Advertising Technology

1) User visits a webpage with an ad slot. 2) Publisher's ad server or SSP creates a bid request containing: user info (anonymised ID, geo, device), page context (URL, category), ad slot details (size, format). 3) Bid request sent to ad exchange. Exchange broadcasts to connected DSPs (10-50+). 4) Each DSP has ~80ms to: look up user in their audience database, match against active campaigns, predict click/conversion probability using ML, calculate optimal bid price. 5) DSPs return bid responses with: bid price, creative (ad), and advertiser ID. Some DSPs no-bid if no matching campaign. 6) Exchange runs auction: In first-price auction (standard now), highest bidder wins and pays their bid. 7) Winning creative URL returned to browser. Ad rendered in the slot. 8) Impression pixel fires — tracked for billing, viewability, and fraud verification. Total time: ~85-100ms. At scale: Google processes 10M+ auctions per second.

Second-price (historical): Winner pays $0.01 more than the second-highest bid. Example: Bids are $5, $3, $2. Winner pays $3.01. Advantage: Bidders bid their true value (truthful bidding). Disadvantage: Revenue unpredictable for sellers. Google used this until 2019. First-price (current standard): Winner pays exactly what they bid. Example: Bids are $5, $3, $2. Winner pays $5.00. Challenge: Bidders have incentive to bid lower than their true value (bid shading). Solution: DSPs use bid-shading algorithms — ML predicts the minimum bid needed to win, avoiding overpaying. Why the shift: Header bidding meant publishers ran multiple auctions simultaneously. In a mix of first and second-price auctions, arbitrage was possible. First-price is simpler and more transparent. Impact: DSPs invest heavily in bid-shading ML models. Google's shift to first-price in 2019 was a major industry change.

Traditional targeting (cookie-based): Third-party cookies track users across websites. User visits a shoe site → cookie dropped → later sees shoe ads on news site. Data Management Platforms (DMPs) aggregate cookie data to build audience segments. How it's changing: 1) Chrome deprecating third-party cookies (Privacy Sandbox). Safari/Firefox already blocked. 2) Apple ATT (App Tracking Transparency) — 80%+ of iOS users opt out of tracking. New approaches: 1) First-party data: Advertisers use their own customer data (email, purchase history). Most valuable. 2) Contextual targeting: Target based on page content, not user data. 'Show car ads on car review articles.' ML analyses page text in real-time. 3) Unified ID 2.0: Opt-in email-based identity system. Encrypted, privacy-safe alternative to cookies. The Trade Desk leads this. 4) Google Topics API: Browser determines user interests from browsing history, shares broad topics with advertisers. 5) Data clean rooms: Advertisers and publishers match data in secure environments without sharing raw data. Impact on India: Less reliance on cookies (mobile-first market uses device IDs more), but ATT affects iOS targeting significantly.

Types of ad fraud: 1) Bot traffic: Automated scripts that load pages and 'view' ads. Cost advertisers billions. Sophisticated bots mimic human behaviour — mouse movements, scroll patterns. 2) Click farms: Rooms full of people or devices repeatedly clicking ads. Common in mobile app install fraud. 3) Domain spoofing: Fraudulent site claims to be a premium publisher (e.g., pretending to be Times of India). Ads bought on premium domains actually shown on low-quality sites. 4) Ad stacking: Multiple ads layered on top of each other. Only top ad visible, but all report 'impressions.' 5) Attribution fraud: Fake installs/conversions reported to steal credit for organic installs. Detection methods: 1) Browser fingerprinting — bots have detectable anomalies in JavaScript execution. 2) Behavioural analysis — ML detects non-human patterns (too-perfect click timing, no scroll). 3) ads.txt / sellers.json — publisher lists authorised sellers, preventing domain spoofing. 4) IP analysis — data centre IPs, VPN detection, geo inconsistency. 5) Device signals — emulator detection, suspicious device IDs. mFilterIt's approach: Real-time analysis of every impression/click with 50+ fraud signals. Blocks invalid traffic before advertiser is charged.

Before header bidding (waterfall): Publisher's ad server called demand sources sequentially — first Google, then Demand Source B, then C. If Google bid ₹20, Source B never got a chance even if they'd bid ₹30. Unfair and suboptimal. Header bidding: All demand sources bid simultaneously in a client-side JavaScript auction (Prebid.js). Every impression gets competitive bids from all partners at once. Impact: 1) Revenue increase: Publishers saw 20-40% revenue uplift because competition increased. 2) Fair auction: All demand sources compete equally — no preferential ordering. 3) Transparency: Publishers see all bids from all sources. Evolution: Client-side header bidding (Prebid.js in browser) → Server-side header bidding (server runs auction to reduce latency) → Google's Open Bidding (server-to-server alternative). Current best practice: Hybrid approach — Prebid.js for key partners + server-side for additional demand. Google Ad Manager unified auction now incorporates header bids. This is how publishers maximise fill rate and CPM.