Transcript Bidding strategy should achieve some goals, typically

Predictive Analytics World
Predictive Keyword Scores to Optimize PPC
Campaigns
Vincent Granville, Ph.D.
Click Forensics
February 19, 2009
CONFIDENTIAL 1
Problem
• Advertisers bidding on keywords on search engines (PPC
programs offered by Google, Yahoo, etc.)
• Bidding strategy should achieve some goals, typically
–
–
–
–
Profit optimization
ROA (revenue on ad spend) optimization
Minimization of cost of user acquisition
Maximization of user lifetime value
Short Term Goals
• Short term ROA can be negative
• Paid + organic search usually provides positive ROI
• Organic search used as a leverage to buy traffic and increase
reach
• KPI’s:
–
–
–
–
Clicks per keyword
Conversions per keyword
Revenue, profit or return
Conversion rate
Issues
•
•
•
•
Keywords with few clicks (“long tail”): difficult to predict
Attaching a conversion to a click: data quality (cookies)
Revenue numbers not known until tomorrow
New bid => Google needs to “learn” how to handle it
– Real time implementation of keyword bidding subject to high volatility
– Focus on end-of-day or bi-weekly algorithm
– Pitfall 1: if max bid is much higher than actual CPC => Google will
eventually notice!
– Pitfall 2: keyword performance can be impacted by “poor” keywords in
same ad group, or by impression fraud / click fraud spikes
– Match type
Keyword Scoring
•
•
•
•
Same as click scoring / credit card transaction scoring
Scores computed at the keyword / ad group level
Response: RPC, Conversion rate, etc.
Independent variables: binary rules
– Actually, there are highly auto-corraleted
• Model
–
–
–
–
–
Logistic regression (ridge or constrained regression)
Naïve Bayes (related to logistic regression via the odds ratio)
Decision trees or combo
Score is predictor of RPC, return or conversion rate, etc.
Conversion blending
Bidding Strategy
• Goals
–
–
–
–
Be able to predict response for keywords with very little historical data
Be able to predict response for new keywords
Conversion rate = f(score)
New bid = g(previous bid, keyword score, ROI, RPC, …)
• Methodology
–
–
–
–
Permanent multivariate (A/B/C) testing
g is a parametric function
A/B/C: each case corresponds to a particular parameter set
Moves in parameter space driven by a simulated annealing algorithm
Examples of Rules
• Text mining rules used in the keyword scoring engine
–
–
–
–
–
–
Length of keyword
Number of terms
Keyword contains “free”, “new”, “2009”
Keyword contains digits
Keyword contain top 1-term word with known response
Keyword contains 2-term word with known response
• Example
– Keyword “used car Honda 2000” contains the 2-term word “used car”
– All keywords containing “used car” have on average a 5% conversion
rate
Issues
•
•
•
•
Keyword cleaning
Each keyword contains multiple 1-term and 2-term words
Scalability
Most keywords contain at least one top term
– With 50MM keywords and 25K top terms, 95% of the keywords
contain one top term (at least)
– Response is not known for most of the 50MM keywords (too
granular), but it is known for each of the 25K top terms (aggregate
level)
– Works with new keywords
Results Test Data
Results: Interpretation
• Keyword score is a good predictor of conversion rate
• Bids are too high on good keywords, too low on poor
keywords
• Simple corrective action suggested
– A/B/C parametric bidding strategy not discussed here
• Cross validation: see next slide
Results: Cross-Validation
• Process 15 days worth of data using score lookup tables
based on training set
• No time period overlap, between training and test
• Keyword overlap
• Large volume of new keywords (“new” means a KW not
found in training set)
• Robustness against missing data / new keywords
• Predictive power somewhat reduced, but still good