This article presupposes a basic knowledge about software systems, stochastic processes and online or internet advertising methodologies prevalent among search/internet engines and content networks.

As mentioned in the description and abstract, this topic delves upon an algorithmic software approach to placing search based and content based advertisements for individual customers and on behalf of advertisement agencies.

This topic is not verbose by design as it accentuates the functionality needed from a software stand point and also the computational ground rules needed for such a system or portal to exist, operate optimally and also learn continually as the ground rules change.

Such a software system or a portal if you will can also be used as an internal tool by advertising agencies for their client submission needs to the various internet engines. This software system also addresses the needs for being search/internet engine agnostic, transparency when it comes to results presentation and arriving at an optimally minimum cost per click which results in the best bang for the buck for the customer.

Given below are some of the software functions that such a system will possess that will aid in realizing the requirements mentioned above and also in the abstract.

Software Functions

• Campaign Objectives and Management
• Real time data aggregation
• Real time enterprise search/internet engine integration
• Search Keyword Targeted Campaigns
• Click to Call Campaign
• Content Network Campaigns
• Site Targeted Campaigns
• Campaign Aggregation and Results Presentation
• Campaign pre-execution planning
• Campaign post-execution optimization
• User Registration and Administration
• Campaign Dashboards and Reporting

Advanced Software Functions

• The software system also provides a way to keep track of changes to search/internet engines and content only sites and networks to come up with a way that enhances the solution offerings periodically.
• Provide a methodology to infuse new keywords or keyword buckets to an existing campaign. This will enhance the optimization that is done post-execution with existing parameters by improving its effectiveness.
• Provide BI solutions to send flags to people internal to the company if a campaign does not fit the ROI metric
• Provide an automated way to read all customer landing pages to monitor changes
• Provide a configurable rules engine that can be used by an analyst to create, modify and add rules to ascertain the word buckets at the start of the pre-execution process for all campaigns.
• Provide a way to apply newly configured rules to selected campaigns as part of a test and also provide a way to apply the tested rule to all campaigns.

In order for campaigns to be organized based on customer needs, keywords are bucketed within a hierarchy with the suffix 1 being the highest within the hierarchy.

• Search Keyword Bucket Analysis

(Rules for forming word buckets before getting related keywords from the search engines and bidding on the words)

• Bucket 1 – Words with brand or company name
• Bucket 2 – Bucket 1 typos
• Bucket 3 – Meta-tags / page descriptors in customer’s website
• Bucket 4 – Competitor brand names
• Bucket 5 – Words on competitor websites

Once the appropriate keywords are ascertained and bucketed into the hierarchy, bids need to be placed with the search/internet and content based engines.

• Search Keyword Bid Management
• The bid management will also follow bucketing hierarchy in that the price sensitivity for the buckets higher in the list will be lower than the buckets lower down the hierarchy
• The bid management of words or word lists within each word bucket will use the prefix ordering algorithm to rank words based on expected-profit-to-cost-ratio (or bang for buck)
• In this algorithm the average expected profit generated converges to near-optimal profits, with the convergence rate that is independent of the number of keywords and can also be extended to a special case where both click through rates and expected profits are unknowns.
• Bid management will also be done using predictors. The two predictors will be search engine based bids and also keyword density based bids. The keyword density based predictor will have the higher rank.

Once the keyword buckets and their associated prices are determined, the software system presents the information to the end customer based the pricing models given here.

• Pricing Models
• Fixed Pricing
• The customer is presented with the plan based on the campaign objectives consisting of the price as determined by the bid management algorithm using the bucket hierarchy.
• Price bundling can also be used to arrive at the final price to the customer for the combination of search marketing and content advertising across search engines and content networks.
• Arbitrage Pricing
• Changes due to campaign optimization can result either in a positive or negative arbitrage which is repatriated appropriately

Once the campaign is live and executing, meaning the software system has placed the necessary bids on the keywords or keyword buckets across search/internet engines and content networks, the real time enterprise integration within the system will start to get results from the providers.

These results can be analyzed to better the efficiency of the campaigns hence the ROI for the customer.

• Campaign Optimization

• Each campaign can have multiple ad packets. Each ad packet has multiple keywords or keyword buckets associated with it. The prices can be adjusted at the campaign, ad packet, keyword bucket and the keyword level.
• Once the campaign is created, it is left to run as-is for the first 2 days
• On day 3, campaign history is checked at a preset frequency for the past 2 days to get the density of visitors for each time frame
• For each of the above time frames, determine the words and the related ads for those words based on the density and time period
• This is done to eliminate the possibility of showing ads with a lower relevance
• Check before and after each time frame to see what the metrics were for the following variables:
• Word, bid density, click thru rates and price
• Have a way to improve the ROI in either or both these ways:
• Bid on lower prices
• Use the metrics that were collected above to find the time frame that has better prices
• Increase the budget
• Tweak the contribution margin of each day of the campaign to increase the spend
• If the ads are to be shown only at a certain period of time, change it to be shown evenly through the day as the increased budget will be able to support such a change.
• Prices will be revisited when the following two conditions are true:
• Competition not there, i.e., if the bid densities are lower
• Increased volume of searches
• i.e., when the traffic estimates of when the bid price was set is lower than the traffic estimate of the time frame in question

The optimization can also be done using the classic 'Gradient Maximization' technique with scale factors that are based on the data collected from the customer, data used to model campaigns and also data collected from search/internet engines from their analytics modules.

At a high level, the scale factors are based on and calculated thusly.

• Click thru rates or ratios
• Price momentums
• Price contrarian
• Expected profit ratios (if available)
• Price to traffic ratios or rates
• Return on investment (if available)

The gradient maximization algorithm maximizes a function by an iterative process. The function of interest has several variables for which we want to find the values that maximize it. To do so, we first initialize the variables and compute the gradient of the function. The gradient provides the direction of change of each variable that increases the function value. We then modify the values of the variables in the direction given by the gradient and calculate the function value. If an improvement is found, we compute the gradient again and continue this process until we get no more increase in the function value.

At this point, we have reached a local maximum. Since we cannot find and be certain of a global maximum in such data, we repeat this entire process with several different initializations of the variables and compare the maximums obtained to arrive at the final value.

Notes