Creating a new type of Research Deliverable with a Custom GPT

Ray Poynter, 24 April 2025

---

Let’s face it, one of the least impressive aspects of insights are deliverables, and in the eyes of the end client, this is the bit they are paying for. Reports are too static, most tabulation tools are too fiddly, and frankly, most dashboards show lots of things you don’t need to know.

Now there is a new, exciting alternative. ChatGPT’s Custom GPTs (and similar products from other vendors) are opening up a new, and I believe, better way of transferring the results of your research to the people who need to use it. (At the moment, there is one caveat, but I am sure that it will soon be dealt with.)

A Deliverable as a Custom GPT
A custom GPT (or just GPT) is something you create by combining an LLM, some proprietary information and a set of instructions for how the LLM and information should be used. In the case of OpenAI and ChatGPT, customer GPTs are a built in, easy to use feature of the platform. To create a custom GPT you need one of the paid licenses (e.g. Pro, Team Enterprise etc). To access a GPT that somebody has built you need a ChatGPT license, but it can be one of the free ones.

An Example
The idea is easier to explain with an example. Here is GPT that we built from the second wave of a NewMR study, which we ran in 2024. The title and subject of the study were “Artificial Intelligence and Insights|.

The GPT builder in ChatGPT allows you to add ‘Knowledge’. Knowledge refers to the items it will refer to when responding to the user. In this case, we included:

• The Presentation
• Some tables from the data collection package
• The questionnaire
• The raw data from Wave 2 (in an Excel file).

As well as Knowledge, you need to add Instructions. Instructions are commands that define how the LLM and the Knowledge should be used to answer the user’s questions.

The key thing with all GPTs is to get the instructions right. In this case the Instructions included: “This GPT is a specialist assistant designed for market research professionals. It acts as a knowledgeable expert grounded strictly in the insights, data, reports, and findings of a market research study conducted by NewMR on the future of AI in the market research industry. It answers user questions using only that information. If a question cannot be answered directly using content from the study, the assistant must explicitly state that it cannot provide an answer based on available data.”

In the next couple of sections I will show you how it works for the user.

Extracting information from the Presentation
We can ask the GPT questions like “How did the views about Synthetic data change between Wave 1 and Wave 2?” To answer this question the GPT will tend to look at the Presentation, as that is the only source of Knowledge about Wave 1.

When we asked this question, the GPT showed some key data and wrote a summary. It also offered to produce a chart, like the one below:

The key reason this approach is such a step forward is that it is user-centric, rather than data-centric. The user gets their answers by chatting with the GPT, in the same way many people are talking to LLMs every working day. This replaces the data-centric tools of the past, empowering users to explore.

Accessing the Underlying Data
For users who want to dig deeper, they can query the underlying data. For example, with our GPT, the user can ask questions like “Only using the records flagged as complete, create a table from the raw data that shows how many interviews were collected during each full week of data collection.” And then they could ask “Show me this data for the three regions with the most responses”. The chart below is the result of these two chats.

Limitations and Alternatives
There are still some limitations, but we expect them to be resolved soon, and alternative solutions are available. For example, with Copilot you can achieve something quite similar with Agents, and products like CustomGPT.ai allow you to build GPTs where you do not need a license with a platform such as ChatGPT or Copilot.

There are perhaps three key limitations worth mentioning:

1. To run a custom GPT made via ChatGPT your client has to have a ChatGPT license (but it can be a free one, the creator needs a paid license).
2. A user could potentially access every piece of information provided in the Knowledge base by asking multiple queries, so careful consideration is needed regarding intellectual property and data privacy.
3. Providing raw data to users without training could lead to misinterpretation or misleading conclusions due to poorly structured queries..

Do you want to learn more about using ChatGPT for Quantitative Research?
This topic (creating GPT deliverables) is just one of the many topics we cover on our course about Using ChatGPT for Quantitative Research.

The next course is scheduled for Thursday, 1st May, where we will cover topics such as project and questionnaire design, data cleaning and preparation, and advanced data analysis. We will also cover how to create new GPT deliverables.

You can book a course by clicking here.