What is Instant BQML?
Instant BQML is a simple form you fill out with details about your Google Analytics and Google Cloud project that quickly & easily assists you in creating an automated, configurable, customized first-party machine learning propensity pipeline for use in marketing activation and optimization.
What does BQML stand for?
BQML stands for BigQuery Machine Learning (ML). It allows users to create and run ML models using standard SQL queries directly within BigQuery.
How does Instant BQML work?
Instant BQML generates JSON instructions which describe a machine learning model and necessary tasks. These are uploaded to CRMint, which orchestrates the analysis in BigQuery and integrates the data with Google Analytics - this data is then used to create audience segments and optimize bids in Google Ads.
What data does Instant BQML analyze?
By default, Instant BQML is configured to analyze the Google Analytics BigQuery export.
Can Instant BQML be used with a Universal Analytics property?
Yes, Instant BQML is compatible with Universal Analytics 360 properties and Google Analytics 4 properties but not Universal Analytics free properties.
What marketing objectives can I optimize using Instant BQML?
Instant BQML offers a range of optimization objectives including Purchase Propensity, Repeat Purchase Propensity, Page Propensity, Event Propensity, Product Propensity, User Property Propensity and Churn Propensity.
What is the difference between propensity modeling and lifetime value modeling?
Propensity models predict the likelihood of future actions from potential customers using Google Analytics data while lifetime value models estimate the total revenue expected from existing customers using CRM data.
Do I need to customize Google Analytics to deploy Instant BQML?
No. Instant BQML uses the default Google Analytics BigQuery export schema and default Google Analytics measurement, however, you have the option to enhance your model with custom measurements to achieve the best results.
What are the default features used in the Instant BQML models and can I add more?
Default features include device information, geographic information and user engagement information. You can also customize the model with additional features!
Are there any best practices for data preprocessing or feature engineering to ensure optimal model performance with Instant BQML?
Yes, you can improve model performance by ensuring data quality, removing outliers and creating new features by combining existing ones - you should also consider adding additional data sources!
What are the prerequisites for setting up Instant BQML?
To deploy Instant BQML, activate the BigQuery export in Google Analytics, link Google Analytics with Google Ads and select an event for optimization. Lastly, decide on deploying CRMint within an existing Google Cloud project or setting up a new project.
Is there an overview or slide deck available to help me understand Instant BQML more?
Yes! You can find the Overview on the Resources page.
Is there a checklist available to help me prepare for Instant BQML setup?
Yes! You can find the checklist on the Resources page.
Is there a 1 sheeter, 2 pager, or brief summary document to help me understand Instant BQML more?
Yes! You can find a 1-sheeter on the Resources page.
What permissions are needed to deploy Instant BQML?
Instant BQML requires the Google Cloud project owner and a Google Analytics Administrator for successful deployment.
Who should be involved in the implementation of Instant BQML?
Consider including people with marketing, Google Analytics Admin and Google Cloud Project Owner expertise to define campaign objectives, provide access and permissions and interpret modelling results.
Why does the CRMint service account need Editor permissions in Google Analytics?
The CRMint service account must be added to the Google Analytics property as an Editor in order to automatically publish audiences.
Why does the CRMint service account need BigQuery Data Viewer permissions?
If you are deploying CRMint in a separate Google Cloud project, you will need to grant BigQuery Data Viewer permissions to the CRMint service account for the GA4 BigQuery Export dataset.
What is CRMint and how does it relate to Instant BQML?
CRMint is an open-source data flow platform that executes the machine learning pipelines generated by Instant BQML - think of Instant BQML as the architect that designs the plans and CRMint as the builder that brings them to life!
CRMint can be deployed using either App Engine or Cloud Run, which version should I choose?
You should use App Engine to deploy CRMint unless you have a Google Cloud Organization set up in your GCP which is required for the Cloud Run version. If you are unsure what a Google Cloud Organization is, or haven't set one up yet, you should deploy CRMint using App Engine.
What is the command to deploy the CRMint application?
The command to deploy CRMint is provided in the Instant BQML form and can be copied and run directly in Cloud Shell.
Do I need to have a Google Cloud project already set up to deploy Instant BQML?
Not necessarily, you can deploy the CRMint application in an existing Google Cloud project or set up a new project specifically for CRMint.
Is CRMint only for use with Instant BQML?
No, while CRMint is used to automate the Instant BQML solution, it can also be used to orchestrate other data pipelines.
Can I export my CRMint pipelines to share with colleagues or use in other projects?
