How to use LLMs to humanize your content and scale your research

One of the major things we talk about with large language models (LLMs) is content creation at scale, and it’s easy for that to become a crutch. 

We’re all time poor and looking for ways to make our lives easier – so what if you could use tools like Claude and ChatGPT to frame your processes in a way that humanizes your website work and eases your day, rather than taking the creativity out of it?

This article tackles how to:

• Analyze customer feedback and questions at scale.
• Automate getting detailed and unique information from subject matter experts.
• Analyze competitors.

These are all tasks we could do manually, and sometimes still might, but they’re large-scale, data-based efforts that lend themselves well to at least some level of automation. 

And having this information will help ground you in the customer, or in the market, rather than creating your own echo chamber.

Analyzing customer feedback at scale

One of the fantastic features of LLMs is their ability to:

• Process data at scale.
• Find patterns.
• Uncover trends that might otherwise take a human hours, days, or weeks. 

Unless you’re at a global enterprise, it’s unlikely you’d have a data team with that capability, so the next best thing is an LLM.

And for this particular opportunity, we’re looking at customer feedback – because who wants to read through 10,000 NPS surveys or free text feedback forms? 

Not me. Probably not you, either.

You could upload the raw data directly into the project knowledge and have your LLM of choice analyze the information within its own interface.

However, my preference is to upload all the raw data into BigQuery (or similar if you have another platform you prefer) and then work with your LLM to write relevant SQL queries to slice and analyze your raw data.

I do this for two reasons: 

• It gives me a peek behind the curtain, offering me the opportunity to learn a bit of the base language (here, SQL) by osmosis.
• It’s another barrier or failsafe for hallucinations.

When raw data is uploaded directly into an LLM and analysis questions are asked directly into the interface, I tend to trust the analysis less. 

It’s much more likely it could just be making stuff up. 

When you have the raw data separated out and are working with the LLM to create queries to interrogate the data, it’s more likely to end up real and true with insights that will help your business rather than lead you on a wild goose chase.

Practically, unless you’re dealing with terrifyingly large datasets, BigQuery is free (though to set up a project, you might need to add a credit card). 

And no need to fear SQL either when you’re pair programming with an LLM – it will be able to give you the full query function. 

My workflow in this tends to be:

• Use SQL function from LLM.
• Debug and check data.
• Input results from SQL query into LLM.
• Generate visualizations either in…