GenAI Best Practices for Creating Audiences

Column Names from Data Sources

GenAI uses the column names and data types from the corresponding data source to generate criteria suggestions. In addition, GenAI has access to a sample of the data values. To ensure that GenAI provides high quality recommendations, columns should be named with terms that are clear and indicate their purpose.

Column names that do not describe the data can prevent GenAI from creating a useful audience. For each column, ensure that the Column Name or the Displayed Column Name is meaningful in relation to the data that is in the column. For example, use “Gender” as a display name for a column that contains gender information, instead of a more generic name such as “Column 1”.

Column names should not conflict with the data type in the column values. For example, avoid naming a column “Country Name” that contains numerical data. This conflicts with the expectation that a country name is a string.

Dates

• Give specific date ranges in the prompt. For example, instead of including the phrase “a recent purchase” in the prompt, use “a purchase within the last 2 months”.
• Confirm that any use of dates in the generated audience is correct. For example, if the prompt includes “people who clicked the email link 20 days ago”, ensure that the generated audience criteria correctly account for the date range.

Reduce Ambiguity

To increase the likelihood of useful results, eliminate ambiguity by being as specific as possible in your prompt. Here are some examples that can introduce ambiguity into the generated AI output:

• Using categories that represent a group of people or a set of common characteristics.
• Expecting GenAI to assign a value to numerical data.

One way to determine if your prompt has ambiguity in it is to consider whether it can be interpreted in more than one way. For example, the phrase “college students” can be interpreted in these ways:

• people enrolled in a college that are between the ages of 18 and 23
• people enrolled in graduate school
• people enrolled in a college part-time
• people enrolled in a college full-time

Instead of using “college students” in your prompt, opt for a phrase that more explicitly matches what you are looking for. For example, you might use “people enrolled in graduate school” instead of “college students” if that distinction matters.

If the name or display name of a column refers to a categorical word, then it can be used successfully in your prompt even if the word contains ambiguity. The context that the column gives can offset the ambiguity inherit in the name of the category. For example, if you have a column that is named “high points”, then using the phrase “high points” in your prompt can give you good results.

Here are some examples of prompts that contain ambiguity and suggestions for how to improve them:

• Instead of including the phrase “good credit score” in the prompt, use “credit scores greater than 670”. By giving a numerical range, you are defining the credit score that you consider to be a good one, instead of relying on GenAI to determine that.
• Instead of including the phrase “young professionals” in the prompt, use “professionals between the ages of 20 and 35”. By specifying the age range, you are not relying on GenAI to define which set of ages the young professional category includes.
• Instead of including the phrase “customers who reached out to support multiple times”, use “customers who reached out to support more than one time”.
• If you have a column that indicates the probability that someone will convert, use “a person with a probability of converting that is .5 or greater” instead of the phrase “more likely than not to convert” in the prompt. By giving a numerical value in the prompt, you are not relying on GenAI to evaluate what “likely” means.

Attribute Values

Troubleshooting AI-Generated Output

The output generated by AI is limited to a certain number of characters. If you receive an error message when trying to generate an audience, try one of these solutions:

• Add a sentence to your prompt that limits the returned attributes to those only in the identity column.
• Be explicit about which attributes you want. For example, add "I only want identity and age" to the prompt.

The content in the prompt does not clear when the audience is generated. As a result, you can continue to refine the prompt without losing what is already entered. However, the criteria listed in the audience are repopulated every time that you generate an audience.

Examples of Good Prompts

A good prompt contains a clear description of the data subset that you want to use. It also lists the information that you want to see in the generated audience.

Here are some examples of good prompts:

• "Give me all of the people between the ages of 25 - 30 and give me their email address and the city they live in"
• "Give me all of the people who made a purchase on January 20, 2024. Include their age and whether they own a second residence"