How NeuralSplit Works

Overview

NeuralSplit is an AI-powered split testing engine that continuously improves your messaging using bandit algorithms. It doesn’t just test and report — it evolves your variants, picks winners, and generates new options, all in real time.

The Optimization Lifecycle

  • You enter your site and NeuralSplit will analyse it to understand your business
  • You create a campaign and add one or more elements (like headlines or CTAs).
  • Each element has multiple variants — manually entered or AI-generated, or both.
  • NeuralSplit distributes traffic using smart algorithms (UCB1, Thompson Sampling).
  • It learns which variant performs best, and shifts traffic accordingly.
  • Once a winner is found, losers are discarded and new variants are generated.
  • This cycle continues forever — constantly improving performance.

Creating Elements: AI, Manual, or Hybrid

NeuralSplit gives you total flexibility when creating content elements:

  • Leave only a name — the AI will generate variants from your campaign context.
  • Provide a few sample variants to guide the AI’s tone and direction.
  • Use a custom prompt to define style, persona, voice, or structure.
  • Combine any of the above for even more control.
  • Or skip AI entirely and set your own variants manually.

Integrating Into Your Site

NeuralSplit is designed to drop into your site or app with minimal effort. Just use the API or one of our lightweight SDKs to start serving and optimizing content instantly.

  • Generate your API key from the dashboard
  • Choose your language or framework
  • Copy the provided integration snippet into your app
  • Tell NeuralSplit what content to show — and when to track a conversion

More frameworks and client libraries are on the way.

WordPress example

<h1>[neuralsplit element="headline"]</h1>

<button onclick="trackConversion()">
  [neuralsplit element="cta"]
</button>

Next.js example

export default async function App() {
  const { headline, subheadline } = await getVariants({
    apiKey: "<your_api_key>",
    campaignId: "<your_campaign_id>",
    elements: ["headline", "subheadline"]
  });

  return (
    <div>
      <h1>{headline}</h1>

      {/* Example usage. Move this to a client component */}
      <button onClick={() => trackConversion({ reward: 1 })}>
        Sign up
      </button>
    </div>
  )
}

How NeuralSplit Decides Winners

NeuralSplit uses bandit algorithms like UCB1 and Thompson Sampling to distribute traffic intelligently. High-performing variants get more exposure, while low-performing ones are quietly phased out.

Once the model is confident enough, it selects a winner — then immediately generates new variants based on that winner to test again.

Why Multi-Armed Bandits > Traditional A/B

  • No fixed test durations — optimization happens in real time.
  • Less traffic wasted on poor-performing variants.
  • Faster time-to-results than traditional A/B testing.
  • Variants evolve automatically — no more one-off tests.

Manual vs. AI Control Modes

You decide how much control you want. NeuralSplit supports:

  • Manual mode: You write every variant by hand.
  • AI-assisted: Provide a few samples or a tone prompt to guide the AI.
  • Fully automatic: Let the AI generate and evolve variants continuously based on campaign context.

Switch between modes per element — you're in control.

Example: Real-World Optimization Flow

Let’s say you want to test a new headline. You create a “headline” element in NeuralSplit. You can:

  • Let AI generate 3 options based on your offer and audience.
  • Start running traffic to all 3 variants with just a few lines of code.
  • After 48 hours, one version performs clearly better.
  • NeuralSplit declares it the winner and generates 2 new variants inspired by it.
  • This loop continues — evolving the headline over time without your input.

Comparison: NeuralSplit vs Traditional Tools

FeatureTraditional A/BNeuralSplit
Setup TimeDays or weeks10 minutes
Variant CreationManual onlyManual or AI-generated
Traffic AllocationFixed ratiosDynamic, based on performance
Optimization SpeedSlowReal-time
EvolutionOne-off testsContinuous loop