Good campaigns are built on insight, and market research is essential for understanding your audience, but gathering insights, especially in B2B, can be time-consuming, expensive, and challenging.
In our on-demand webinar, ‘How to Get AI to do Your Market Research’, we explore how AI can support B2B marketers with conducting market research. We will cover:
- What are synthetic personas?
- Using AI to work out what to askÂ
- The risks of using synthetic personas
- How to create a synthetic persona: demonstration
- The six steps to synthetic successÂ
Register to view our webinar on demand by clicking here, and why not get in touch to let us know if our insights helped you.
Napier Webinar: ‘How to Get AI to do Your Market Research’ Transcript
Speakers:Â Mike Maynard
Hi everyone, and welcome to the latest Napier webinar. So, we’re going to talk about how to get AI to do your marketing research. And really, you know, the objective here is to get synthetic personas to do your market research. So we’re going to talk a bit about synthetic personas, what they mean, how you can create them. And then what we’ll do is, you know, really understand how maybe you can do some market research with AI.
Now, the one thing to say is, this is not a perfect solution for market research. There’s certainly some downsides. So we’ll cover some of the downsides as well, and then clearly, if you’ve got any questions, I can hopefully cover those at the end to explain whether or not it might be suitable for your particular market research needs.
So here’s the agenda for the session today. We’re going to run for about 25 to 30 minutes. I’m going to give you a brief introduction to synthetic personas and talk about the risks and also some of the best practices. Then we’re going to talk about how we create a synthetic persona.
And for those of you in the UK in the best traditions of Blue Peter, for those of you outside the UK, probably in the best traditions of live TV, I’m going to try and do a demo of creating a synthetic persona, a very, very simple one, not necessarily one you’d necessarily want to create for market research, but to give you an idea of how to approach things. We’re going to look at where to get more information. So we’ve got a reference slide on that, and then hopefully conclude by thinking that AI could be a way to help us do market research.
So firstly, let’s look at synthetic personas and what they are. Well, a persona basically tries to humanize typical characteristics. So you know, anyone who’s worked for, for example, HubSpot, which is very big on personas, or a lot of marketing automation research, will develop personas. It’s like your ideal customer profile in some ways. So it will decide to just describe a person that you’re trying to reach in your audience, and it will get the typical characteristics. So all synthetic personas are doing is basically creating AI versions of your personas.
So they’re creating AI versions of a particular person you’re trying to reach. It can be built on large amounts of data. It can also be built on relatively small amounts of data. But the neat thing about using AI is you can actually interact with this synthetic persona. So you can choose to interact with the person you’re trying to reach. And obviously, when we look at market research, that can be hugely helpful, because you can ask your market research subjects the synthetic persona questions about what would be more effective.
Sometimes people call these synthetic customers as well. And what we’re seeing is that in B to B synthetic personas are getting used quite widely. And the reason is really simple, is in many B to B markets, it’s incredibly hard to put together focus groups or get responses to surveys. So whether you’re doing qualitative or quantitative research, it’s actually really, really hard to get results in B to B because people don’t have time and don’t want to share their opinions. It’s easier in consumer part of the reason being, consumer markets tend to be much, much bigger in terms of number of audience members. But also there is an issue.
Around confidentiality and focusing on what people consider to be doing their work, rather than going off to a research focus group. And so it’s really hard to do market research. And actually what we find is market research can often be the last person that somebody visited in sales, or it can be, this is my opinion of what I think the market wants. And that’s obviously dangerous, because you’re imposing your opinion. You’re taking, you know, short term views and potentially making quite big business decisions based upon it. So trying to do market research is good. It gives you a much more independent view. It opens you up to more ideas, and it potentially stops you from making very big mistakes.
So what we can do is we can take these synthetic people, these people that we’ve created through AI, and we can do different things. We can do qualitative testing.
