Disclaimer: this blog references a pretend government service. This is purely fictional and has been created for illustrative purposes only. 

In part 1, I ran an experiment using Gemini Deep Research to see if it could help with content design when there’s no time or budget for user research. The fictional task was to create an application form for a UK government service called Borrow a Digital Device.

It took nine versions of the report, multiple prompts, and a fair bit of negotiation (and repetition) to get something that stuck to the brief. But the final version pulled together a decent set of insights and recommendations — and I wanted to see how useful they’d actually be in practice.

As before: anything I typed into Gemini will be shown in blue. Anything it gave me back will be in orange.

How useful would the insights be on an actual project? 

None of the insights in the report would be new or surprising to an experienced content designer, or other user-centred design (UCD) professional. However, the ‘Actionable Insights and Recommendations for Form Design’ would serve as a useful checklist to remind a team working on such a form about everything they need to bear in mind. For example: 

Website Language and Content Comprehension: Employ plain English, avoiding jargon, technical terms, and acronyms to ensure clarity for users with varying levels of digital literacy and cognitive abilities.21 Utilize short sentences and paragraphs, and structure content with clear headings and subheadings to improve readability and scannability.21 Provide clear and concise instructions for each step of the application process, minimizing ambiguity and potential for errors.22 Incorporate hint text within form fields to offer additional context or explanations without overwhelming the main label.22

I also like how the report really emphasises the importance of going above and beyond with explicit instructions – not taking for granted that users in this scenario would understand how to complete the form. 

It is also really nice to have all the insights from so many different sources pulled together in this format. To do your own desk research and compile a report on this scale would take several days, so it is definitely a useful tool to save some time. 

However, having a closer look at some of the insights and the sources it has cited, I am not confident that the report is an entirely accurate reflection of the sources it cites. For example, the report states in the section ‘Frustrations Experienced by Older Adults’ that: 

The language used on government websites is often perceived as complex and filled with jargon, making it difficult to understand instructions and the overall purpose of the content.

The source for this is given as an ‘Offline and overlooked’ report published by age UK in 2024. However, looking through this report, I cannot see any evidence for this insight. If decisions were being made off the back of this, it could be problematic. It could also be really embarrassing if this was quoted to stakeholders and someone challenged it. 

I’m assuming that Gemini has made this leap due to there not being much publicly available research about older users interacting with government services. This would also explain the problems I was experiencing with earlier versions, where it was reluctant to remove references to the studies from 2003 and 2006. It’d be interesting to look at the validity of insights for topics where it has more sources to draw from – I’m assuming it would be more reliable. 

Pushing it further

I was interested to see how much more Gemini Deep Research would be able to offer, so I decided to give it a further prompt related to my scenario:

Can you also tell me what information we should be capturing in the form?

One of the ways it approached this task was to do an analysis of existing digital inclusion initiatives. It scoured the web to find application forms for similar ‘borrow a device’ services (mostly councils and universities) and created a useful comparison table of what information they ask for.  

It also had a section about how to determine eligibility for the service for the different user groups, for example: 

To assess elderly individuals with limited digital skills, the form could include an optional question about the applicant’s age, followed by a self-assessment of their digital skills. This could involve asking them to rate their confidence in performing basic online tasks or to indicate which tasks they find challenging.17 Questions about their current access to digital devices and the internet could also help determine their level of digital exclusion.

I was also really impressed by its suggestions about other questions you could consider: 

While not strictly necessary for eligibility assessment, including questions about the intended use of the borrowed device can provide valuable insights for the ‘Borrow a digital device’ service. This information can help the service better understand the needs of its user base, potentially tailor support and resources, and measure the overall impact of the program.6

This showed a really good understanding of service design and its focus on user needs. 

It also had a dedicated section about how to ensure the form was accessible, which included some insights that were broader than simply meeting WCAG 2.2 level AA guidelines: 

The language used should be inclusive, employing person-first language and avoiding any discriminatory or stereotypical terms related to age, disability, or other protected characteristics.55 Consideration should be given to potential time limits on the form, as these could disadvantage some users.47 

Would I recommend using Deep Research for desk research?

Using Deep Research for desk research provided several key benefits:

  • Rapid information gathering: It saved me countless hours of manual research, quickly compiling information from diverse sources.
  • Breadth and scope of sources: It is likely to find insights you’d have little chance of discovering if you were doing the desk research without an AI tool.
  • Structured output: The research was presented in a clear and organized format, making it easy to digest and apply.
  • Inspiration and ideas: It sparked new ideas and perspectives that I might not have considered otherwise.

However, as with any AI tool, it should be used with a healthy level of scepticism. Make sure you carefully check through all the sources, and spend a bit of time validating that any important insights are an accurate reflection of the source material. As I found, I’d recommend being extra cautious if your research involves a fairly niche topic, where it might be struggling to find relevant insights. 

It was also frustrating to use at times, like when it repeatedly said it couldn’t help me by removing the out-of-date sources, but then randomly changed its mind. However, it is still a really new tool and I imagine this will happen less frequently as the technology improves. 

It also goes without saying that any form of desk research is no substitute for real user research. Drawbacks include:

  • Lack of empathy: AI cannot replicate the empathy and nuanced understanding gained from interacting with real users.
  • Limited context: Deep Research relies on existing data, which may not fully capture the specific context of your service or target audience.
  • Potential bias: AI models can inherit biases from their training data, which could affect the research output.

Tips for using Gemini Deep Research

Based on my experiment, here is my advice for anyone who wants to use Deep Research for desk research:  

  • Craft your prompt carefully. Give as much context as possible about your role and the problem you are trying to solve.
  • Think about different angles, for example, I asked the follow up question about what questions should be included in the form and it led to some useful insights and a comparative analysis of the questions asked in similar services. 
  • Ask it to add citations to the report.
  • Specify any criteria for the sources you want it to look at, for example, how long ago they were published, or if you want it to just focus on research from a particular country.  
  • Carefully look at the sources it has used to make sure they are relevant, from trustworthy sources and follow your specified criteria – garbage in, garbage out!
  • For any insights that will inform your decision making, take the time to check if the insights in the report are accurate.  
  • If you get a ‘I’m sorry, I can’t help you with that at this time’, keep asking! Or, if it really seems stuck, start a new session.