1. Technical Field
This invention relates generally to the field of advertising. More specifically, this invention relates to predicting user response to advertisements.
2. Description of the Related Art
An advertisement creative describes any type of advertising content or image that advertises a product or service. Advertisements displayed on a web page are called impressions.
The payment model for displaying impressions can be a flat fee, or more likely, a combination of a fee for displaying the impression and a fee for any instance where a user clicks on the advertisement, i.e. a click-through. Some payment models even include a conversion fee. See, for example, Google® U.S. Publication Number 2008/0103887. A conversion occurs when a user both clicks an advertisement and purchases either the product or service being advertised.
Typically, different advertisers bid for the same ad space. Because a user is more likely to click on a targeted impression, publishers have developed a variety of ways to personalize the impressions. Google®, for example, sells ad spaces that are paired with keywords entered into Google®'s search engine. The pairing results in a higher click-through rate. For example, if a user types “bird seed” into Google®'s search engine, ads relating to the sale of bird seed are served.
This method of pairing ad space with search terms, furthermore, is especially advantageous for companies that sell specialized products because the ad space is cheaper than for popular terms, e.g. “car,” but the click-through rate is much higher because it is more likely that the user is looking to purchase that specific item. For example, people who own parrots frequently buy foraging toys to keep the parrots entertained. When a user enters “foraging toys” into the Google® search engine, only a few sponsored links appear with the search results because the term is rare. However, a user looking for these toys is much more likely to click on one of the links than a user that employs “car” as a search term.
The drawback to these methods is that although the advertisements are targeted, they only reflect one dimension of a user. Google® developed a more detailed mechanism for personalizing search results. See, for example, U.S. Publication Number 2005/0240580. In this approach, the server orders search results for a user according to information gleaned from the user's Internet activity, e.g. previous search queries, uniform resource locators (URLs) identified by the user, anchor text of the identified URLs, general information about the identified documents, the user's activities on the identified documents, sampled content from the identified documents, category information about the identified documents, the user's personal information, and the user's browsing patterns. This approach is limited, however, because it only tracks a user's activities when the user is logged-in to Google®. Furthermore, because the system is predicting user behavior based on that user's previous behavior, the prediction is only useful for predicting that the future behavior conforms to previous behavior. This method cannot make predictions about new areas for which the user develops an interest.
In one embodiment of the invention, user data is collected from a variety of sources, e.g. Internet activity. The data is compiled and segmented according to subject matter. The segments are used either in a behavioral model or organized according to pre-defined rule segments. User data is scored against the behavioral model or rule segments in real-time. The closest matching advertisements are served on the web page. The user's reaction to the advertisements is recorded by the cookie and transmitted to the server to further refine the segments.
In one embodiment, the invention comprises a method and/or an apparatus for collecting user data, creating behavioral segments, scoring ads according to the segments, serving targeted ads to users, and recording user responses to further refine the behavioral segments.
The client 100 includes a computer-readable medium 310, e.g. random access memory (RAM), coupled to a processor 300. The processor 300 execute s computer-executable program code stored in memory 310. Other embodiments of a computer-readable medium 310 include, but are not limited to, an electronic, optical, magnetic, or other storage or transmission device capable of coupling to a processor, e.g. flash drive, CD-ROM, DVD, magnetic disk, memory chip, ROM, etc.
In one embodiment, the system includes multiple client devices 100 that communicate with a server 110 over a network. The network comprises the Internet. In another embodiment, the network is a local area network (LAN), a wide area network (WAN), a home network, etc. In one embodiment, the network is implemented via wireless connections.
Data Collection
The Website host installs a cookie on the client 100 that tracks a user's behavior on that host's Website, e.g. eBay can collect data for each user. The cookie is associated with a specific ISP address. As a result, when different hosts install cookies on the same client 100, the data is reconciled according to the ISP address once the data is compiled. In one embodiment, the cookie records user profile information from the website, e.g. age, location, income, educational status, job category, gender, etc. The cookie receives data from a Java script that runs on the client 100. The Java script is displayed as a one by one pixel on the client's 100 display.