So basically this is interacting with the persona and asking questions to get responses. So you can go and ask a synthetic persona how would be the best way to approach you with a headline for an ad for a particular product, for example, very, very simple example, and you can engage and get good feedback. Now, this is actually just like running focus groups in B to B, particularly when you run focus groups. One of the most important things to remember is that, by definition, your sample is skewed. It’s skewed to the sort of people who are prepared to give up their time to participate in focus groups. So as with doing standard market research using synthetic personas, is also skewed to the way that you’ve created these personas, so they might not always be representative of the audience as a whole, and that is a big risk of any kind of market research is making sure you’ve got a representative sample.
But having said that, it’s really great to be able to talk to people who represent, you know, at least a proportion, if not all, but at least a proportion of the audience, and discuss what would be most effective. And it’s a really effective way of brainstorming, because you’re effectively brainstorming with the audience you want to reach. So qualitative testing is very interesting.
Quantitative testing, which, to be honest, is something that’s come in more recently than qualitative is where companies are building, really an army of synthetic personas. And so what you do is you try and profile your audience by building personas in roughly the same proportion they occur in the real world that represent the different kind of views and biases that you see with each of the audience members you’re trying to reach. So you’re trying to build effectively a sample of the full population. Qualitative is much smaller. It tends to be, you know, a small number of personas. And you’re not trying to be completely representative. You’re trying to get good ideas. It’s much more about the kind of brainstorming and creative approach than it is about trying to get firm data. Quantitative research is really trying to make sure that you build personas that represent the whole population you’re trying to reach, and then you use them to tell you how effective different things would be.
So you now understand that within the audience you’re trying to reach, there are different people you’ve built that in you’re trying to represent it, obviously with quantitative research, if you don’t build your army of synthetic personas to be completely representative of the audience you’re trying to reach, you’re going to have problems, because you’re going to get results that are skewed, and the difficult thing is really understanding what that audience thinks about so you need to Make sure that these personas represent the entire population, having said that, of course, if you were to do survey based research, again, you’re only getting responses from people who, by definition, are the people who will fill in a survey form. So there is a group of people who won’t respond to surveys, and they may have different views. So you may not get entirely representative views when you do proper or real human research, you know. So the same problem still exists, whether it’s synthetic or a real person.
So let’s have a look at qualitative. And I think qualitative is really the place where I would strongly encourage anyone to start quantitative requires building a lot of personas, and it also requires a pretty good understanding of your audience to make sure your persona army represents your audience. So qualitative is great, because you can build personas that really represent people you know in your audience, and then you can ask them questions. So here’s an example. So we, for example, might be running a campaign for a client, and we want to test two headlines. So 64 bit risk processor for more performance or risk 564, bit processor with larger cash and faster processing. Maybe not perfect headlines, but they’re two different approaches, and what you can do is you can ask your synthetic persona which of the headline would be more compelling.
And here we see the answer. It’s built from a synthetic persona that we created, and they’re very definite. You know, the longer more specific headline made more difference, and that was because it was specific. It differentiated the product. It was what they called Market relevance, but it gives important information. If you’re building systems with processors, knowing the architecture is quite important, the performance benefits are clearer and there’s no ambiguity. So clarity is good. If you’ve ever marketed to engineers, you’ll know that being specific, being clear, absolutely matters to that audience. So actually, this is a pretty good response. I mean, I would say this is a pretty well defined rationale behind why that headline would be better than the other.
So we can see we’ve got some really useful information from our, you know, just one synthetic persona. And obviously, you know, if you’ve been marketing for many years, you fully understand an industry, maybe you’re going to say that actually, you know, I don’t need this. I know this already. But for people who are newer to the industry, perhaps less experience, perhaps have less technical knowledge, this can be incredibly helpful in making sure that you tweak the campaigns you’re building to really resonate with the audience. So it can be incredibly useful just asking simple questions of a synthetic persona.
But there are risks. So crudely speaking, when you build a synthetic persona, you are basically building it with two things.
