Dynamics in Clickthrough and Conversion Probabilities of Paid Search Advertisements

Anoek Castelein, Dennis Fok, Richard PaapSSRN 2019 · DOI

TL;DR

This paper introduces a dynamic Bayesian model to understand how clickthrough and conversion probabilities in paid search advertising change over time. By accounting for these dynamics, advertisers can develop more effective strategies and better estimate ad performance, leading to improved profitability.

What Problem Does This Paper Address?

Search engine advertising (SEA) is a crucial marketing channel, but understanding how ad performance metrics like clickthrough rates (CTR) and conversion rates (CR) evolve over time is challenging. These probabilities are influenced by various factors, including changing consumer preferences, new product launches, and competitor actions. Traditional models often treat these probabilities as static, failing to capture their dynamic nature. This lack of insight hinders advertisers from optimizing their strategies, such as setting appropriate bids and crafting effective ad copy and landing pages. Furthermore, issues like data sparsity (many keywords with little traffic) and endogeneity (bid and ad position influencing each other) complicate accurate estimation.

Methods

The authors propose a dynamic Bayesian model that addresses these challenges. Key features include:

  • Time-Varying Parameters: The model explicitly allows CTR and CR to change over time, capturing both permanent and transitory dynamics through AR(1) processes. This helps distinguish between short-term fluctuations and long-lasting shifts in performance.
  • Handling Data Sparsity: By linking keywords based on common semantic characteristics, the model can leverage information from related keywords to improve estimates for sparse ones.
  • Addressing Endogeneity: The model simultaneously estimates consumer behavior (CTR/CR), search engine behavior (ad position allocation), and advertiser behavior (bidding). It uses instrumental variables (lagged performance metrics) to correct for endogeneity between ad position, bids, and performance.
  • Bayesian Inference: A Gibbs sampler with Polya-Gamma data augmentation is employed for efficient Bayesian inference, naturally handling missing data.

Datasets

The study utilizes a unique dataset from a Dutch online retailer selling laptops. This dataset contains historical performance data for 14,710 keywords measured daily from January 2014 to March 2016. It includes metrics like impressions, clicks, conversions, average ad position, and cost-per-click (CPC). Semantic characteristics of keywords (e.g., specificity, brand, match type) are also incorporated.

Key Insights and Ideas

  • Dynamics are Crucial: The research strongly emphasizes that CTR and CR are not static. Ignoring these dynamics leads to suboptimal advertising strategies.
  • Permanent vs. Transitory Shocks: The model's ability to differentiate between permanent and transitory shocks is vital. Permanent shocks require a more fundamental revision of SEA strategies, while transitory shocks might be managed with more frequent adjustments.
  • Endogeneity Matters: The findings confirm that ad position and bidding are endogenous. Purely predictive models that don't account for these relationships will likely misinterpret causal effects.
  • Keyword Specificity and Brand Effects: Different levels of keyword specificity (e.g., generic vs. retailer-specific) and brand-level shocks have distinct impacts on CTR and CR, with some effects being more persistent than others.

Key Results

  • Substantial Time Variation: The empirical application reveals significant and persistent time variation in clickthrough and conversion probabilities, underscoring the need for dynamic SEA strategies.
  • Shock Persistence: Shocks to clickthrough probabilities tend to be permanent or highly persistent across various keyword types. For conversion probabilities, brand-level shocks are often more transitory.
  • Model Superiority: The dynamic model significantly outperforms a static model in both in-sample and out-of-sample forecasting, as evidenced by highly negative log Bayes factors.
  • Position and Bidding Endogeneity: The model confirms the presence of endogeneity, with positive correlations between unexplained parts of CTR and ad position, and a negative correlation between position and bid error terms.

Limitations and Future Work

  • Computational Cost: Acknowledged limitation is the substantial computation time required, particularly for drawing auxiliary Polya-Gamma variables.
  • Bid Proxy: The study uses Cost-Per-Click (CPC) as a proxy for bids due to data limitations. Actual bids might contain more information.
  • Missing Data: While the model handles missing data, the possibility of data missing not at random (e.g., advertisers not bidding on poorly performing keywords) could affect results.
  • Future Work: The authors suggest extending the model to include correlations across time-varying parameters for CTR and CR, and using latent factors instead of pre-specified keyword characteristics to explain ad performance. Determining optimal bids, which requires additional information on spillover effects and cross-channel impacts, is also noted as a separate, complex problem.

Summary updated

BibTeX
@misc{castelein2019dynamics,
  author = {Anoek Castelein and Dennis Fok and Richard Paap},
  title = {{Dynamics in Clickthrough and Conversion Probabilities of Paid Search Advertisements}},
  year = {2019},
  doi = {10.2139/ssrn.3439291},
  howpublished = {SSRN},
}
Figure