In one embodiment, the cookie sends data directly to the server 110 that contains the data collection component 400. In another embodiment, the cookie sends data to a third-party server, e.g. the distributed datastore provided by Akamai of Cambridge, Mass., which transmits the information to the server 110. A third-party server insulates the rest of the system from being bombarded with data from the multiplicity of clients 100. Other advantages of using a third-party server will be apparent to a person of ordinary skill in the art.
The cookie collects data for two different groups: for advertisers 140 and for the data collection component 400 stored on the server side profile 110.
Data collection for advertisers 140 is triggered by a beacon that is embedded in the ad space. The beacon responds to a pre-determined rule. For example, using
Data is stored in the data collection component 400. In one embodiment, the cookie records the category and event of each website visited by the user. The category refers to the source and type of website or search terms, e.g. car. The categories are segmented into more precise categories to encompass both the highly specific and the general. For example, Range Rovers is a specific example of a sport utility vehicle (SUV), which is a type of car. The event refers to an action taken by the user. For example, when the user visits a website advertising cars, the cookie records whether the user browsed cars, bought a car, bid on a car, searched for cars, etc.
In another embodiment, the cookie records the frequency of a user's visits and the frequency of events, e.g. the number of times a user searches for a Range Rover. The frequency can be characterized three ways: velocity, intensity, and persistence. Velocity refers to the rate at which a user visits a web page. For example, a user may visit a car website infrequently when he's considering buying a car and then more frequently when he's ready to purchase. That behavior is characterized as an increasing velocity. Intensity refers to how many times a user visits the website and tracks the images that the user views while using the website. Persistence measures whether the website is regularly visited by the user, e.g. eBay, Amazon, a favorite blog or whether the website was a one-time occurrence.
The cookie collects other pieces of information that are useful for categorizing user behavior. For example, each website that the user visits is tagged with a description, e.g. financial section of the New York Times. The cookie collects these tags. In addition, the cookie collects the browser speed because users with higher browser speeds are more likely to be earlier adaptors of technology.
This information is stored on the cookie and transmitted to the server side profile 110. The cookie can only store 4K of information. The cookie transmits this information to the server side profile 110 in real time, but because the cookie can be used as a back-up data storage device for determining which ad to serve to a user, the cookie must be kept relevant. Thus, the cookie continually discards the least valuable data elements. The value of the data is determined by frequency and time. Thus, if a user visits the NY Times daily, this information is kept on the cookie. However, if the user visited the Washington Post only once in the last month, that information is discarded.
Another source for information is proprietary data that is particular to a party and is only used to generate behavior predictions for that party. For example, a cell phone manufacturer installs a cookie on a user's machine when the user visits the manufacturer's website. The cookie transmits data about the user's activities while on the website, e.g. searching, completion of a registration form, etc. This data is important for gauging the user's level of interest, e.g. researching cell phones, intending to purchase a cell phone, already owns a cell phone, etc.
The information is used, along with other segmented data such as the user's activities on Amazon® to determine what type of ads to serve to the user while the user is on the manufacturer's website. For example, if it is clear that the user is about to make a purchase of a cell phone, the ad can offer a 10% discount. Furthermore, because the compiled data may include demographic information, previous purchases from other websites, etc. the behavior prediction can be even more specific for the user and predict down to the dollar how much the user is likely to pay for a cell phone. Because information gleaned from the cell phone manufacturer website is proprietary, it is kept separate from the rest of the information and is not sold to competitors.
Lastly, the data collection component 400 stores response data to advertisements selected by the ad scoring component 430. The response data includes the type of response, e.g. impression, click, and purchase and transformations associated with the response, i.e. time between impressions, time between clicks, and frequency of purchases.
In one embodiment of the invention, the server side profile 110 is a distributed data environment where multiple servers capture the data from cookies.
