Description of Methodology

This audit is focused on Google's pay-per-click, video impressions, TrueView views, and video viewability advertising systems. Google provides these advertiser-facing solutions through Google Ads and publisher-facing solutions through AdSense and YouTube.

Only the Video Viewability Report is submitted for MRC Video Metrics accreditation.

Google advertisements may be administered to users through the following products or services: AdSense for Content (AFC), AdSense for Domains (AFD), AdSense for Search (AFS), Ad Exchange (AdX), YouTube, and Google.com. AFC relates to advertisements displayed on the pages of a partner site, where the context of the information on the page is used to determine and display relevant advertisements. AFD relates to advertisements displayed on the pages of a particular domain, where the domain name itself is analogous to a search query. AdX relates to advertisements displayed on participating partner sites, where the context of the page and real-time bidding is used to determine and display relevant advertisements. YouTube relates to advertisements displayed on YouTube.com or in the YouTube app, where the context of the video and search queries are used to determine and display relevant advertisements. AFS and Google.com relate to advertisements displayed as paid results within the context of search engine query and results.

Google's non-video impression-based advertising solutions, such as Google Marketing Platform, and systems which measure clicks for non-commercial purposes (such as Google search) are outside of the scope of this audit. Other systems outside the scope of this audit include related support and management systems such as Google Analytics. In addition:

• Google Ads report builder and other dashboard video metrics are not part of the accreditation process.
• Other device types are also not part of the accreditation process.
• Video campaign and App campaign Clicks are not a part of the MRC accreditation process for Clicks.
• Dashboard metrics segmentation (for example, demographics) beyond display and search campaign totals and device type segmentation, are not part of the MRC accreditation process.

The measurement methodology is based on all click activity recorded, and does not utilize sampling for the purposes of click measurement. Only stage 2.2 of the IAB click referral cycle (measured clicks) are directly observed by Google. With respect to the click-referral-cycle, upon receipt of the initial click transaction by the ads redirect server, Google Ads records the click and issues a non-cacheable HTTP 302 redirect to the browser based on the location established by the Advertisers for the specific advertisement. This constitutes the measured click. The click measurement methodology is the same across all device types (desktop, mobile, tablet) and for browser and mobile apps unless otherwise noted.

Furthermore, currently, parallel tracking is mandatory for Search, Shopping, Display, and Video campaigns. Parallel tracking sends customers directly from your ad to your final URL (landing page) while click measurement happens in the background (without sending them to the tracking URLs first).

Ads can be displayed on mobile devices that are supported by the Google Mobile Ads SDK (see a list of currently supported platforms).

A known limitation of this method of measuring clicks is that a network interruption may cause a user who successfully receives a 302 redirect to not be able to view the resulting advertiser web site.

The counting methodology utilized is the multiple-click-per-impression method. Consequently, to avoid inappropriate counting of navigational mistakes (for example, multiple clicks per user), we require that the time between a given click and a previous click on the ad impression is greater than a specific period of time.

Logs are generated and processed in real-time, storing all data associated with observed HTTP transactions. Numerous variable and heuristic techniques are utilized to implement the click filtration systems, which will not be enumerated here to protect their security.

Both Google and their partners deliver the advertisements to users; however, Google maintains control over and performs the processes related to measurement and advertiser reporting of click activity. When a user clicks on an advertisement, whether delivered by Google or a partner, via any one of the products administering the product, the click activity is tracked by Google Ads through the ads redirect servers.

Measurement of click activity is based on the Google Ads click measurement methodology, which utilizes a technology infrastructure to manage and monitor click events. A click is recorded (measured) when Google Ads has received an initiated click and sends the user an HTTP 302 redirect to the advertiser landing page or website (or other intermediate server such as an advertiser's agent). These measured click events are recorded to data logs within an event file system. The data log files are then accumulated, edited and compiled through fully automated processes to produce click measurement and advertiser reporting. The editing process includes the process of filtering erroneous or corrupt data, identified non-human traffic including robots and other automated processes, and other identified invalid click activity. The filtered clicks are considered invalid, which means they are not billable to the advertiser. Google prepares click reports for advertisers which can be directly accessed by the respective advertiser.

