An Agentic Campaign is one that learns and improves on its own, without requiring manual intervention after launch. You add your experiences, choose a strategy, and an AI agent handles the rest: who sees what, when they see it, and how traffic is allocated across every audience segment.
Instead of running one A/B test at a time and locking in a fixed winner, the agent continuously gathers evidence, updates its understanding of what works best for each audience, and shifts traffic accordingly.
The two things the agent does
1. Decides who sees what (Contextual Allocation)
The agent decides which experience each visitor gets based on their full behavioural context: intent level, device, channel, landing page, product affinity, shopper mindset, and more.
The key difference from standard A/B testing: there's no single winner. Different experiences can win for different audiences. A social proof message that works brilliantly for low intent mobile visitors might underperform for returning high intent desktop customers, and the agent learns this automatically.
Think of it as running hundreds of optimisations simultaneously, not one test with one winner.
2. Decides when to show it (One-Shot Optimisation)
For experiences that should fire once per visitor; email capture, discount overlays, exit-intent pop-ups, the agent doesn't just decide who sees it. It optimises when in their journey to show it.
The hypothesis is that the same intervention has very different effects depending on where a visitor is in their session.
Example (Browse Abandonment):
Imagine you launch a Browse Abandonment campaign with three experiences plus a "show nothing" baseline:
A - 10% discount overlay
B - Free shipping reassurance
C - Email capture with first-order offer
Baseline - show nothing
On day one, the agent splits traffic roughly evenly across all four to start gathering evidence. By day seven, it's learned that:
High intent visitors from paid social convert best on C
Desktop visitors who've viewed ≥3 products convert best on B
Low-intent discoverers on category pages are best left alone - Baseline wins
There's no single "winning experience" - each combination of audience and moment gets its own answer, and those answers update as new data arrives.
How it learns
When a campaign first launches, the agent starts by spreading traffic fairly across all your experiences. This is the Broad Exploration phase: gathering data, not yet making strong calls.
As patterns emerge, it shifts more traffic toward better-performing experiences while keeping some traffic on the alternatives to keep learning. This is the Narrowing In phase.
Once the agent is confident enough to direct most traffic to the best-performing experiences, the campaign moves into Optimisation. Eventually, when learning has stabilised across the full audience, it reaches Always On - the steady-state phase where the agent continues to monitor and re-score, but allocation patterns have largely settled.
The mechanism behind this is Bayesian statistics. Rather than working from a single estimate and a p-value (as a standard A/B test does), the agent maintains a full probability distribution over each variant's true performance rate — updated continuously as new data arrives.
Thompson Sampling uses those distributions to make allocation decisions: it draws a plausible rate from each variant's distribution and routes traffic toward whichever looks best in that draw. This means allocation shifts naturally as evidence accumulates, without a fixed end-date or manual decision point. A campaign that would take six weeks to reach significance in a fixed A/B test will typically route 80–90% of traffic to the winner within two to three weeks.
The agent also uses shared learning: insights from one visitor group can inform decisions in another, which helps it optimise faster when data is still sparse.
All of this happens automatically. You don't need to set thresholds, pick a winner, or make manual decisions. The agent re-scores every audience × experience combination once per night using the previous day's evidence. That's why allocations shift day-by-day rather than hour-by-hour, and why the fastest-learning campaigns tend to be ones with high-volume, clear-signal goals such as conversions or revenue per user.
The constraints the agent works within
The agent isn't running without boundaries. A handful of parameters are fixed at platform level, and a few others are defaults you can change. Together they shape what the agent is allowed to do, how quickly it can act, and how much evidence it needs before committing.
What you decide:
Experiences per campaign: up to 5, with at least 2 as the practical minimum.
Optimisation speed: how quickly the agent commits as it learns:
Conservative - learns cautiously.
Balanced - recommended for most campaigns.
Aggressive - for high-traffic, short-duration campaigns like a weekend flash sale. May commit on an early signal, so results arrive faster but may be less optimal.
Optimised control share: enabled by default. Starts at the midpoint of a 5–50% range and moves toward the minimum when winning and the maximum when losing -with certainty setting how far. You can disable optimised control or change the min/max bounds from agent customisation.
Custom prompt (optional): tailor the agent's reasoning with a short instruction (up to 1,000 characters). Influences what data the agent uses to define targeting and how it groups users into segments - doesn't change the matching criteria in Global Rules.
Guardrails: track up to 3 extra metrics alongside the goal. Guardrails appear in the report with the same detail as the primary goal but don't influence what the agent is optimising towards. Transactional goals (Revenue Per User, Conversion Rate) are automatically tracked as guardrails when you're optimising for something else.
What the agent builds for you:
Segments: up to 500 contextual combinations per campaign. This is the cardinality cap; the agent iteratively reduces its feature selection to fit under this limit so each segment stays sample-rich enough to train on.
Timing variations: up to 75 moment-based triggers per campaign, on strategies where timing optimisation is enabled.
What it needs before it can go live:
1,000 users matching the strategy's eligibility rules, plus 10 goal events within those users. Both thresholds must be cleared. The agent checks every four hours until they are. Until then the campaign is staged and ready but not firing.
How often it updates:
Nightly retraining on a rolling 90-day training window of behavioural data. That cadence is why you'll see allocations shift day-by-day rather than hour-by-hour, and why the learning window always reflects recent trading conditions rather than historical behaviour that's no longer representative.
The control variant
Every agentic campaign keeps a control group. These are visitors who see your default site experience, with no personalisation applied. This is how the campaign measures uplift.
By default, the size of the control group is managed automatically using Optimised Control, which adjusts the control share over time as confidence in results grows.
Control as an agent experience
You can also add your control as a variant within the agent experiences - there's a 'Show as Control' option when adding a variant. If none of your experiences are outperforming, the agent will shift traffic toward whichever option is performing least badly, including the control if you've added it as a variant.
What you'll see in the report
The Agentic Report gives you visibility into where the agent is in this process at two levels:
Campaign phase
New - campaign hasn't started learning yet.
Broad Exploration - traffic is roughly even across experiences while the agent gathers initial evidence.
Narrowing In - the agent has identified signal in some audiences and is shifting traffic accordingly, while still gathering evidence elsewhere.
Optimisation - the agent is confident enough to direct most traffic to the best-performing experiences. The report shows a rough estimate of how long until the campaign reaches Always On.
Always On - learning has stabilised. The agent continues to re-score nightly.
Performance signal
The Overview header shows the headline metric (e.g. Revenue Per User) against control, the relative uplift, the Probability to Beat Control, and a Certainty score reflecting how confident the model is in its current results given the volume and consistency of data collected.
The detailed Performance Breakdown shows total uplift and relative uplift with credible interval ranges, plus an Annual Projected Uplift figure. To keep projections credible, the annual projection is hidden until the selected analysis window is more likely than not to beat control.
A note on confidence: low traffic on a variant doesn't always mean it has lost. Because the agent uses shared learning across segments, a variant may have low traffic because the model has already learned, from similar segments, that it's unlikely to work for that audience.