Yes, you can easily export your pipelines as JSON files directly from the CRMint UI.
Can CRMint handle data processing tasks beyond machine learning, such as data cleaning and transformation?
Yes, CRMint can handle various data processing tasks by executing custom SQL scripts within BigQuery.
Can CRMint be used to orchestrate data pipelines for other Google Cloud products, such as Google Cloud Storage or Cloud Spanner?
Yes, while Instant BQML is focused on BigQuery, CRMint can orchestrate jobs that interact with other Google Cloud Products
Can I manually run specific jobs within a pipeline or does the whole pipeline need to be executed?
You can run individual jobs within a pipeline as needed by clicking the "run job" button in the CRMint UI.
Can I edit my pipeline in CRMint after I've enabled the "Run on Schedule" toggle?
Yes, but you will first need to disable the schedule to make any changes to the underlying BigQuery scripts within the pipeline, this ensures that any manual edits are not overwritten the next time the pipeline runs automatically.
How can I troubleshoot issues with my CRMint pipeline, are there any debugging tools available?
Yes - CRMint offers logging for every job within your pipeline and you can view the status of each job (running, success, failed) in the CRMint UI. You can also enable "dry run" mode for BigQuery jobs which allows you to test your queries without incurring costs.
My prediction pipeline failed - where can I view the logs to see what went wrong?
You can find detailed logs for each job within your pipeline by navigating to the "Logs" tab in the CRMint UI. These logs will provide insight into any errors that may have occured during the pipeline run.
Does Instant BQML create a new user property in my Google Analytics 4 property for every pipeline I create?
Yes, the prediction pipeline will create a new user property with a unique name based on the BigQuery namespace you provide in the Instant BQML form.
Do I need to create a user property in Google Analytics before running my Instant BQML pipeline?
No, a job within the prediction pipeline will automatically create a new user property in your GA4 property to store the propensity scores.
What scope is used for the custom dimension created by Instant BQML?
The custom dimension (user property) created by Instant BQML is user-scoped, allowing the propensity scores to be associated with individual users and persist over time.
What is measurement protocol and how is it used by the Instant BQML prediction pipeline?
Measurement protocol is a method for sending data directly to Google Analytics servers. The Instant BQML prediction pipeline uses measurement protocol to send the propensity scores it generates to your GA4 property, allowing you to leverage the data for marketing activations.
What does the "Publish/Update GA4 Audience" job do?
This job uses the Google Analytics 4 Audiences API to create and publish new audience segments in your GA4 property based on the boundaries defined in the "Audience Boundaries" job. By default, the job creates three audience segments. A high audience segment consisting of your top 25% of users, a medium audience segment composed of your next 50% of users, and a low audience segment consisting of your bottom 25% of users.
Does the prediction pipeline create the predictive audiences in Google Analytics?
Yes, a job within the prediction pipeline uses the Google Analytics 4 Audiences API to automatically create and publish your predictive audience segments.
Why does the Event Propensity pipeline have additional jobs compared to the other pipelines?
The additional jobs, "Conversion Values" and "User Value Map" are specific to the Event Propensity objective and are used to calculate a value for each user based on the predicted conversion rates of similar users. This scaled value is sent back to Google Analytics as an event parameter that can be used to optimize Smart Bidding in Google ads.
Why does Instant BQML calculate conversion values for my audience segments?
Instant BQML calculates the conversion rate of each audience segment to your selected marketing objective in order to understand the relative performance of each segment.
What is the significance of the VBB_MULTIPLIER parameter and how does it impact the scaled value assigned to users?
The VBB_MULTIPLIER is used to scale the conversion values to a range suitable for bidding in Google Ads, you can adjust this parameter based on your business needs.
Does the scaled value created by the User Value Map job overwrite the value of my existing conversion actions?
No, the scaled value is sent to Google Analytics as a new event parameter and does not overwrite the values of any existing conversion actions or goals.
Why is a moving average used to calculate the scaled value rather than the actual daily conversion rate?
A moving average is used to account for daily fluctuations in conversion rates and provide a more stable value that reflects longer-term trends, this ensures more stable bidding behaviour.
Can I use my Instant BQML audiences with Smart Bidding strategies other than Target ROAS?
Yes, the predictive audiences can be used with any Google Ads bid strategy including Maximize Clicks or Maximize Conversions.
How can I use the prediction pipeline outputs to optimize bidding in Google Ads?
