The wide availability of large language model APIs transformed the expectations of what AI products should be. With their borderline magical power, it became easier than ever to develop a proof-of-concept for a powerful AI tool quickly. Turning it into a final product, especially one that is a joy to use, is a completely different and increasingly more challenging task. This is a story of how our AI-powered domain name search grew from a minimal demo to what we think is the best domain name recommender in the world – Name Sentinel.
When ChatGPT became a thing, Alicorn’s team worked on (and with) AI solutions for the domain name industry for more than a decade. From work automation to content classification and portfolio valuation to domain name recommendation. As the industry went through its twists and turns, many things changed – but its centerpiece remained the same: the first step of your domain name search.
The search box
The “Google it” of domain names. You type a domain name you want, click the search button, and hope it will be available. If it is not, you might get a list of recommended available names, but these recommendations are rarely useful. There is not a lot of insight you can squeeze from someone’s one-word search to recommend anything worthy, other than suggesting a different extension or generic additions to search terms.
Frankly, just like the Google search, the search box remained unchanged because it did its job so well. It was obvious what you must do to check for domain name availability. If you know what you want, there is no better way to do it. For the last dozen years we have been testing, collecting data, and changing things to make the search box faster, prettier, and more fun to use while sticking to the same old user interface – and we will probably continue to do so for at least a dozen more.
A new type of users
The thing is, our data showed that things are shifting. As the time passed, a new generation of users emerged. Domain name searches happen way earlier – when our users are unsure what the domain name they want exactly is. Finding a domain name became part of naming their businesses and products. Even if they know what they want to search for – the domain name is often already taken.
Alicorn’s largest client in the domain name industry is .ME. During many brainstorming sessions with their team, one idea often popped up. What if someone can describe to us what they want so we can help them find a perfect domain name – not by spitting out a “take it or leave it” list of good names, but by turning this search into an iterative process where they can refine suggestions, change their mind, give additional input, and even save the progress so they can continue it later?
In 2020 we started experimenting with language models like BERT, exploring if and how they can help experiences in the industry. It became clear that the technology that will make this kind of experience possible is just around the corner. We started working on the UX and AI workflow for a recommender based on language models. Taking lessons from brand sprint workshops, our designers explored how to replace a simple text box with the right combination of short questions and checkboxes which will help the user form and refine their conversation with AI.
With these older models, things felt almost right, and when OpenAI GPT APIs came, everything just clicked. For the first time, we had recommendations that were overwhelmingly worth considering. Just like that, with a switch of a single building block, the first working demo of Name Sentinel, one of the earliest LLM-powered domain name searches out there was born.
The road from a working demo to a final product is hard, and that goes to extremes for an AI product. Contrary to most software development, developing an application with machine learning components is a gentle art of “good enough”. Getting from what is good enough for a demo to what is good enough for a final product is hard work, requiring tweaks and novel approaches on every corner.
Proof of concept is not a product
When you make a generative AI application, you face two dangerous opponents: hallucinations and slop. Hallucinations are the overconfident responses from an AI model that has nothing to do with truth. If you have used ChatGPT you have probably gotten an incredibly convincing response, but it turned out it was made up of imaginations and lies. Whenever you have a generative AI component inside your software – you run a risk of this happening. If your product allows for iterative refinement, the risk increases.
This is why we surrounded our AI components with guardrails and safeguard routines – to get correct, quality responses effectively all of the time.
Slop is a different beast. To get quality generative responses, you have to give your model time to think, and there is no way around it. When you try to solve issues with AI components interacting with each other, not only does the performance hit increase, but the chance of low-effort output gets greater. We tested our demo with various international crowds and found out that they did not compare it to other generative domain name tools but to a plain “search-box” experience. With this mindset, AI search felt sluggish and users were not too keen to engage in iterative, conversational refinement when it meant waiting around half a minute for the next batch of results to appear. We had to retain suggestion quality, while dramatically improving on search performance, and getting there required a hard detour.
Reinventing the basics
We took a lesson from nVidia. In video game graphics race, framerate is probably the most important metric. The more frames (pictures) per second a graphics card can draw, the smoother and more impressive the experience gets. They created an amazing feature called DLSS frame generation that takes two consecutive rendered frames and uses AI to generate one that goes in between them, effectively doubling the framerate.