In one embodiment, the server side profile 110 receives data from cookies, user profiles, and other sources 120. The other sources 120 include information that is not collected via the Internet. In one embodiment, other sources include information about purchases made through catalogs that are organized according to demographics, telemarketing information, etc.
Segmentation
The data collection component 400 transmits the data to the segmentation component 410 for segmentation. In one embodiment, the data is segmented according to four categories: demographics, contextual, integrated user profile, and historical data. Demographic data includes the location of the user, age, race, income, educational attainment, employment status, etc.
The subject matter searched by users is categorized using a data tree structure. In a data tree, the subject matter becomes more subdivided as the tree branches until the subject matter can no longer be divided any further, at which point the subject matter is referred to as a leaf node. For example, if a user is searching for a specific kind of watch, the categories may proceed from the following: jewelry and watches→watches→wristwatches→military watches→Czech military watches. Other methods of organizing data will be obvious to a person of ordinary skill in the art.
Behavioral Modeling
In one embodiment, the system includes behavior modeling for predicting a user's actions. The modeling component 450 groups segmented behaviors of multiple users together. For example, one group includes people that are interested in Czech military watches. Depending on the behavioral model, this “interest” can be defined as people who purchased Czech military watches, people who searched for those terms, people that purchase military watches, etc. The model also takes into account user frequency, i.e. velocity, intensity, and persistence.
The modeling component 450 makes associations between groups of users and the categories of interest. For example, people interested in Czech military watches may also be interested in typewriters or antique cars. People interested in buying designer purses may be interested in purchasing fashionable clothing or new televisions. In one embodiment, these categories are correlated using regression analysis.
Once the models are complete, users are compared to the models using real-time data to predict the user's similarity to user groups and, as a result, the user's likelihood of being interested in certain categories. For example, a user that purchased a laptop designed in the last year is grouped with other laptop purchasers. Those purchasers frequently purchased plasma televisions. Thus, the user being compared to the groups is likely to purchase a plasma television as well. As a result, the system uses cross-marketing to target a larger number of users while maintaining a high likelihood of success.
The user's behavior prediction is modeled in real time. The modeling component 450 queries the data collection component 400 for the user's data to predict the user's future actions. If the modeling component 450 is unavailable, the modeling component 450 queries the cookie, which provides up to 4k of the user's previous activities.
Ad Scoring
The prediction of a user's behavior is used by the ad scoring component 430 to predict the user's reaction to impressions for different categories. For example, in the above example, the user is likely to click on ads for laptops, but is not likely to click on ads for window treatments. The ad scoring component 430 assigns a score that represents the likelihood of a positive response to an ad. If an advertiser provides multiple ads, each ad is scored according to a segment and the ads are prioritized according to the highest score.
Matching is a function of the ad score and an advertiser's bid. For example, in one embodiment the scale for the score is between 0 and 1. Advertiser A has a score of 0.3, meaning that there is a low correlation between the segment and the advertisement, but the advertiser is willing to pay $1.00 for each impression served. Advertiser B has score of 0.9, but is only willing to pay $0.50 for each impression served. The score is multiplied by the bid, i.e. Advertiser A=0.3 and Advertiser B=0.5. Thus, even though Advertiser B is paying less for the ad, Advertiser B's ad is served because it is much more likely to result in a click through.
The publisher typically charges both for displaying the ad and for additional actions, e.g. click-through, conversion, etc. Thus, in the above example, even though the publisher receives less money for displaying the impression, the publisher makes more money because the user is more likely to click on the ad.
Once an ad is selected, a log file is generated. The log file identifies the ad, a full data profile, and a full segment membership at the time of the ad call. The log file is stored as part of the cookie and is transmitted to the server 110.
The ads are retrieved from an ad database 440 and transmitted to the publisher 130. The ad database 440 can be stored on the server 110 or provided by an advertiser 140. The ads are sent to the publisher for insertion into a web page space.
