Are Synthetic Personas the New Normal of User Research?

Artificial intelligence has transformed every aspect of marketing – from content creation to image generation to consumer research. We’ve now reached a point where AI models are using AI-generated data to train themselves. Why? Because they’ve run out of human data; machines aren’t learning from people but from other machines. Of course, AI assistants are not yet perfect but they are getting better every day.

So, it’s not a surprise that large language models can replace real-world people in surveys, interviews, and research studies. Market researchers are already using generative AI tools like ChatGPT to create synthetic personas that can mimic, describe, and (sometimes) successfully predict human behavior. The goal is simple: reduce the cost of marketing while improving user experiences.

Synthetic personas do away with traditional consumer research, which marketers believe is critical to understanding customers and selling their products. However, these methods are expensive and time-consuming. More importantly, they require big budgets and human resources – people who are willing to participate in a survey and answer your questions. With marketing costs on the rise and people becoming sensitive about the kind of data they share, market researchers often find it difficult to collect quality customer information.

It is this bottleneck – collecting, processing, and analyzing consumer demographic and behavioral data – that synthetic personas try to solve. Although they cannot replace qualitative and quantitative research involving human interactions, they can supplement your research methods by helping you get a general idea of the reaction your products or marketing campaigns will receive and make informed decisions.

What Is a Synthetic Persona?

A synthetic persona is a virtual customer created using AI and machine learning algorithms based on user demographics, psychographics, and behavioral data. It provides an accurate representation of a specific audience segment, describing their likes, dislikes, thoughts, preferences, needs, and challenges. Though synthetic users seem similar to customer personas, they are different in the sense that you can directly interact with or talk to them.

They are no different from AI humans created by companies using model generation and sophisticated puppeteering. Such avatars can track eye movements, facial expressions, and pupil dilation to understand human emotions and offer helpful instructions (or in some cases, sell their products or services).

Depending on the goal of your research study, you can create a diverse set of synthetic personas using ChatGPT or dedicated persona generation platforms like Delve AI. They are fairly easy to create and are cheaper than recruiting and conducting research with real people. You have complete control over the data fed into the system and can ask follow-up questions to take the interview forward.

Note: ChatGPT takes a lot of prompting to understand the right customer segment and adopt the related user persona.

For synthetic users to work, you need to ensure that there are standards in place. Your data needs to be validated before it’s fed into the AI systems. The resulting personas also need to be cross-checked and compared with real use cases before you use them.

How synthetic customers work

A synthetic persona is a methodology wherein you feed a selected amount of customer data into large language models (LLMs) like GPT-3 and prompt it to adopt the “persona” of an actual customer based on the data you provided. This data mostly contains demographic information like age, gender, and location, along with psychographic attributes, such as personality traits, attitudes, lifestyles, and interests. The persona generated is then used in customer surveys to give out human-like responses.

Now, the accuracy of the synthetic personas is directly proportionate to the quality of the training data. The higher the quality, the better your personas. If your input data is generic and inaccurate, the output will also be the same.

Data concerns aside, you can use synthetic users to simulate a diverse range of customers and ask them about their needs, pain points, and challenges regarding your product or service. You can use them in qualitative research to get user insights, build product designs, and test ad creatives across multiple target audiences. The great thing about synthetic personas is that you can interview respondents from the comfort of your home or office.

Synthetic data and user research

In the past, the telephone was an essential tool for conducting surveys, interviews, and business conversations. Phone surveys became a great way to gather insights from people. With the advancement of technology and the arrival of the Internet, you could then conduct online surveys.

Fast-forward to the 21st century, and we have synthetic consumer data generated by AI and ML algorithms. You know that LLMs digest a huge amount of training data from books, academic journals, newspapers, and more to synthesize human-like responses in a text-based format. They can summarize text, answer questions, and even write poems. LLMs are continuously learning entities—they learn from their mistakes.

This data processing and learning ability lets them analyze consumer data, build synthetic users, and help out in user research by generating responses. Although synthetic data is derived from real-world events, it’s not the same as the data you collect from human survey respondents. It’s gaining popularity with companies nowadays since they can test scenarios and conduct interviews without worrying about data privacy laws.

Cost and time also play a big part. As explained before, quantitative and qualitative research have high costs associated with them – you need to process data, gather respondents, set a time, and conduct interviews to discover trends and patterns in consumer behavior. Synthetic data easily helps researchers develop new workflows and simulate several scenarios.

AI-powered virtual assistants can carry out user research by interviewing people on your behalf – sometimes in multiple languages, 24/7. They can analyze consumer responses to ask personalized follow-up questions. Synthetic personas additionally help quant researchers study unreachable demographics and fill in missing data and surveys using advanced predictive capabilities. Niche audience segments can be analyzed, with no additional costs involved.