The problem with this approach was that instead of showing the frame the card has rendered, you have to wait for the next one to do all the needed calculations, effectively delaying what the user experiences by 3-5 hundredths of a second and making experiences feel sluggish as a consequence. Instead of giving up on the AI technique, nVidia looked where else they could reduce latency to make up for the one they introduced. They changed how their cards communicate with the CPU and optimized the signal processing for input devices and displays, shaving milliseconds on every step to get to the point where the delay they introduced isn’t noticeable anymore.
This kind of reinventing the wheel in domain name searches is what we needed to make our tool feel great to use. We set a target metric – for this to truly work we had to get the time that goes from clicking on the search button to getting your results from 30 seconds down to less than 3. Sure, making AI respond quicker is one piece of the puzzle, but getting there also included creating software that processes terabytes of data daily to create local, blazingly fast structures for bulk domain name searches, along with creating components that cleverly control input, expectations, and output from the LLM model so we can intercept the output and send its parts to processing before the response is fully generated.
The costs of getting AI to work this fast outweigh the costs of using AI language models in the first place, but the response from testers immediately showed that it is worth the price. They found it a joy to use. Making things so responsive cut the time for thinking about how the tool works and made users give in to the magic of collaborating with AI to find a perfect name.
Some interesting, unexpected upsides also materialized: users registered significantly more domain names per search and all this complex preprocessing allowed our clients to be at peace that an AI tool would not overwhelm their WHOIS servers (a common, less expected issue that appears at the end of development – when an AI tool of this kind is made public).
And how about premium names?
A good recommender tool must not overlook the domain name aftermarket. Premium names and marketplaces for reselling already registered names are an important part of the ecosystem. AI recommender that doesn’t work with these lists is incomplete and not ready for the market. The problem is that the way a “normal” domain name AI recommendations work is fundamentally incompatible with how premium names should be handled.
If you have a large list of available premium names, this tool must never recommend anything outside of this list. The magic cog in the machine that comes down to “Hey, ChatGPT – think of some great domain names for my user” doesn’t work anymore. You cannot pass a list of available names to the LLM because it surpasses input limits by many orders of magnitude (these lists are sometimes a couple of millions of names long). To make things harder, the fact that domain names are short strings makes them a really tough challenge for retrieval-augmented techniques.
Our engineers came up with a complex solution (they wrote an 11-page scientific paper to explain how it works) that is the complete opposite of the previous AI search. Generative components are put in a completely different part of the process. To save search time and reduce hallucinations, components late in the search process are replaced by another, non-generative machine learning module.
This resulted in premium name searches that are as fast (and just as important – as cheap) as regular AI searches, producing highly relevant tiered lists of names that require zero human input or classification to be added to the search pool.
Putting things into production
Closed beta testing of Name Sentinel started in late 2024, where we conducted user research to optimize the user experience, tweak the backend stuff, and work on the best ways to integrate it into vendor websites. With every round of testing, we saw a confirmation of this tool’s worth. Great general satisfaction scores, a measurable increase in satisfaction with domain name suggestions, and an increase in number of domain names registered per search.
The first public implementation of it became available as part of Domain.me website in February of 2025, with premium name search coming soon. The moment economists joined the meetings to talk about market strategy, positioning, and pricing, we knew the little “let’s wait for the right time” prototype grew into a full-fledged product.
Recognition and Results
Time is money — and our tool turns a rough idea into a concrete, available domain name ready for registration in just a few seconds.
This leap forward did not go unnoticed: our technology was recognized by one of the leading industry magazines. We were featured in Domain Name Wire in an article titled “New company aims to speed up AI-powered domain name search” (May 2025).
Even more importantly, direct use of the tool by one of our clients led to a 230% sales increase in the target demographic, a tangible proof of the business impact and conversion strength behind Name Sentinel.
Call it a good intuition, having the right people at the right place at the right time, or just sheer luck – the story of Name Sentinel is the story of how a decade of experience in the industry can turn into a cutting-edge product when the right time comes.