What you’re trying to do is you’re trying to provide actually information around that personality, around that persona, so you’re trying to guide the model. But you’ve also got a large language model, and that will probably pull in data that’s been trained on as well. So we’re providing data both through the large language model training, which we don’t control, that’s been done by open AI or somebody else, and also by the The information we provide. And I’ll show you how we do this. But the risk is, and a classic programming phrase is, garbage in, garbage out. So the quality of the persona is going to be defined by the quality of what you provided the information.
You do have the risk that AI will make assumptions, and you do have a lot of risk associated with personas and market research, because personas are simply simplifications. They’re not actually real people. They’re built upon what you understand about those real people. And you don’t understand everything, and quite often, you know what you’re really doing is basically you can end up feeding in some training and then just regurgitating that training out, in which case you’ve got no real benefit of using AI other than a nice, chatty interface.
So that’s really important. The other issue is that, you know, customers don’t necessarily know what they want. So you know, one of the things is, if we look at this, we can ask AI, you know, what did Henry Ford say about faster horses? There is a famous quote according to chatgpt, which says that Henry Ford often said, If I asked people what they wanted, they would have said faster horses. So synthetic personas, if you program them with what people know, they’ll often say what I’ve got already. But a bit better. Now there’s a couple of problems with that, because actually people didn’t want faster horses. We haven’t turned into a world where everybody is in a horse and cart with much bigger and faster horses. We’re all driving cars, as Henry Ford rightly worked out. And also Henry Ford actually never said this. It’s a myth. It’s a widely propagated myth on the internet. And actually, what’s happened is it’s so widely propagated that AI has been trained on this myth, and now AI repeats it as though it’s fact. It’s not true. So you have to be careful with AI, because hallucinations and poor training can result in errors.
So how do we avoid these errors? What are the best practices for developing synthetic personas?
Well, we have six steps to creating a synthetic persona. So six things you need to be successful.
The first thing, the most important thing, is use research rather than intuition. So what you want to do when you build a persona is you want to base it on fact, not on what you assume. If it’s just what you assume, you might as well make the decisions yourself, because all you’re doing is feeding your biases into an AI and then getting the AI to use those biases to answer questions. Maybe it sounds a bit more credible. It absolutely isn’t. It’s just using your biases. Is so use facts. Don’t use intuition. Do some research. And this is a challenge with synthetic personas, is that you have to do research before you can really build a good persona.
So you have to do research in the real world before you can build that. You can also find research data. And what I’d strongly recommend is using multiple sources of data, so the more information you can get, and the wider range of sources, the better the person is going to be. And we need a demo. We’re actually going to do a demo using a single source of data. It’s just for convenience. It’s not what we recommend.
We certainly recommend generating multiple personas, trying different things in the way you set up personnel. Personas, because that will give you slightly different views. And also, typically, we’re all running marketing campaigns that focus on buying committees or decision making units, and those have different people with different motivators, different drivers. And so we want to generate at least personas for each of the members of the of the decision making unit, and ideally multiple personas that reflect some of the subtle differences within each personality. In that decision making unit, you need to eliminate sources of bias. You know, the classic source of bias is, well, I’m going to understand what personas think and why they buy products, and so I’m going to ask my customers.
And that’s fine, but you’ll get personas based upon people who have already bought your product. You’re not going to get personas based on people who are not customers. Typically, if you want to grow your business, what you want to do is understand what the non customers think and how you have to change them to make them buy your product. Very simple, acquire more customers. So surveying your own customers is not necessarily a great way to build synthetic personas. If you’re looking to expand market share, definitely validate your persona through human feedback or additional data analysis. You know the least you can do is at least have a chat with your personas to see if it’s giving you reasonable answers. And then finally, I would say, don’t just rely on a synthetic persona. It sounds like a really easy way to answer all your questions, but you’ve got to include it in other research methodologies. So you’ve really got to understand how your persona answers fit with everything else. And other research can be a whole range of things. It can be, you know, surveys you’re actually running out to the customer base or to prospects, but it can also be feedback from sales teams and much more anecdotal research as well. So make sure that you are openly discussing the results and getting feedback from people who really understand the customers, because they may be able to spot issues that you don’t.