Creating Synthetic Users with Delve AI: Plus Other ChatGPT Prompts

A good way to understand customers is to listen to them – build a rough customer profile and try to empathize with them. But how do these three processes work? You get human feedback to listen, create a buyer persona to characterize, and develop a customer journey map to empathize. The fourth step, interaction, works with chat personas.

Persona by Delve AI is a persona-based marketing software that helps you create personas for your website, competitors, and social media audiences. It covers all four aspects mentioned before – consumer data, personas, user journeys, and interactive user chatbots.

Here’s how: our software gathers your first-party (web analytics, search console, and CRM data) and second-party (social analytics and competitor intelligence) data, then enriches it with 40+ public data sources (eg. Voice of Customer data from reviews, ratings, feedback, online forums, and newspapers) to create data-driven personas for you.

Three to seven personas are generated depending on your business, with B2B persona segments highlighted in green and B2C personas highlighted in blue.

Each persona gives you persona details, distribution, and sample user journeys to ascertain the traits, characteristics, and behaviors of your desirable and non-desirable audience segments. For example, an ecommerce business will able to identify bouncers, passive browsers, cart abandoners, and high-value buyer segments marked with blue and red respectively.

PERSONA DETAILS offer the age, gender, location, job profile, goals, motivations, pain points, values, interests, jobs to be done, and hobbies of that particular segment. Additionally, you get insights into their favorite social networks, brands, ecommerce sites, tools, TV shows, movies, YouTube channels, podcasts, events, and more.

The DISTRIBUTION TAB groups the segment by age, gender, location, language, marketing channels, online activity timings, actions, and decision-making stage. SAMPLE JOURNEYS display the total number of visitors, along with their sessions, page views, conversions, and decision-making stage the user is in. Best thing about it? Your AI-generated personas and user journeys are automatically updated with real-time customer data.

Now, let’s get into our soon-to-be-launched chat with personas feature. It offers users the ability to interact with high-value customer segments online. The dashboard is similar to WhatsApp or any other messaging app. Have doubts related to your new product or marketing copy? Just visit our platform, select your target persona, and pop the question. You can also directly ask a question via Slack – all you have to do is integrate your Slack workspace with Delve AI.

ChatGPT prompts to create synthetic personas

ChatGPT can process vast amounts of unstructured data – numerical data, job profiles, literature, research papers, forums, reviews, and websites – and present it logically. As such, it is a free alternative to create interactive personas that you can use in consumer research.

It’s a simple exercise really:

• Upload your customer data into ChatGPT
• Direct it to list down the possible segments
• Select the audience segment you want to target
• Prompt ChatGPT to adopt the persona of that segment

You can then ask the persona to be a participant in a survey session. Note that ChatGPT personas are only useful when you have broad customer profiles and not a niche audience. Unless, of course, you have enough qualitative and quantitative data for ChatGPT to form assumptions.

Here’s how your prompts can play out:

Prompt 1: Analyze and structure the customer data provided, focusing on key demographic, behavioral, and psychographic attributes. Summarize and highlight any clear trends or patterns across different customer attributes.

Prompt 2: Using this structured data, list the possible customer segments. Name each segment and briefly describe their defining characteristics, including their interests, purchasing behaviors, and other notable demographic details.

Prompt 3: From these segments, focus on [Segment Name]. Give more details about their major motivations, typical pain points, and potential brand interactions.

Prompt 4: Adopt the persona of a typical customer from the [Segment Name] audience. Describe their day-to-day lifestyle, interests, key values, and specific needs. Assume this persona’s character for further interaction.

ChatGPT cannot understand meaning so it cannot be used to create patterns of meaning that synthetic users communicate to you in a study. It does not help you understand the different approaches people take to reach a goal or complete a task. AI cannot use products if you think about it – it can just imagine. You can run tests to validate your product ideas but every opinion might just be a positive one – ChatGPT personas aim to please the interviewer.

Emporia Research conducted a comparative study between three groups of survey responses: LinkedIn-verified respondents, synthetic users generated by AI based on target LinkedIn profile data, and users based on AI-generated personas representing IT decision-makers. Their study discovered that B2B synthetic users generated by AI display a strong positive bias in comparison to real respondents. They follow a herd mentality; the quality of insights is not that great either.

Synthetic Personas in Action: Use Cases and Examples

Marko Sarstedt exemplifies the use of synthetic personas in research and design in his research on silicon samples. He used them to get feedback on survey questions and refine packaging design. Besides his applications, synthetic personas can also be used for other marketing and product-related use cases.