If you've selected Event Propensity as your marketing objective, the prediction pipeline will calculate conversion values that can be used for value-based bidding. Create a new conversion event in GA4 based on the event that was sent during measurement protocol so that you can import the Conversion Event into Google Ads to optimize your bidding strategies.
What types of marketing activations are possible in Google Ads using Instant BQML?
You can bid on predictive conversion events, target the high, medium, and low propensity audience segments with unique bidding strategies and creatives and tailor conversion values for specific audience segments.
Does Instant BQML offer support for Google Ads value rules?
Yes! you can use value rules to adjust the value of your conversions based on your Instant BQML audience segments.
Can Instant BQML be used with Google Ads automated rules to optimize campaigns based on propensity scores?
Yes, while the audiences are the recommended approach for optimization you can also create automated rules that are based on your custom dimension.
How can I evaluate the performance of my Instant BQML model?
You can evaluate the performance of your model both in BigQuery ML and by analyzing the performance of your predictive audiences and conversion events in Google Analytics 4. Furthermore, you can use the Instant BQML Performance Insights Looker Studio template to visualize your model's predictive performance.
Can I use data from Google Analytics and Google Ads to inform each other in a feedback loop?
Yes! this is actually a core benefit of Instant BQML. For example, you can import the predictive conversion event created by your pipeline into Google ads and bid to it with a value-based bid strategy - the results can then be analyzed in GA4 to evaluate performance and inform future iterations of the model.
Can you provide examples of how businesses have benefited from using Instant BQML?
Webautomation saw a 4.5x increase in conversions and a 70% reduction in CPA, Worten achieved a 3x ROAS increase and a grocery store chain experienced a 200% conversion rate lift and a 75% reduction in cost-per-conversion.
Where can I find the Cloud Project ID for my GA4 BigQuery Export?
You can find this in Google Analytics under Admin > Product Links > BigQuery Links
Where can I find my Google Analytics BigQuery Export Dataset Location?
You can find this in Google Analytics under Admin > Product Links > BigQuery Links > Completed link details > Default Location for dataset creation.
Which Google Analytics BigQuery Export type should I select?
You should select the option that matches the BigQuery Export configuration you have selected in your GA4 property
What if my Google Analytics 4 BigQuery Export is only scheduled daily?
The Instant BQML form will tailor your pipeline’s SQL queries to ensure they are compatible with your BigQuery Export setup and will pull from the correct tables.
Where do I find my Google Analytics Data Stream information?
You can find your data stream information in Google Analytics under Admin > Data Collection and Modification > Data Streams
Where can I find the Google Analytics API secret?
You can find your API Secret in Google Analytics under Admin > Data collection and modification > Data streams
The value in the API_SECRET parameter is different from the one in my Google Analytics 4 property - which one should I use?
You should use the API secret value generated by Google Analytics and shown in your GA4 property settings.
Where can I find the Google Analytics Property ID for my GA4 property?
You can find this in Google Analytics under Admin > Property > Property Details
Where can I find the Cloud project ID for CRMint?
This is the Project ID where you've deployed CRMint and can be found in the Google Cloud console under IAM & Admin > Settings.
What should I name my pipeline?
You can choose any name, but it must start with a letter, be 24 characters or less and only contain letters, numbers or underscores - the form will suggest a name for you.
Are there any costs associated with using Instant BQML?
While the tool itself is free, there are costs for the Google Cloud services it uses, like BigQuery ML and data storage. CRMint offers a dry-run feature to estimate costs before execution.
I see the prediction pipeline has a job called "Clean Up Predictions", what does this job do?
This job automatically cleans up old prediction tables in BigQuery in order to manage storage costs and prevent clutter.
If I have multiple Google Analytics 4 properties sending data to the same BigQuery project, can I use Instant BQML to analyze them separately?
Yes you can! simply fill out a new Instant BQML form for each GA4 property.
What is the benefit of enabling Global Explain in the BQML model options?
Enabling Global Explain provides feature attribution reports directly in Vertex AI which can help you better understand how different features are influencing the model & predictions
What does the "OPTIONS" section of the BQML model do?
The options section is where you define the type of model you want to create and set specific parameters to configure the training process.
The model references a "target_label" - what is this?
The target label is the column that contains the values your model is trying to predict, it is typically a binary 1 or 0 that represents whether or not the user completed the desired action (ie made a purchase).
Can I use Instant BQML to create predictive audiences based on custom events that I have set up in Google Analytics?