Just a note on the wearing of bias. I mean, I mentioned that, you know, if you ask your customers, you’ll get a different response to if you ask non customers. I mean, a very famous case is Amazon. Amazon was very worried about their hiring, because they realized that typically, their management team was built of white men, and basically, you know what they were doing was continuing to hire white men and promote white men. And they realized, just look at the numbers, it couldn’t be right. There had to be some bias in there. So being engineers, or having a lot of engineers at Amazon, what they decided to do was use an AI tool to remove that bias completely. So they trained the AI tool by basically supplying it CVs of people who’ve been successful at Amazon and people who hadn’t. And of course, all they did was train the AI tool to realize that if you’re white and male, you’re going to get promoted, and if you’re female or if you’re non white, then you’re not and so the AI tool then started recruiting white men doing exactly the same thing as what Amazon was doing with humans. So you’ve got to realize that if humans have biases, training an AI tool needs to remove those biases.
So how do we create a synthetic persona? Well, there’s lots of ways to do it. There are more complex approaches. But if you’re starting, if you’ve not done this before, there’s a very simple approach you can create, and that’s just basically using chat GPT, so you know, probably the most ubiquitous AI tool around there. And what it does is it lets you build a simple synthetic persona. And so what I’m going to try and do is do a quick demo to show you how to build a typical synthetic persona.
So if I open up chat GPT, I can click on explore gpts, and I can click here to create a GPT. Now, suddenly we’ve got a GPT net. Now this is basically a customized model that’s using specific instructions to determine how it behaves.
And so what I’m going to do is I’m going to magically train it with a marketing persona. So we’re going to create a synth marketer here.
So I’m going to paste in some instructions about you’re a marketing professional. You love inbound marketing. You’re a typical HubSpot user, but you believe in a variety of channels. You work for an engineering company, so we’ve got a whole explanation here. And also because what I want to do is make this work. Well, I’m going to feed it HubSpot 2025, state of marketing report as well, and tell it to use that. So we’re going to put that in there.
And okay, it’s uploaded, so now we can start programming it. Chat GPT doesn’t always follow the same routes, but I’m really hoping this is going to work well.
So we just have to wait a minute. As you can see, there’s a little model here whilst it starts setting up, and it recommends we try it out. So on this side of the screen here, we can try it out. And what we can do is we can ask it questions.
So what I’m going to do is start asking it some questions. Does AI improve marketing?
It’s going to have a look. And you know, maybe not surprisingly, AI improves marketing. But what it’s doing is it’s referring back to the content that we had here. So it’s actually telling us, if we look at the state of marketing report, AI’s got all these benefits. So this is exactly what we want to do. We’ve basically skewed the standard chat GPT model to be much more likely to use information that’s in the HubSpot model and in the explanation. So this is really super interesting and super important, because what we’ve done is we’ve created a persona that is just like, what is the aggregate persona? So the overall result from the state of marketing report, so it’s effectively the average of the people that were surveyed in the state of marketing pool.
We can, you know, ask it all sorts of things about that.
But what we could do is, let’s say, for example, I want to make millions, and I’m going to do it by creating an email marketing tool. So if I was to offer an email marketing tool including AI feature. Which one AI feature would be most important to include?
And let’s see what it says here.
And so here, what it’s saying is AI powered personalization and predictive content. Again, this is maybe not that surprising. It’s a pretty typical feature, but what HubSpot is doing is explaining why, sorry, what the persona is doing is explaining why it thinks it’s a good feature. So you can see, if you’re a marketing person, we can create a marketing synthetic persona, and we can get some quite interesting answers.
But what we can do is we can also move away from, perhaps, areas that we’re comfortable with, and we can move into other areas. So one of the things I’ve created previously is an EMI expert persona.