Marketing, segmentation, and branding

You can create two types of synthetic personas: those representing individual customers and those representing groups of customers. Each enables you to get information about user behavior and preferences without accessing private customer data. It’s a good starting point for professionals who are not familiar with their target audience. You can develop different customer segments to test your campaigns and identify which ones respond best. Consequently, you can customize your marketing assets, content, and copy to suit the preferences of your target audience segments.

Simply upload your marketing plans, graphic designs, videos, and ads to get feedback in the campaign development process. After uploading your material into the model, ask the synthetic user for inputs, like:

• Does this video ad invoke happiness or sadness (replace it with your desired emotion)?
• Where would you like to see this ad? On billboards, Facebook, Instagram, or LinkedIn?
• Do you like the overall message that the ad copy tries to convey?
• How likely are you to respond to this email campaign?

Furthermore, add performance data of past marketing campaigns to check what worked and what didn’t – it gives the AI model more context to work with. You will have to tweak your prompts to get efficient outputs, but the results are well worth it. Who knows, it might just help you create an ad that goes viral!

In product development, design, and testing

Synthetic users allow product teams to empathize with their users and create new products, features, and functionalities that align with consumer needs and expectations. You can simulate usage scenarios to identify potential problems and ensure that your product functions well in different situations.

Companies can use them to validate their product designs before launching them in public. For example, they can test how well the user interface and interactions work in a relatively risk-free environment (since they won’t have to test it on actual customers). Doing so will help them make refinements to the overall product user experience (UX), greatly reducing the likelihood of issues popping up when the product is officially launched, thus bringing in new users and reducing customer churn rates.

Question: If you can’t conduct research, is it okay to use synthetic personas? Maybe, maybe not. You can always learn something new, as opposed to not learning anything at all. Yet, basing critical product or marketing decisions on synthetic data is not ideal.

Suppose your new feature is a hit with your synthetic customers. You launch it in the market but it fails with the actual users. What happens then? You waste resources and lose money that was better spent elsewhere. So, always take their responses with a grain of salt.

The Challenges with Synthetic Users

No matter how fancy AI gets there are way too many limitations of synthetic users than meets the eye. Synthetic personas are built using machine learning models, so they are very useful in scenarios where actual customer data is scarce or limited. However, there are some things you need to know before incorporating them into your workflow.

To begin with, synthetic personas have a tough time capturing the full range of human behavior and actions; their responses can be too simplified, generic, or logical. Human beings are not always logical – we are influenced by various internal and external factors (like our personal experiences, personalities, opinions, and values) that govern the way we function and respond to things.

Artificial intelligence and machine learning systems, the foundations of synthetic personas, are not yet able to comprehend these factors. For instance, AI respondents always want to please the interviewer, fully complete any given task, and answer questions in the affirmative (a phenomenon called sycophancy). They provide one-dimensional responses, and seem to care about everything; providing a long list of needs, challenges, and interests. People don’t have such lists – they care about specific things, and other things don’t even cross their minds.

In their paper titled, “A Manager and an AI Walk into a Bar: Does ChatGPT Make Biased Decisions Like We Do?” Chen et al. found that ChatGPT is often unable to replicate effects that characterize customer behavior, like the sunk cost fallacy. People continue bidding on an item even after the price has gone beyond what it's worth, just because they want to win the bid. There is further no hierarchy in their responses – how do you understand what’s important and what’s not?

Another concern is the data used to create synthetic personas; if the training data is biased, the responses will contain and perpetuate stereotypes. Data scarcity also makes it difficult to represent underrepresented or marginalized users. Validation is tricky too. It's not easy to test synthetic users against real-world situations and outcomes. They need to be constantly monitored, updated, and verified with real user data. We are talking about data that’s anonymized and does not reveal sensitive customer information.

Even though you have all these things crossed off your checklist, integrating synthetic users into your existing systems can prove to be difficult. Technical and data problems aside, convincing skeptical stakeholders about their usefulness might be your biggest challenge. Worse, your company might get accustomed to using synthetic customers for user research and not want to invest in the real thing.

The Future of Synthetic User Personas

Synthetic personas reveal unknown insights about your customers. But can we trust synthetic data? Is it reliable, is it ethical? Who knows. Some say they are the future of user research, others say they lack the nuance and variation generally found in human respondents. Yet, in a world where market research costs an arm and a leg, synthetic personas are a cost-effective way to understand your customers.

You can learn about a new user group, what interests them, and the way they behave online, and develop proto-personas and user journeys that can help you with real users. Like every other invention, synthetic user personas come with a list of pros and cons, but if done well, the former far outweighs the latter.