Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Bidding Models Complex & Simple By Sandeep Dey


Published on

From the SMX Advanced Conference in Seattle, Washington, June 22-23, 2016. SESSION: The Mad Scientists of Paid Search. PRESENTATION: Bidding Models Complex & Simple - Given by Sandeep Dey, @sdey - Groupon Inc, Staff Data Scientist. #SMX #11B

Published in: Marketing
  • Be the first to comment

  • Be the first to like this

Bidding Models Complex & Simple By Sandeep Dey

  1. 1. #SMX #11B @sdey Optimizing Bidding for different Scenarios Bidding Models Complex & Simple
  2. 2. #SMX #11B @sdey Developing a Bidding Strategy What do you want to do ? 1.  Branded vs Transactional 2.  Reach vs Efficiency 3.  CPA vs CPC
  3. 3. #SMX #11B @sdey Target: Maximize Volume •  Concentrate on keywords which have low CPC but high volume •  Queries with lower transaction intent e.g. Facebook can sustain huge traffic even at 0.01$ •  Bidding becomes less important in such a scenario, budget maintenance becomes more important •  Bid to get Top of Page
  4. 4. #SMX #11B @sdey Target : Get revenue & new customers from spend •  Set up efficiency targets & budgets •  Go for transactional keywords •  Could be lower volume •  Mine tail keywords •  Bidding is the most important part of this strategy
  5. 5. #SMX #11B @sdey Theory of Bidding, Cost, Revenue & Efficiency Bidding Theory
  6. 6. #SMX #11B @sdey •  Clicks increase as bid increases (linear) •  Revenue increases as bid increases (linear) •  Cost increases as bid increases (quadratic) •  Efficiency increases as bid increases (linear) How does efficiency change with bids bid
  7. 7. #SMX #11B @sdey •  Maximize volume while hitting efficiency targets •  Profit is not the metric we are optimizing for Aim of the game : Hit target efficiency
  8. 8. #SMX #11B @sdey •  PLA •  Revenue follows a logit function •  EGR follows a crazy slope •  Broad Match Keywords •  Cost follows an exponential trend with bid increase •  Revenue follows a log trend with bid increase Not all keywords or programs are created equal
  9. 9. #SMX #11B @sdey Implementing Basic Bidding into an account The Basics
  10. 10. #SMX #11B @sdey Revenue per click x Efficiency target OR Revenue per click ROAS target Bid Calculation Basics for a Keyword Measures •  Revenue last 30 days •  Clicks last 30 days Parameters •  Efficiency target (EGR) •  cost / revenue •  ROAS target •  revenue / cost
  11. 11. #SMX #11B @sdey •  Different products sell at different times. •  Use exponential decay on revenue and clicks. •  Revenue from N days ago has decay ^ N weight •  Start with 0.95 decay factor Seasonality : Use exponential decay
  12. 12. #SMX #11B @sdey •  Down jackets & sweaters sell the most from Nov to Jan •  The sales volume decreases by half at the end of Feb •  Without a decay , the bids will follow seasonal trends very slowly Seasonality : Use exponential decay YEAR
  13. 13. #SMX #11B @sdey •  Difference between bid (max cpc) & cost (avg cpc) •  Bids have to be incremented by a bid headroom factor to prevent hyper efficiency •  Bid headroom = max cpc/ avg cpc Bid headroom
  14. 14. #SMX #11B @sdey •  Say you want to pay $.90/click •  Bid up at 1.10$ •  Bid up by 20% ( max cpc x clicks / cost ) Bid headroom
  15. 15. #SMX #11B @sdey •  Choose a time period to collect Revenue & Clicks (30 days) •  Keywords with clicks > threshold (100 clicks) can be bid with individual info •  Bid = revenue/clicks x efficiency target x bid headroom High traffic keywords
  16. 16. #SMX #11B @sdey •  Group multiple keywords to get enough clicks above a threshold •  Grouping can be based on adgroup & campaigns •  Grouping can be also based on category of the keywords. Low traffic keywords : Grouping keywords
  17. 17. #SMX #11B @sdey Low traffic keywords : Grouping keywords •  E.g. group electronics into a single category •  Group Apparels into a single category if needed
  18. 18. #SMX #11B @sdey •  Increase date range for low traffic keywords •  Less frequent keywords : longer time ranges •  Beware too long time ranges can’t follow seasonality Low traffic keywords : Increase date ranges
  19. 19. #SMX #11B @sdey Issues to keep in mind while implementing a bidding strategy The Troublemakers
  20. 20. #SMX #11B @sdey •  Keep a lookout for bad search queries & keywords •  If they have lower bids, they will reduce bids for the group they are clustered in •  Solution : Negative keyword them or be aggressive in their bid reduction Keywords which don’t convert
  21. 21. #SMX #11B @sdey •  Addition of bid modifiers would upset your bidding •  A negative bid modifier tends to make you hyper-efficient •  Final CPC are not equal to what is calculated •  Solution: measure the overall effect of bid modifiers on CPC and adjust Bid Modifiers
  22. 22. #SMX #11B @sdey •  Remarketing lists for search ads (RLSA) •  +100% increase for previous visitors •  Bid is set at 1$ , but CPC >1$ •  Solution : Figure out how audiences affect the CPC in total Bid Modifiers
  23. 23. #SMX #11B @sdey Machine Learned Models for Bidding for an ecommerce Advanced Bidding
  24. 24. #SMX #11B @sdey •  Transactional Text Ads Bidding •  Campaign Strategy centered on either landing on Product Page or Search Page •  Modeled into a Machine Learning Problem with targets as revenue per click Advanced Bidding for Ecommerce
  25. 25. #SMX #11B @sdey Target feature : Future Revenue Per Click Input Features : •  Historical Revenue Per Click •  Margin & Price •  Discounts •  Matchtype •  Keyword Relevance to Product Title Performance : •  Revenue lift of +20% with cost increase of +10% as compared to baseline Predictive Modeling at a keyword level