Yes, as long as data for the custom events is included in your GA4 BigQuery Export they can be used to create predictive audiences.
What does the Format Data Job do?
The Format Data job in the Vertex AI training pipeline formats the raw Google Analytics 4 BigQuery export data into a table that will be used to create a Vertex AI Dataset.
What does the Create Vertex AI Dataset job do?
This job creates a new dataset in Vertex AI based on the structured data created in the "Format Data" job. The dataset will then be split into training and testing data and used to train the ML model in the next step.
What does the Vertex AI Trainer job do?
This job uses Vertex AI's tabular training functionality to train a new machine learning model based on the data formatted in the "Format Data" job.
The Vertex AI pipeline includes a "Format Predict" job, how is this different from the "Predict" job in the BQML pipeline?
The Format Predict job in the Vertex AI pipeline performs a similar function to the predict job in the BQML pipeline, however, rather than using BigQuery ML to generate predictions, it formats the data required by the Vertex AI Batch Prediction job.
What does the Vertex AI Batch Prediction Job do?
This job uses the model created by the Vertex AI trainer job to generate a batch of predictions. The predictions are then written to a table in BigQuery.
What is the long term vision for Instant BQML?
We plan to enable business users to leverage Google Cloud's AI capabilities through a simple interface, with plans to expand to additional use cases and Google products!
What is the role of the Google Analytics Data Stream selection in the Instant BQML form?
The data stream selection specifies which type of data will be used to train the model (ie web or app) and provides the necessary information for the Measurement Protocol API.
What does the Instant BQML training pipeline actually do?
The training pipeline is responsible for creating your machine learning model within BigQuery, it creates a dataset to store outputs and uses a BQML script to create the predictive model based on your GA4 BQ export data.
What does the "Create BQ Dataset" job in the training pipeline do?
This job creates a new dataset in BigQuery to store the outputs of your machine learning pipeline if it doesn't already exist. It uses the BQ_DATASET parameter value you provided in the Instant BQML form as the dataset name which by default is "analytics_" followed by your GA4 property ID.
What does the Event Propensity Training job do?
This job uses BigQuery ML to create a predictive model based on your Google Analytics 4 data and the marketing objective you selected.
What BQML model type is used by Instant BQML by default and can I change this?
The default model type is a Boosted Tree Regressor, but you can modify the SQL query in your training pipeline to create other models such as a Logistic Regression or AutoML Regressor.
What is the significance of the _TABLE_SUFFIX filter in the BQML queries?
The table suffix filter limits the data used in the model to the last 12 months, this ensures the model is trained on fresh, relevant data.
What is the recommended cadence for running the Instant BQML training and prediction pipelines?
The training pipeline should ideally run weekly to capture the latest customer behavior trends while the prediction pipeline should run hourly to provide fresh scores.
Can I automate my Instant BQML pipelines?
Yes, and automating them is highly recommended! Once a pipeline has been run successfully in CRMint you can automate it by toggling the "Run on schedule" switch.
What does the "schedules" section do in the CRMint pipeline JSON?
The schedules section defines the automated schedule that CRMint uses to run your pipelines.
What does the Instant BQML prediction pipeline do?
The prediction pipeline uses your trained BQML model to generate propensity scores for users and ingests them into Google Analytics 4. It also creates audience segments based on these scores and a new conversion event for specific use cases.
How often should the prediction pipeline run?
The prediction pipeline should ideally run hourly. This ensures the scores in your GA4 property are up to date and reflect the latest user interactions.
What happens when the prediction pipeline runs?
When the prediction pipeline runs, it completes several jobs including generating predictions with your BQML model, formatting the data and sending the scores to GA4 via the measurement protocol, calculating audience boundaries, creating and publishing predictive audiences, and creating a new conversion event in GA4 (for certain objectives.)
I see the prediction pipeline references a "calculated_visitors" table - what is this used for?
The calculated visitors table is created by the prediction pipeline and contains a list of users who visited your website or app on the date specified in the "calculated_fields" table, this ensures that propensity scores are only generated and sent to Google Analytics 4 for relevant users.
What is the purpose of the "calculated_timestamps" table created by the prediction pipeline?
This table is used to assign a timestamp to the measurement protocol events that are sent to Google Analytics 4. It selects the latest event timestamp for each user from your GA4 BigQuery Export and adds 1 - ensuring the new event is processed as a 'fresh' event by Google Analytics.
How does the User Value Map job use the data from the Conversion Values job to assign values to users?
