Always-on optimization

Your page, getting better while you sleep.

Q: Why did you re-test an old variant?

That was a self-check. The loop periodically re-checks a prior result to catch winners that won for the wrong reasons — novelty, a one-time event, or an instrumentation glitch. If a result can’t hold up on fresh traffic, the loop reverts it and records the correction. It’s a built-in sanity check.

Q: Why did my loop pause?

Several rounds in a row came back inconclusive — none produced enough evidence to confidently promote a challenger. That usually means the page has no strong signal right now (already converting well, traffic too low for the effect size, or the selected element types aren’t where the lift lives). The loop pauses to flag this rather than burn credits. Resume any time from the loop detail page — nothing is lost.

Q: Why aren’t Images / Layout available yet?

Coming soon. Continuous Loops v1 ships with Headlines and Copy because that’s where AI-authored challengers are most reliable today and the lift-per-round signal is cleanest. Image and layout variants need a different generation pipeline (asset selection / regenerative DOM rewrites) — it’s next.

Q: How is this different from a multi-armed bandit?

Bandits dynamically shift traffic toward whichever arm looks like it’s winning. That maximizes short-term reward but sacrifices the clean causal claim — you can’t cleanly say "B beat A by X%" when the allocation kept changing while the result formed. Continuous Loops use a champion-challenger 50/50 split. When a round resolves you get a defensible "X beat Y by Z%, P > 0.95, CI excludes zero" claim you can put in a report, defend to a stakeholder, or feed into the next round’s prompt. We trade a little short-term efficiency for results you can trust and explain.

Q: What does it cost?

Creating a loop: 15 credits. Each round transition (resolution + next challenger authored): 10 credits. Enterprise: unmetered. Free plan: not available — any paid plan (Starter / Growth / Scale / Enterprise) can launch loops.

A Continuous Loop is an endless champion-challenger test. AI writes a challenger, the round resolves with a defensible Bayesian winner, the winner becomes the new champion, and the loop authors the next challenger — round after round, until you pause it.

Start a loop → See how it works

Bayesian, not guesswork Self-checks for false wins Launch once, walk away

app.variflask.com / loop

Pricing page · headline

Continuous Loop

Running · Round 7

🏆 Champion 3.6% CVR

“Ship A/B tests in minutes, not sprints”

⚡ Challenger 4.3% CVR

“Stop guessing. Let AI run your A/B tests.”

Probability to win

96.4%

95% CI on lift

+4.1% … +28.7%

Round complete · ready to promote

One launch. An endless cycle.

You pick a page and the element types to test (Headlines & Copy today; Images and Layout coming soon). From there the loop runs itself — five steps, repeating forever.

1
AI authors a challenger

A fresh, on-brand variant is written from your page context and what past rounds learned.
2
Clean 50/50 round

Champion vs challenger, split evenly — never a moving allocation that muddies the result.
3
Bayesian resolve

Promote, hold, or call it inconclusive — only when the evidence clears every gate.
4
Winner = new champion

The victor takes the throne. A loss is logged; the champion holds the line.
5
Repeat

The loop authors the next challenger and goes again — round after round, until you pause.

Step 5 loops back to step 1 — forever

Start improving

Point a loop at your page and let it run — round after round, each one earning its place.

Start improving →

Every round has to earn it

Bayesian

not guesswork

Full-cycle

rounds, not flukes

High bar

confidence to promote

Self-checking

catches false wins

How the loop protects you

The loop is deliberately conservative. Every round must clear all of these gates before a winner is promoted — and the loop watches itself for drift and burnout.

Runs a full cycle, not a fluke

Every round runs long enough to average out day-of-week swings and the short-lived novelty bump a fresh variant can get — so a result reflects steady-state behavior, not a honeymoon.

Enough data to be sure

A winner is only called once a round has collected enough traffic to give the result real statistical power, scaled to the page so low-traffic pages are not called early.

A high Bayesian confidence bar

A challenger is promoted only when the Bayesian model is highly confident it is genuinely better than the champion — not just numerically ahead on the day.

Real effect, not just probable

We look for evidence that the true effect is positive, not merely that the variant is likely ahead — so a win reflects a real improvement.