Rule Building
In another embodiment, a rule building component 420 generates rules for serving ads. Multiple rules can be used by the same segment by connecting rules using Boolean operators, i.e. and, or, not, etc. The rules can be part of a nested query, i.e. subqueries are defined by using parentheses. Rule combinations are associated with a unique segment identification (ID) and a user-generated segment description. A graphical user interface (GUI) is displayed for building rule combinations.
The event type 520 signifies the action that the user performs. For example, if the cookie were tracking a user on the eBay® website, the event type is viewing an item, browsing in general, searching for a specific item, watching an item, bidding on the item, or purchasing the item. If the user is searching for items on the Amazon® website, the user might place the item in a shopping cart or purchase the item. Purchase is defined as the achievement of a pre-defined goal. Thus, purchase could be paying money for an item or a user submitting a telephone number or home address. Recency 530 connects the event type 520 with the frequency 540. For example, a publisher may want an impression served to all users that bid on an automobile in the last month.
In one embodiment, additional rules are generated by selecting the “add a category” button 550. These rules are connected using conventional Boolean terms. These rules constitute an ad campaign.
Feedback
When a publisher 130 designates a space on the web page for an advertisement, the space is associated with an ad code. If the ad space is large enough, it may even require multiple ad codes. Different websites with the same publisher receive different ad codes.
The ad codes are used to track the impressions served. A list is generated of the ad code, the impression, the time it was served, and the reaction, e.g. whether a user clicked on the ad. This is helpful for tracking users' reactions to the ads. Studies show that a user is more likely to purchase something when they see an advertisement for it multiple times on different websites. Tracking the impressions allows advertisers to present the ideal purchasing situation.
Impressions are also tracked to determine behavioral characteristics. For example, are people more likely to make discretionary purchases on the weekends, at work, etc. Furthermore, because the ad code is associated with a particular website, the user's visit to a particular website may help characterize the user's needs. For example, if the user is recorded visiting Kelly Blue Book, the user may be interested in purchasing a car. Behavioral characteristics can even include visiting patterns of an advertiser brand by an advertiser category. For example, with enough information, the system can predict that a user that buys clothing at the Gap® is likely to buy clothing at Old Navy® but not at Banana Republic®.
Lastly, impressions are tracked to prevent problems such as user fatigue. For example, a user may be less likely to make a purchase if the same ad is served more than four times in one day.
Displaying Advertisements
Now that the individual components have been described, it is possible to lay out the steps for generating segments according to the system illustrated in
A cookie is installed 700 on the clients 100. The cookies transmit 705 data to a server 110. The data collection component 400 receives 710 data from the clients 100 and other sources. The segmentation component 410 generates 715 segments for all user data according to categories. In one embodiment, the modeling component 450 generates 720 predictions of user behavior. In another embodiment, the rule building component 420 generates 725 rules.
The ad scoring component 430 contains a database of ads 440. Each ad is associated with the price that the advertiser is willing to pay for displaying the ad and/or the click-through cost. The ads are scored 730 against either the predictive model or the rules, depending on the selected embodiment. If one advertiser has multiple ads selected for the same spot, the ads are prioritized 735 according to the ad score.
In one embodiment, the response is recorded 845 as part of the log file. In another embodiment, the response is recorded 845 directly in the cookie. The cookie transmits 850 the response to the server 110. The response is received by the data collection component 400 and used to further refine the segments. The feedback mechanism reinforces accurate predictions.
As will be understood by those familiar with the art, the invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. Likewise, the particular naming and division of the members, features, attributes, and other aspects are not mandatory or significant, and the mechanisms that implement the invention or its features may have different names, divisions and/or formats. Accordingly, the disclosure of the invention is intended to be illustrative, but not limiting, of the scope of the invention, which is set forth in the following Claims.
This patent application claims the benefit of U.S. provisional patent application Ser. No. 61/102,317, Turn Segment (Rule) Builder Requirements, filed Oct. 2, 2008, the entirety of which is incorporated herein by this reference thereto.
Number | Date | Country | |
---|---|---|---|
61102317 | Oct 2008 | US |