Click measurements can be reported aggregated by geographical location (not subject to MRC accreditation) and device type. Geographical location is based on the user’s IP address or from a publisher-provided location (publishers must obtain user permission to provide such location). Note that some traffic may be routed through a service provider’s proxy servers and so might not correctly reflect the user’s actual locations (for example, mobile carriers may proxy mobile traffic). Device type classification (computer, tablet, and mobile devices) is based on information from the HTTP header-using libraries operated by Google.

In some AdSense implementations, partners render ads on their own site subject to their own design and formatting rules and control the clickable area around the ad. In these implementations, adjustment of this area is beyond Google's control. In regular AdSense implementations, Google both controls the clickable area as well as renders the ad impression directly to the end user.

Google allows Google Ads users to create video campaigns, upload and manage creatives, and set bidding strategies and related targeting for their campaign. Google Ads video ad content must be hosted on YouTube; however, these video ads can appear on YouTube and on video partner sites and apps across the Google Display Network (GDN).

Google’s proprietary Interactive Media Ads Software Development Kit (IMA SDK) is integrated directly into the YouTube video player, the YouTube mobile app, or video partner sites and apps to facilitate communication between the video players and the ad server for video measurement. Google maintains two versions of the IMA SDK, one which supports Flash and the other which supports HTML5. The IMA SDK is a Video Ad Serving Template (VAST) (versions 2.0, 3.0 or 4.0) with a compliant tag implementation used to measure both linear and non-linear video ad content to serve and track digital video ads. The IMA SDK also supports Video Player Ad-Serving Interface (VPAID) (version 2.0) that allows the video ad and video player to communicate with each other, and Video Multiple Ad Playlist (VMAP) that allows multiple ads to be played within the video ad content.

All measured YouTube video ads included in the video viewability report are delivered in-stream. For video ad impressions, measurement utilizes the count-on-begin-to-render methodology. When properly implemented by the video ad content publisher, the Google Ads IMA SDK solutions are consistent with the Video Impression Guidelines requirements regarding post-buffering initiation of the measurement event. TrueView in-stream ads are often referred to as “skippables” as they have a skip button and give viewers the option to skip the ad after 5 seconds and run in-streams (pre, mid, or post-roll) of a video. TrueView views are a cost per view format, which means we only charge the advertiser when the viewer “views” the ad. TrueView views are not related to viewability. With TrueView in-stream ads, you pay when a viewer watches 30 seconds of your video (or the full duration of the video if it's shorter than 30 seconds) or interacts with your video, whichever comes first. A view is defined for TrueView in-stream ads in the following ways:

Google tries to identify and filter both general and sophisticated invalid traffic continuously through data-based identifiers, activities and patterns. This identification and filtration is done across clicks and video impressions, and includes non-human activity and suspected fraud. However, because user identification and intent cannot always be known or discerned by the publisher, advertiser or their respective agents, it is unlikely that all invalid traffic can be identified and excluded from the reported results proactively. To protect invalid traffic filtration processes from becoming compromised or reverse-engineered, no details of specific filtration procedures, beyond those detailed here, will be disclosed, other than to auditors as part of the audit process.

Both specific identification (including obeying robot instruction files, filtration lists, and publisher test activity) and activity-based filtration methods (including analyzing multiple sequential activities, outlier activity, interaction attributes, and other suspicious activity) are utilized in filtration.

