The Decision
A platform processing millions of search moments daily can’t afford to average them.
Identity Digital operates one of the world’s largest domain registries — a business where search and discovery are the revenue engine. Their team understood something most registrars hadn’t fully acted on: the moment a first-choice domain is unavailable isn’t a failure state. It’s a decision point. Handle it well, and a user stays, explores, and converts on an alternative. Handle it poorly, and they’re gone in under ten seconds — likely to a competitor with a more intelligent recovery experience.
At Identity Digital’s scale, even a fractional improvement in that recovery rate carries significant commercial weight. The business case wasn’t hard to build. The harder question was the right one: not whether to invest in smarter recommendations, but what “smarter” actually required — and whether the default instinct (more capable model, more compute) was actually the answer.
The Problem, Reframed
The existing system was doing exactly what it was designed to do. That was the problem.
The legacy recommendation engine wasn’t broken in any obvious way. It performed competent keyword matching — surfacing domains that shared strings with what the user typed. The issue was conceptual, not technical: domain search is not a string-matching problem. It is an intent-interpretation problem.
A user searching for “bluemountaincoffee.com” might care most about the brand feeling, the category signal, the geographic anchor, or simply securing a credible .com at a reasonable price. Any combination of those things could produce a correct recommendation — but a system that only mirrors what was typed will always produce recommendations that feel mechanical. Because they are. It was solving the wrong problem, efficiently.
“We weren’t losing customers because our search was slow or broken. We were losing them because it was literal when it needed to be interpretive.”
Identity DigitalProduct Leadership [Placeholder — replace with attributed quote]
Three failure modes compounded each other. Keyword-only matching produced irrelevant alternatives, accelerating drop-off. Drop-off happened precisely at the highest-intent moment in the funnel — a user who had already decided to search for a specific domain is more committed than a casual browser. And the conventional fix — heavier models, higher accuracy — introduced latency and compute costs that threatened to make the economics unworkable at scale.
The Approach
The instinct was to reach for the most powerful model. Apexon’s team resisted it.
Most AI engagements default to the same pattern: select the most capable available model, build a wrapper, measure improvement. The problem with this pattern isn’t that it fails — it’s that it leaves significant value unrealised and builds in cost and latency risks that surface later, usually at the worst possible time.
Apexon started with a different question: what does a correct recommendation actually need to understand — and what’s the minimum reliable cost to understand it? This led to a heterogeneous architecture rather than a single-model approach. The right model for each specific job in the pipeline, selected through structured benchmarking across accuracy, latency, and cost — not defaulted to because it was the most capable option available.
The core of the system combined Retrieval-Augmented Generation, high-dimensional semantic embeddings, and agentic routing logic. Rather than passing all search intent through one model, the system routes different intent types to different interpretive strategies. Someone searching for a personal brand is reasoned about differently than someone searching for a product category or a geo-anchored business. Relevance improved not by adding more compute, but by adding more precision about what kind of understanding each query actually required.
The solution was deployed entirely within Identity Digital’s existing AWS infrastructure — no new services, no additional overhead, no security boundaries to negotiate. A globally distributed team delivered it in six months.
What Changed
Sessions that used to end in abandonment now extend into exploration.
The most telling shift wasn’t a conversion number — it was a behavioural one. Users who previously hit an unavailability wall and left were now staying, engaging with alternatives, discovering options they hadn’t considered. The unavailability moment stopped functioning as a dead end and started functioning as a discovery gateway.
Conversion increased. Session engagement deepened. Cost-to-serve improved — a direct outcome of the benchmarked model architecture rather than reaching for the most expensive option. And the platform now has compounding returns: every interaction refines the next recommendation. This wasn’t delivered as a finished product — it was architected as a foundation for continuous iteration, which is the only honest way to build AI systems in a landscape that doesn’t hold still.
Outcomes
Conversion ↑
Significant uplift at the unavailability moment — the highest-intent point in the funnel, previously treated as a write-off
Engagement ↑
Higher average session time; abandonment replaced by active discovery behaviour
Cost-to-serve ↓
Benchmarked model selection outperformed default-to-most-capable on both accuracy and economics
6 months
Full delivery, globally distributed team, zero additional infrastructure beyond existing AWS footprint
Architecture decisions
RAG over fine-tuning. Domain naming conventions shift constantly. RAG reasons over a live knowledge base; a fine-tuned model is a snapshot that decays.
Agentic routing over single-model. Personal brand intent, product category intent, and geo-anchor intent require different reasoning — not averaging through one model.
Benchmarked heterogeneity over best-available. Structured testing across accuracy, latency, and cost revealed which tasks needed capability and which didn’t. The outcome: better performance, lower spend.
Why these choices, not the obvious ones
Fine-tuning
Why Retrieval-Augmented Generation instead of fine-tuning?
A fine-tuned model encodes a moment in time. Domain naming conventions evolve continuously — new TLDs launch, naming trends shift with culture and market, creative patterns emerge and fade. A RAG system reasons over a continuously updated knowledge base, which means recommendations stay relevant without retraining cycles. In a domain this dynamic, recency isn’t a nice-to-have.
Single-model inference
Why route intent to different reasoning strategies rather than average through one model?
Domain search intent is not uniform. Someone securing a personal brand name is optimising for memorability and uniqueness. Someone searching for a product category domain is optimising for discoverability and convention. Someone looking for a local business domain has a geographic anchor that matters. Averaging these through one model produces recommendations that are plausible for no one. Routing intent to the appropriate reasoning strategy produces recommendations that feel correct for the specific person asking.
Default to most capable
Why benchmark across models rather than use the most powerful one available?
The instinct to use the most capable model is understandable. It feels like a defensible decision. But capability has a cost — in latency, in compute spend, in architectural complexity — and that cost compounds at the scale Identity Digital operates at. Structured benchmarking revealed a clear pattern: for certain bounded, well-defined tasks in the pipeline, smaller models matched or exceeded the performance of larger ones at a fraction of the cost and latency. The result was a heterogeneous architecture that outperforms a single-model approach on every relevant dimension.
The broader implication
Domain registrars are converging on AI-assisted search. The question is whether you build it thoughtfully or reactively.
The technology to do this isn’t scarce. The rigour to design it correctly — economics-first, intent-precise, architecturally honest about what each task actually requires — is rarer than it should be. That rigour is the difference between a recommendation engine and a conversion engine. Between a feature shipped and a capability compounded.