And here what I can do is I can ask the EMI expert something. So I don’t know how many people on the call of EMI experts. But you know what we want to do is we work for an EMI company. We make connectors that filter to remove EMI. How do we promote these things? What are the most important things that make you choose one product over another?
I’m guessing some people are going, I’m not really sure on this. However, our synthetic persona definitely has an opinion. And as with all demos, there’s a nervous wait, as there’s a bit of a delay on Chat GPT, just wait for an answer here.
Okay, and thank you. Chat GPT, you’ve managed to prove it doesn’t always work, so I’ve just pasted that in, and here we go.
So basically, we now have our synthetic persona, which is an engineer who understands EMI telling us what’s most important, and it’s telling us in order as well, which is really useful. Now we’re going to look at this and we’re going to realize that there’s not necessarily anything super deep about what it’s saying, but that doesn’t mean it’s not useful. So what we have from our EMI expert synthetic persona is there’s three. The three primary things are, firstly insertion, loss, performance and attenuation. Secondly, mechanical, electrical, and thirdly, quality and reliability.
Now this is interesting, because actually what it’s doing is it’s it’s almost working through the process an engineer goes through when they’re selecting a product.
So basically the first thing you need to do is get rid of the EMI. And so to get rid of this interference, you have to have something that gets rid of the interference sufficiently to get the interference low enough to meet requirements. And that’s what point one says. So the first thing you need to say is, this is the performance in terms of removing EMI.
The second thing it talks about is mechanical and electrical compatibility. So fundamentally, if I’ve got a space for a connector, this connector has got to fit in there. It’s also got to be the right voltage, the right current, and it’s also got to meet the right temperature range and things like that. So again, feature based stuff, but important things in that process.
And then lastly, it’s talking about quality and reliability, and it’s highlighted the importance of supplier reputation and certifications. So if I was building a marketing campaign, you know, I might want to talk about, you know, the performance of our filter, so the insertion loss. I might want to talk about the compatibility with mechanical and electrical, and I might want to talk about my credibility. Now, if you’ve attended any of our AI webinars before, you’ll know that the way AI responds to things is it actually introduces some randomness. And this is really interesting, because the first two answers are pretty much exactly the same as I got this morning when I did a dry run of the webinar. Answer number three, whilst it’s a very good answer, it’s actually different to what I got this morning. So this morning, you know, the third priority was cost.
So it’s often worth asking the same synthetic Persona The question two or three times, because sometimes you can get different answers that can be useful. So even without creating multiple personas, you can get multiple answers, because inherently, in these large language models, there’s some randomness.
So hopefully that’s got you intrigued. They were very, very simple approaches, so really simple ideas, and what we’re trying to do when we create a persona is we’re really enhancing that large language model with some specific data and some specific behavior that we know a typical persona will display. And so we’re making, in this case, chat GPT, a much better model of what someone who we’re trying to reach would actually say, but doing it, as you can see, with relatively little effort. Now, as I said, Generally, we recommend multiple sources of data, a lot more information, rather than just saying, build it all off HubSpot report. But even then, we got some good answers.
So how do we how do we get this information? How do we create it? Well, obviously, you know, the first thing people are going to say is, I’m doing synthetic personas because it’s really hard to do surveys and get responses, and you’re telling me I’ve got to get do surveys and get responses before I create the synthetic persona. That kind of feels like it’s really difficult, and it is, but the great thing is, is you can pull in lots of information, so you can put in public research. You can do your own research.
You can actually do interviews. You know, one interesting research project I saw a while ago was a university that have basically created a questionnaire, got people to answer it, used AI to summarize and categorize the questionnaires, and then developed multiple synthetic personas from those questionnaires. You can also do AI interviews as well. And again, I saw another research bit more recently, where a university had basically got an AI tool to do 1000 interviews and use that internet interviews as source material to create 1000 different personas. Obviously, all this can be quite time consuming, and quite, you know, difficult to do, but definitely, if you spend the time, you can create really good personas, and you can create personas that represent different customer segments.