In addition, the following parameters apply to the filtration methodology:

• Third-party filtration is not used by Google.
• Robot instruction files (robots.txt) are employed.
• Sources used for identification of non-human activity: Google uses the IAB/ABCe International Spiders & Robots List as well additional filters based on past robotic activities. The IAB Robots List exclude file is used.
• Activity based filtration processes: Activity-based identification involves conducting certain types of pattern analyses, looking for activity behavior that is likely to identify non-human traffic. Google's Ad Traffic Quality team has systems in place to determine any suspicious activities and does such activity based filtering appropriately.
• All filtration is performed 'after-the-fact' and passively. That is, the user (browser, robot, etc.) is provided with their request without indication their traffic has been flagged, or will otherwise be filtered and removed as Google does not want to provide any indication to the user agent that their activity has triggered any of Google's filtering mechanisms. In some cases frontend blocking is also utilized, when it is likely that the resulting ad request may lead to invalid activity. Historically less than 2% of ad requests are blocked.
• Processes have been implemented to remove self-announced pre-fetch activity.
• Processes are in place to allow publisher test clicks and video impressions. These processes support publishers adding a specific tag to an ad request to indicate that the ad request is a test request and should not be counted for any billing or official accounting purposes.
• When inconsistencies or mistakes are detected, processes exist to correct this data and provide refunds to advertisers. These refunds are reflected in the billing summaries. The corruption of log files is extremely rare, but in cases where this may occur, processes exist to recover them.
• Processes have been implemented to remove activity from Google internal IP addresses.
• Filtration rules and thresholds are monitored continuously. They can be changed manually, and are updated automatically on a regular basis.

All partners that display Google Ads on their content are required to adhere to our program policies, which prohibit invalid activity. Learn more about AdSense Program policies.

Google filters for invalid traffic on an ongoing basis, and will review any business partners that receive high amounts of invalid traffic. Partners who continually receive high amounts of invalid traffic may have their account suspended or closed.

Google Ads reports the total number of clicks, the total number of impressions, and subsets of this data (for example, clicks, impressions, and clickthrough rates, by campaign, ad group, and keyword) to advertisers, and similar data corresponding to site statistics to publishers.. The scope of the audit process covers the click and advertiser reporting for Google Ads. These figures may fluctuate to an extent during the course of the month and are not considered finalized until they are frozen at month end. After this time, the reported clicks will not be adjusted, however, credits may be given to advertisers if Google deems it appropriate.

Google Ads includes the capability for advertisers to see the total number of daily clicks filtered (marked invalid) for each campaign. Google Ads doesn’t report general invalid traffic and sophisticated invalid traffic totals separately to prevent this data from being reverse engineered to optimize invalid traffic. Approximately 80% of total invalid click traffic is estimated to be general invalid traffic.

Comprehensive unit test procedures are utilized to ensure the accuracy of reported data in the Google Ads and AdSense frontends. These are the primary mechanisms utilized to ensure that data from backend databases are conveyed accurately in user-facing reports. In addition, user feedback is carefully monitored to discover and correct any errors which may make it through to a release. Numerous automated systems are in place to ensure the proper operation of all machines and software reporting data to Google Ads users. The content of the reporting, however, is primarily verified through the aforementioned unit tests and user feedback.

Electronic records relating to click activity are retained indefinitely. However, two data fields, IP and cookie IDs, are anonymized after a specified time period (9 months for IP addresses and 18 months for cookie IDs).

You can derive Gross Clicks metrics by reporting on Clicks and Invalid Clicks (GIVT & SIVT) this will allow you to calculate Gross Clicks.

Dashboard metrics presented outside of Search and Display Click campaign totals and device type segmentation are not submitted for MRC accreditation.

Search clicks from Connected TV (CTV), unknown and other device types, which are not MRC accredited, may be commingled with accredited desktop clicks and total clicks metrics on Google Ads reporting dashboards. Traffic from unaccredited device types is estimated to be less than 1% of search click traffic.

Google Ads reports the total number of video viewable impressions, viewability metrics (outlined below), and subsets of this data to advertisers.

