted prices always higher than cleartext Time shift: More recent cleartext prices are higher than the ones last year Comparison of CPM costs for the different IAB categories in our dataset and the 2 probing adcampaigns. Internet Measurements Conference [IMC'17] - P. Papadopoulos

How much do advertisers pay to reach you? Internet Measurements Conference [IMC'17] - P. Papadopoulos 27

Encrypted Vs. Cleartext prices “It’s safe to assume that encrypted prices follow the same distribution with cleartext prices.” price distribution of encrypted prices (A1): → distinctly different → about 1.7x higher median value than Comparison of price distributions between cleartext prices (A2) cleartext and encrypted, for different time periods and datasets (D vs. A1 and A2). Internet Measurements Conference [IMC'17] - P. Papadopoulos 28

How much do advertisers pay to reach you? 100% • Cumulative cost from encrypted prices: 90% 80% cannot surpass cleartext (still dominant). 70% 60% CDF 50% • some users more costly than others 40% 30% 20% 10% • median user costs 25 CPM 0% 10 -2 10 -1 10 0 10 1 10 2 10 3 10 4 (73% of the users cost < 100 CPM) Cumulative user cost (CPM) • 2% of users cost 10-100× more to the cleartext est. encrypted cleartext (time corr.) total ad-ecosystem than the average user! Cumulative CPM paid per user in our year-long dataset (2015) Internet Measurements Conference [IMC'17] - P. Papadopoulos 29

In summary… • Methodology to measure cost of advertisers per individual user: • At real time and on user’s side based on the user’s profile • Leverage ad-auctions and RTB’s price notification channel • Construct a model to estimate encrypted prices using as features auction’s metadata and user’s leaked info. (> 82% accuracy) • Our methodology is tested using a year-long dataset of 1600 real mobile users Internet Measurements Conference [IMC'17] - P. Papadopoulos 30

Takeaways ü Encrypted prices are 1.7x higher than cleartext ü Median user costs 25 CPM per year ü Taking into account several different factors: (HTTP+HTTPS, management and intermediaries costs, mobile+desktop traffic) → overall user ad-cost in the range of 0.54-6.85€ per year → cheaper than most users think (10s of euros*) * Your browsing behavior for a big mac: Economics of personal information online . WWW’13 Internet Measurements Conference [IMC'17] - P. Papadopoulos 31