Some Thoughts About Copilot Pro Custom GPT Builder and the Future (Premium)

Following the quiet release of four custom Copilot GPTs for image creation, vacation planning, cooking assistance, and fitness training two weeks ago, Microsoft this week made its Custom GPT Builder available to all Copilot Pro subscribers. So I decided to take a look, though I feel like the GPT I’d like to build, one that utilizes only Thurrott.com as its data source, isn’t currently possible.

Copilot GPT Builder is well-hidden, at least for now: If you visit the Copilot website, you’ll find a new “See all Copilot GPTs” link under Copilot and the four custom GPTs in the right rail. To me, this link suggests the availability of more custom GPTs, but that’s not the case currently. Instead, the pop-up that appears provides access to a new “Create a new Copilot GPT” option in addition to those four custom GPTs.

The resulting experience provides two ways to create a custom GPT: You can do so conversationally, where Copilot prompts you to give it some starter information—a name, plus one or more prompts that describe what you’re looking for—so it can then step you through the rest of the process. Or, you can switch over to the Configure tab and just fill out a form.

I started with the conversational wizard, as will most, I’m sure. In time, I suspect I’ll be familiar enough with this tool to jump right into the Configure tab. But the nice thing about the wizard is that it fills out the form for you, so you can see how that works.

I was explicit: I want a GPT that searches my website thurrott.com. But as I quickly found out, getting there might be impossible. No matter how I configured this thing, it would poll the entire web for the information it needed. More on that in a moment.

Copilot recommended the name Thurrott Search Assistant, but I changed that to Thurrott GPT. Perhaps Thurrott.ai or similar would be better, I don’t know. First things first.

Then, I was asked for the main purpose of my GPT. To this, I responded, to answer questions about the personal technology topics that we cover on Thurrott.com.

After it updated the GPT profile, I was asked about the emphasis. More of a focus on software or hardware? Should it avoid any brands or topics? It should focus on both software and hardware. The emphasis is Microsoft productivity solutions but we also cover Apple, Google, and other personal technology topics.

After another profile update, it asked about rules and guidelines, and tone. Are there phrases it should avoid? To this, I wrote it can be formal or informal depending on the topic, in keeping with how I approach my own content. It should avoid politics and other topics not related to personal technology.

Understood, I was told, after another profile update. Should this GPT sound friendly, professional, enthusiastic, or something else? It should be friendly and professional, I wrote.

Perfect! It said, updating the profile yet again, listing the details.

It was time to test the GPT. So I clicked the “Preview Copilot GPT” button, sure that I wasn’t going to get what I asked for. The presentation is interesting—I’d like a custom “T” graphic there, and a less wordy description—but fun to see.

That said, I wasn’t disappointed—well, I was disappointed, but you know what I mean—when I asked my first question—how much does Microsoft 365 cost?–and it responded using data collected from Microsoft, PC Magazine, and How to Geek.

Wa-waa-waaah.

So I prompted it with, please only provide information from thurrott.com. Not surprisingly, that worked: The resulting information came only from my website.

But I never could get it to just default to only using thurrott.com, which was the point of this exercise. What I’d created, I guess, was a very shallow film over plain vanilla Copilot. Maybe less than that: I’m pretty sure I could just ask Copilot the same questions and get the same (types of) responses. (Every response is generated on the fly and, like snowflakes, no two are literally identical.)

So I asked. How can I restrict the data source to thurrott.com only?

The answer was a non-starter: You can upload PDFs, text files, and other files to Copilot. You can “upload the content of your website as a knowledge file,” too, whatever that means. But my site, which I suspect isn’t particularly voluminous compared to the rest of the web, has about 18,000 articles and whatever number of forum posts, comments, and other content. I can’t really “upload” it to Copilot.

I tried to force the issue. Turning to that Configure tab, I could see how my answers to the wizard’s questions had filled out the Instructions list. For example, there were bulleted items like “I can be formal or informal depending on the topic” and “I should maintain a friendly and professional tone.”

So I added one of my own: I should only look for content from thurrott.com. Then, I saved it, and tried again. Nope. A question about Epic v. Apple resulted in links to Wikipedia, Justia.com, CNBC, and other sources, but not thurrott.com. When I asked it to again use only my site, it did so, again. But having it default to thurrott.com seems beyond its (or my) capabilities.

Hm.

As I write this, I’m flying somewhere over the United States on our way home from Mexico City and my connection is, at best, sporadic. But I can see from the Microsoft Learn website that Copilot Studio, a more full-featured and commercial tool for building Copilots (i.e. GPTs) does what I want: You can “let your copilot create responses in real time with generative answers and information from a website you choose.” You can also “add other data sources, including internal resources like Sharepoint or OneDrive, and public websites.” So what I want is possible today. Just not with the tool I’m trying to use (or, again, maybe it’s me).

To be clear, what I’m trying to create here isn’t a glorified website search feature: Yes, our search sucks, but a GPT can summarize information from across a spectrum of articles, as I imagine it, and that can be really valuable. As with more general web searches, many site searches today are almost certainly from people looking for an answer. This can be specific—how do I enable Shake in Windows 11?—or general, as in what’s going on with Sonos and Google?

But I am also interested in this notion of a personal GPT given my decades of work archives. And in playing around with this limited Custom GPT Builder tool today, it occurs that a public front-end to that might be interesting. That is, imagine you came to Thurrott.com to find out about a specific topic—say, a sidebar feature that used to be in Windows—and the search results, for lack of a better term, included an AI summary, a standard list of article links, and some kind of summary/list of content from my archives? Some of this would be raw, from notes I might have taken years ago, while some would be from fully formed articles or books from the past. Tell me that’s not interesting.

I think it is. But it’s also science fiction, at least for now.

In the meantime, we have to start somewhere. So let me know if you have any ideas about restricting a GPT to content from a specific site. That seems like a logical enough first step.