The User Value Map job joins the predicted value of each user from the "Predict" job with the scaled value from the "Conversion Values" job to assign each user a value based on the conversion rate of the segment they are in.
How does the Instant BQML prediction pipeline determine which users to send scores for?
The prediction pipeline identifies users who visited your website or app the previous day by querying the GA4 BigQuery Export and only sends scores for these users to ensure the data in GA4 is fresh.
What is the purpose of the "Format Scores" Job?
The format scores job prepares your data to be sent back into Google Analytics 4 using the Measurement protocol. it creates a table with the parameter name, user ID and values formatted as expected by Measurement Protocol.
How do I access the predictive scores generated by Instant BQML?
The scores are sent to your linked Google Analytics 4 property via the Measurement Protocol and will appear in your realtime report as they are created. A user-scoped custom dimension is also created to store the scores for audience creation!
Can the Instant BQML prediction pipeline analyze data from Firebase?
Yes! if you have a Firebase app linked to your Google Analytics 4 property and are collecting data via a Firebase data stream you can use Instant BQML to analyze this data.
What is Vertex AI and why is it used in the Instant BQML pipeline?
Vertex AI is Google Cloud's platform for managing and deploying machine learning models. Instant BQML leverages Vertex AI to train models which enables advanced ML capabilities like Explainable AI.
What if I need help deploying Instant BQML?
You can reach out to the Instant BQML User Group (instant-bqml-users@googlegroups.com) or our Instant BQML Certified Group (instant-bqml-certified@googlegroups.com) for support. If you have a Google Cloud account contact, please reach out to them for more guidance.
How can I certify my organization to appear on the Partners page?
Please complete our Certification program. You can find more details on our Resources page.
How much data do I need for Instant BQML?
To effectively use Instant BQML, begin by linking Google Analytics 4 (GA4) with BigQuery, selecting an objective with at least two outcomes (e.g., purchaser vs. non-purchaser). The performance of Instant BQML enhances with more data, leading to better outcomes. After initializing your Instant BQML pipeline, audience segments will begin to improve, reaching full potential in 30 days, though they are usable immediately and refine over time with new data.
How long does it take for audience segments to reach full size?
Instant BQML pipelines with Google Analytics create audience segments that achieve full size within 30 days, due to their default 30-day membership duration. User propensity scores are updated daily and integrated into Google Analytics, ensuring segments reach their full potential in at least 30 days.
In BigQuery, what tables are generated for Instant BQML analysis?
Instant BQML creates some intermediary tables to assist with analysis. Those include: predictions, calculated visitors, calculated timestamps, calculated fields, conversion values, user value map, and a measurement protocol formatted table. Note that the conversion values and user value map are only generated for event propensity pipelines.
What does the predictions table in BigQuery contain?
The predictions table (named like: %pipeline_name%_predictions_%platform%) contains user IDs mapped to the predicted conversion probability according to the model.
What does the calculated_visitors table in BigQuery contain?
The calculated_visitors table (named like: %pipeline_name%_calculated_visitors) contains a mapping of user IDs who visited the property on the previous day and will be scored & integrated in the next Prediction pipeline run.
What does the calculated_timestamps table in BigQuery contain?
The calculated_timestamps table (named like: %pipeline_name%_calculated_timestamps) contains a mapping of the previous day's user IDs and their last known event timestamp according to the BigQuery export. This timestamp is eventually used in the measurement protocol payload (plus a variable number of microseconds based on user properties) to add or update the propensity score for the user.
What is the performance_insights table in BigQuery?
This table provides a high-level analysis of your model's predictions by displaying the conversion rate for propensity scores, categorized as High, Medium, or Low. You can also visualize this table in the Instant BQML Performance Insights Looker Studio report.
What is the scored_users_log table in BigQuery?
The scored_users_log table provides a history of the scores assigned to each user by the prediction pipeline. It includes the user ID, timestamp for measurement protocol event, user property & event score assigned, and a unix microsecond timestamp of the system time of the last time the user was scored.
How do I access the CRMint UI?
Accessing the CRMint UI depends based on the version of CRMint you are using. If you have deployed the App Engine version of CRMint, the simplest way would be to go into Google Cloud, search for App Engine, and click on the URL in the upper right hand corner on the Dashboard page (this URL follows a general pattern of your_Google_Cloud_Project_ID.appspot.com. If you are using the Cloud Run version of CRMint, you can run `crmint cloud url` from your Google Cloud Shell once the CRMint command-line interface is installed.