There is a lot of information about synthetic personas, so there are a few links here. If you want to grab a screenshot, feel free. Obviously, if you’re you know, if you wait when we publish the webinar, we’ll give you a link to the replay and also to the slide, so you’ll be able to access these as well, but feel free to take a screenshot now and go and learn a bit more about synthetic personas.
So finally, you know, what do we what do we think? Have we found that AI can help you?
Well, if we look at this, you know, marketing research in B to B is really tough. It’s hard to get respondents, hard to get panelists. It’s time consuming, and, you know, it can be very expensive, and often people don’t have the budget to do that, so actually trying to do everything through.
Market research with real customers or real potential customers, is difficult, so synthetic personas is definitely a good way forward. And to make it better, there’s data that exists to create synthetic personas. You might be able to find some data. So as an example, quite often, companies will do surveys. I know our clients, for example, have done surveys on attitudes of people who are involved in IoT design, you can very easily take that public survey data and put it into a synthetic persona. But also you can ask sales and one of the easiest ways to create synthetic personas is to get a group of sales people and get them to tell you what they think personas were, you know, actually believe in if you merge that all together, so you take answers from four or five sales people, you merge that all into one briefing, you’ll probably get quite a good profile of a typical customer.
Obviously, be careful of bias, and in particular, be careful of being biased to your customers if you’re looking to grow your market share, because what you really care about, if you’re growing market share, is people who are not your customers. And also, you know, the one thing I’d say is, whilst you can create simple personas, and it’s quite fun, and sometimes it can be quite useful, particularly for brainstorming, it’s really worth getting some advice, because it’s very easy to create personas that actually don’t represent your your audience, and if you do that, you can end up making bad decisions, because the people advising you these robot personas actually aren’t represented with the people who are going to be your customers.
So thank you very much for listening.
If you’re interested, we have another webinar coming up next month, the impact on AI in SEO and search advertising. So this is really talking about the impact of generative AI results that appear at the top of search pages when you search for something you know, Google and other search engines are now trying to give you the answer using AI. What does that mean for your business? Does that mean that mean all your traffic to the website is going to stop? We’re going to investigate that now. If anyone’s got any questions, please feel free to put them into the chat now, and I’ll just take a look and see what we’ve got and what we can answer.
There’s just one question I can see at the moment, so please feel free to add another one.
So I’ve been asked, how many personas do you need to take to make the project to create, to make the project worthwhile?
And this is a great question, and I think the answer is actually not that easy, because it really depends on what you’re trying to do. So if you’re trying to create, you know, some qualitative research, you know, maybe trying to say, you know, which headline is better, how would you write this? Actually creating one synthetic persona can be super helpful, particularly if you’re not an expert, or the person running the campaigns is not an expert on that market. So sometimes one persona can work really well. If, however, you’re looking to build a launch for a new product, or even build product specifications, I’d really say, you know, relying on one synthetic persona, just as asking one customer is incredibly dangerous, so what you need to do is build a portfolio of personas that represent the range of different views in your customer base.
So hopefully that’s been helpful. I’m just having a look, and we don’t seem to have another question, which, with any luck, means that everybody’s found this useful.
We have also reached the end of our half an hour. So what I’d recommend is, if there’s questions that come up, or if you’d like to know more about creating synthetic personas, my contact details are here. Email me, Mike at Napier, B to b.com and I’d love to talk about this. Anyone who knows me well knows that I’m a absolute geek, and these technical things, you know, really interest me. We’ve had some really good positive results from creating synthetic personas, and we’ve also, to be honest, seen some of their pitfalls as well. So please feel free to get in contact. I’d love to chat to you about it.
Thank you very much for listening to the webinar. I really appreciate it. I hope that we’ll get to meet again and the next webinar in March and talk about AI and search results. Thank you very much. Please feel free to send me an email if you’ve got any questions.
