How to Use ChatGPT as a Search Engine – Fight AI With AI

Using ChatGPT to Crawl Search Engines

One day in a blogging Facebook group, I made a joke about wanting to create an AI-powered search engine that crawled Google. Wouldn't it be nice to have a search engine that delivered good results in Top 10 SERPs-style format, while cutting out the garbage like ads, AI Overviews, out-of-date Reddit threads, etc.?

Laughs were had, but at the core of any good joke is a kernel of truth. The current argument in artificial intelligence is that machine learning isn't stealing, it's acceptable training on publicly available data. Well, I just so happen to know a few sites that have a lot of data (data that is almost entirely third-party, for the record), and AI could really organize the data better for how I like to consume it.

Yes, I'm talking about search engines, like Google.

At first, I was just going to let the joke ride. But eventually, my curiosity got the better of me, and I logged into ChatGPT to see what it could do as a what-if. I put in the following prompt using ChatGPT-4o model:

“Search Google for the term “where to stay in Edinburgh” and bring back a curated list of the best 10 websites with URL and meta description after searching the top 100 results.”

I was expecting an error message saying that ChatGPT couldn't search a specific website, but something surprising happened- it brought back results.

In fact, for not having any further direction, not only did ChatGPT provide me a list of 10 websites, all without AI Overviews, ads, crappy product carousels, etc., the returned list of SERPs was, dare I say, good. Like, really good. Not quite Google at its best good (which I'd peg as circa 2017-2019), but better than the current iteration of Google that AI literally crawled to generate this result.

So, I started tailoring my requirements more to see how much it could be improved.

The first issue I noticed was that the links weren't clickable, only text, so I wanted to see what I could do next to refine and create a listing that looked like old school SERPs:

“Search Google for the term “where to stay in Edinburgh” and bring back a curated list of the best 10 websites (with clickable URL and meta description) after searching the top 100 results. Make the links clickable, and try to only include local sources over national/international sources.”

Oh my, it was better than I expected. Within just a few more adjustments to my initial query, I trained an AI chatbot to crawl search engines for me, take out the kind of sites I don't want to see, and provide better results- all at the expense of waiting a few extra seconds to generate and the monthly subscription fee I was already paying for.

But, I refined this further with other query adjustments, such as:

• “Prioritize local websites over national and international sites.”
• “Do not include user-generated content sites like Reddit, TripAdvisor, and Wikipedia.”
• “Search Google, Bing, Yahoo, and other search engines.”
• “Search the top 100 queries and return the best results.”
• “Prioritize content that was published and/or updated in the last 12 months.”
• “Do not show more than one result from any given source domain.”
• “Never show results from [top-level-domain].”
• “Provide a 2-3 sentence summary of each article.”
  • This was not meant to re-create AI Overviews giving me the answer to my query directly, but rather to form a better Meta Description to help me figure out which articles I'd want to click on.

In just a few revisions, I essentially trained ChatGPT to be Google, while actually being better than Google, without having to actually load Google myself.

You know what? Machine learning is fun!

Is ChatGPT Actually Searching Google?

Now, you may be wondering if AI tools like ChatGPT can actually search Google. Part of me wondered if they had their own inherent algorithms to organize websites, and if they were just ignoring that element of my queries. This wouldn't have surprised me in the slightest.

Unfortunately, it also wouldn't be as funny than if it was actually searching a selected search engine directly.

So I did a sanity check with the query “Search Google for the term “[query went here]” acting as a private brower and return the top 10 links in order” and compared the results with the same Google results as shown in a private browser on my page.

It hit 10 out of 10 on the first try. 

Honestly, I was surprised by that, so I tried some more. The next few queries weren't perfect, but still hit 7 or 8 out of 10. In one query, it skipped some results, but otherwise still picked up 10 within the top 15 or 20 if I kept scrolling. I tested another query that I knew had some really outlandish results that absolutely shouldn't be there, and, yes, it picked those up, too.

So while I can't say with absolute certainty that ChatGPT can search Google upon command, mostly because I can't verify via source code, my gut feel is that it absolutely can- if my limited test can be extrapolated out at least.

This Chat Is Not Without Its Faults

Now, the bad news. As of now, this chat is good, but not perfect.

Yes, some of the results are much, much better than the slop you'd normally expect from a conventional search. But at the same time, there is still significant room for improvement.

A few of the biggest pitfalls here were that the chat would periodically revert back to just telling me the answer, rather than providing links (so I'd have to remind it only to return SERPs style links, which it would in an updated response) and that it would ignore select requirements (like not including UGC sites, forgetting to link the title to the destination page, or only showing one domain in results).

Still, a few updated prompts later would fix these, for a short while at least, and I'd have a pretty good working search engine all within ChatGPT.

So, how can you get started with something similar? Try this prompt to kick off a new chat:

“I want this chat to behave like a search engine and return 10 unique website links for any query provided in SERPs style layout. The title should be a clickable link to the destination website, and a 2-3 sentence description of each article should be provided thereafter.

Search the top 100 results on Google, Bing, and Yahoo, and return the 10 most appropriate for the search intent. Prioritize sites that have named authors, with local expertise, and do not include international brands or UGC sites like Reddit, Quora, TripAdvisor, or Yelp. Prioritize articles that have been updated in the last 12 months. Do not repeat websites within the list- only one listing per source. Do not include sites with AI content (images or text).

Start with “[replace here with a query]” to test an example.”

This query has, for me at least, got me close to a functional search engine in just a single send. Once the chatbot is trained with this as a saved setting, I would only need to insert queries like “where to stay in Edinburgh” and it would automatically return results in SERPs format.

The most common mistake I've seen here is that the title often comes back without a clickable link, but you can try a simple command like “regenerate with URLs clickable to the destination site” or some variant like that to improve the query. Similar variations can be tested for other issues like those highlighted above- although getting rid of multiple listings from the same domain has been the hardest for me so far.

It may take a couple iterations, but odds are good that starting here will get you on the path to having a working search engine within ChatGPT!

So, who needs AI Overviews when you have literal AI? Apparently the answer is no one, and I'm honestly okay with it.

Building What AI Search Should Have Been From the Start

To finish this one, I really only want to say that it feels like search companies missed the true power of AI here, and it's going to hopefully bite them in the butt long-term.

The problem stems from the fact that no one wanted another company to corner the world's information. Wikipedia already does that. Most users simply wanted search engines to stop being garbage and offer individually curated results for their needs and interests.

Given how much data Google (and, let's be honest, literally every tech company) has amassed on me in the last few decades, you'd think a better AI play would have been to curate SERPs to my browsing habits- matching content types and sources to the kinds of media I historically consume.

Do I want to see a TripAdvisor review from 11 years ago where someone complained the hotel restaurant was too far from the elevator? Literally never. Do I want to see a private Reddit user's thoughts on the best restaurants in a city, when that list was published five years ago, and half the places have closed? Also no, and I'm not sure anyone wants to read content about closed businesses at all.

It is almost like a machine could, and I'm just spitballing here, learn that about me, and adjust SERPs on the fly to present a curated list of results from media types I prefer (and repeat to create a unique environment for each of the billions of users out there).

Unfortunately, that would've only made too much sense. Instead, we're quickly approaching a world where these companies want to control the world's data, and this is where I opt out in support of an open web.

Thankfully, my AI prompt above fixed a lot of these issues in just a few sentences of refinement, and gave me the search engine experience I (mostly) wanted from the start.

So, why exactly do I need to use Google anymore? Truth is, I don't, because AI will search it for me instead. I'm here for it.