For the purposes of MRC accreditation, only the Google Ads metrics listed in download only video viewability report are in scope. Other reporting of video impression and viewability metrics across the front-end tools are excluded from accreditation, for example, the Campaign reporting front-end.

The metrics in the downloadable video viewability report are reported Total Net of GIVT and SIVT across Desktop, Mobile Web, and Mobile In-App environments. GIVT metrics are also reported for clicks. Approximately 79% of total invalid video impressions traffic and 18% of total invalid TrueView views is estimated to be general invalid traffic. Due to the nature of TrueView views ad format implementation, the GIVT percentage will be lower. GIVT metrics for Google Ads clicks and video are not MRC accredited. These metrics have been submitted and are pending accreditation for Desktop, Mobile Web, and Mobile In-App environments.

The following Video ad formats are measured and reported in the video viewability report for Video ads running on YouTube and Google Display Network. All other ad formats, which aren't described below, are excluded from the video viewability report, including ads served in the YouTube Kids mobile application.

• Skippable in-stream ads: Skippable in-stream ads are 5 seconds or longer and play before, during, or after other videos. After 5 seconds, viewers have the option to skip the ad. TrueView views are only available in the Skippable in-stream ads format.
• Bumper ads: Bumper ads are 5-6 seconds long and play before, during, or after other videos. The viewers won’t have an option to skip the ad.
• Non-skippable in-stream ads: The non-skippable in-stream ads can be up to 15 seconds long and play before, during, or after other videos. The viewers won’t have an option to skip the ad.

GVP in-stream ad placements served on mobile app environments are as follows:

• In-stream placements from IMA SDK (e.g. a pre-roll video ad within a Paramount+ stream)
• Rewarded app ads (e.g. a full screen video ad that rewards user with an extra life on a game")
• App interstitials (e.g. a full screen video ad that plays before a level on Candy Crush)

Google uses supervised machine learning techniques¹ through methods such as Classification (e.g., Neural Network approach), in which the model will predict invalid traffic (IVT) by making a yes/no decision about whether an event is invalid, and Logistic Regression, in which the model scores various activities and then an IVT decision is made based on score thresholds. Supervised machine learning models may also use tree methods and graph methods. 

Data sources used for machine learning include logs of queries and interactions (“ads logs”), non-logs data that can be joined with ads logs, and a variety of other supplementary proprietary signals. Google relies on hundreds of data sources of varying sizes: the total number of records per data source ranges from thousands to trillions, depending on the data source. Traffic-based models are required to be evaluated with a minimum 7 days of traffic as input data.

For active defenses Google maintains monitoring procedures over the traffic signals (training data) feeding into the models, which trigger alerts for human intervention if certain threshold bounds are not met. As a result minimal, if any, reduced accuracy is expected. 

Models are continuously retrained when appropriate and practical, and model performance is regularly or continuously assessed. As a result (similar to our monitoring procedures above) minimal, if any, reduced accuracy is expected.

Biases in machine learning training and evaluation data are minimal and if they are material the IVT defense would not be approved. All machine learning projects (“launches”) go through a cross-functional review process before they are approved. As part of this process, bias for the model(s) and corresponding data are evaluated, and projects must meet predetermined ad traffic quality criteria before being approved. Continuous monitoring is in place to detect the emergence of bias in models, which in turn trigger alerts and model evaluation, analysis, and updates.

Google applies a mix of machine learning and/or human intervention/review techniques on all traffic. For some defenses Google relies on ML-based lead generation followed by human review. Other defenses start with human review data and use ML to generalize. Our application of machine learning and human intervention/review techniques is evolving, and our usage shifts according to multiple criteria, including alerts, escalations, and organic fluctuations in types of invalid traffic that may emerge. As a result, the distribution is not in steady state, and the “level” of reliance on either machine learning or human intervention/review fluctuates over time.

¹Supervised machine learning relies on labeled input and output data, meaning that there is an expectation for what the output of a machine learning model will be.