Self-checks for false wins

The loop periodically re-checks results to catch wins that came from novelty, seasonal blips, or measurement quirks — and unwinds a promotion that cannot hold up on fresh traffic.

Guards against noise over time

Across many rounds the loop is designed to keep the share of "winners" that are really just noise low — so compounding rounds do not compound false positives.

Pauses when there is no signal

If several rounds in a row come back inconclusive, the loop pauses itself and emails you — that is it telling you this page may have no strong signal to chase right now, so your credits are not quietly burned.

A defensible record of every round

Each round leaves a clean, explainable result — "X beat Y by Z%, P > 0.95, CI excludes zero." Put it in a report, defend it to a stakeholder, or feed it into the next challenger’s prompt.

Promotions, holds, and inconclusive rounds — all logged with the stats behind them.
Self-checks flagged inline, so reverted false wins are visible, not hidden.
Cumulative lift you can actually attribute — compounding, round over round.

Round history

8 rounds · +21.4% cumulative

Round 1 Promoted
+12.0% · P 0.97
Round 2 Held
champion defended
Round 3 Promoted
+6.8% · P 0.96
Round 4 Inconclusive
no signal in budget
Round 5 Promoted
+4.3% · P 0.95
Round 8 Self-check
champion confirmed

Why not just a bandit?

Bandits chase short-term reward at the cost of a clean answer. Continuous Loops trade a little efficiency for results you can trust and explain.

Multi-armed bandit

— Allocation shifts toward whatever looks like it’s winning.
— Maximizes short-term reward.
— No clean causal claim — the split kept moving as the result formed.
— Hard to report "B beat A by X%".

Champion-challenger loop

Clean 50/50 between exactly two variants.
A defensible "X beat Y by Z%, P > 0.95, CI excludes zero."
Self-checking — regression rounds catch false wins.
Results you can put in front of a stakeholder.

Frequently asked

Why didn’t my winner promote on day 3?

A round runs long enough to average out a few effects that make short reads unreliable: day-of-week seasonality (Tuesday traffic differs from Saturday), the novelty bump a fresh variant gets for a little while, and shifts in who is visiting in the first days of a test. Reading too early bakes those in and overstates the lift.

The exception: when a win is genuinely overwhelming, the loop can settle sooner — running longer would add cost without meaningfully changing the answer.

Why did you re-test an old variant?

That was a self-check. The loop periodically re-checks a prior result to catch winners that won for the wrong reasons — novelty, a one-time event, or an instrumentation glitch. If a result can’t hold up on fresh traffic, the loop reverts it and records the correction. It’s a built-in sanity check.

Why did my loop pause?

Several rounds in a row came back inconclusive — none produced enough evidence to confidently promote a challenger. That usually means the page has no strong signal right now (already converting well, traffic too low for the effect size, or the selected element types aren’t where the lift lives). The loop pauses to flag this rather than burn credits. Resume any time from the loop detail page — nothing is lost.

Why aren’t Images / Layout available yet?

Coming soon. Continuous Loops v1 ships with Headlines and Copy because that’s where AI-authored challengers are most reliable today and the lift-per-round signal is cleanest. Image and layout variants need a different generation pipeline (asset selection / regenerative DOM rewrites) — it’s next.

How is this different from a multi-armed bandit?

Bandits dynamically shift traffic toward whichever arm looks like it’s winning. That maximizes short-term reward but sacrifices the clean causal claim — you can’t cleanly say "B beat A by X%" when the allocation kept changing while the result formed.

Continuous Loops use a champion-challenger 50/50 split. When a round resolves you get a defensible "X beat Y by Z%, P > 0.95, CI excludes zero" claim you can put in a report, defend to a stakeholder, or feed into the next round’s prompt. We trade a little short-term efficiency for results you can trust and explain.

What does it cost?

Creating a loop: 15 credits. Each round transition (resolution + next challenger authored): 10 credits. Enterprise: unmetered. Free plan: not available — any paid plan (Starter / Growth / Scale / Enterprise) can launch loops.

Launch once. Let it compound.

Continuous Loops are available on every paid plan. Point one at your highest-traffic page and let the optimization run itself.

credits to create a loop

credits per round transition

∞

unmetered on Enterprise

View plans & start → Back to docs