The Nudge: AI-Driven A/B Testing Flywheel

Exec Briefing | March 2026

The Opportunity

Three conversion moments drive all revenue. Each is underoptimised:

A 20% improvement at each stage compounds to ~70% revenue uplift with no additional traffic spend.

Recommended Stack

Keep what’s there. Add one layer.

Omniconvert handles what to show. PostHog answers why users behave that way. Claude closes the loop.

The Flywheel

PostHog detects significance threshold met
        ↓
Claude reads results + session replay summaries + learnings library
        ↓
Writes conclusion: what moved, why, confidence level
        ↓
Generates 3 ranked hypotheses for next test
        ↓
Posts to Slack: "Test concluded. Proposed next experiments ↓"
        ↓
Editor approves (one click) → Omniconvert experiment auto-created
        ↓
Repeat

Human input: ~10 minutes per week to approve hypotheses. Everything else is autonomous.

Experiment Roadmap

Prioritised by revenue leverage, not technical complexity.

Priority 1 — Trial → Paid Conversion

Highest leverage. This is where money is lost.

Test 1.1 — Pricing transparency on CTA

• Hypothesis: Hiding price (£4.99/mo) until mid-signup causes drop-off surprise
• Variant: Add “Then £4.99/month, cancel anytime” under “Start Free Trial” button
• Expected: +15–25% trial completion rate
• Effort: 30 minutes in Omniconvert

Test 1.2 — Trial length

• Hypothesis: 7 days isn’t long enough to experience the best perks (events are weekly)
• Variant A: 14-day trial | Variant B: £1 first month
• Note: £1 trial often beats free — signals intent, reduces tyre-kickers
• Expected: +20–35% trial→paid (monitor churn at day 60)
• Effort: Low (Stripe config + Omniconvert)

Test 1.3 — Specificity of benefit bullets

• Current: “Discounts at 100+ restaurants and venues”
• Variant: “Average 40% off at restaurants like [3 recognisable names]”
• Expected: +15–30% click-through to trial
• Effort: Copy change only

Test 1.4 — Urgency via member events

• Hypothesis: Abstract perks don’t convert; concrete upcoming events do
• Variant: Replace static benefit list with “Next member event: [real upcoming event] — 48hr early access for members”
• Dynamic, updated weekly via Omniconvert personalisation
• Expected: +20–40% on trial starts
• Effort: Medium (requires content automation)

Priority 2 — Referral

Word-of-mouth is the natural distribution channel for a brand built on insider knowledge. The audience self-identifies as taste-makers — that’s the asset to activate. In an era of declining SEO discovery, referral is the most defensible acquisition channel. Rewardful is already installed.

The mechanic to test first: double-sided referral

• Referrer gets 1 month free. Referred friend gets 14-day trial (vs. standard 7)
• Both sides win → share rate 2–3x single-sided programs
• Rewardful handles tracking; Omniconvert handles the in-product prompts

Test R.1 — Trigger timing

• Hypothesis: Most referral prompts fire too early (at signup), before the user has experienced value
• Variant A: Prompt after first perk redeemed (“You just saved £X — know someone who’d love this?“)
• Variant B: Prompt at day 10 of trial (post-value, pre-renewal decision)
• Control: Prompt at trial signup
• Expected: +40–80% share rate vs. control — timing is the biggest lever in referral
• Effort: Medium (Omniconvert trigger + Rewardful link injection)

Test R.2 — Reward framing

• Current default assumption: free month
• Variant A: “Free month” (monetary framing)
• Variant B: “Unlock a secret member dinner for a friend” (experience framing)
• Hypothesis: Experience rewards resonate more with this audience than cash-equivalent discounts — identity consistency with the brand
• Expected: +20–35% on referral conversion rate
• Effort: Low (copy + reward config)

Test R.3 — WhatsApp-first share flow

• Hypothesis: London young professionals share via WhatsApp, not email
• Current: Generic share link
• Variant: Primary CTA = “Send via WhatsApp” with pre-written message (“I’ve been using this for London restaurants — you’d love it, here’s 2 weeks free”)
• Pre-written copy removes the blank-page friction that kills shares
• Expected: +50–100% share completion rate on mobile
• Effort: Low (Omniconvert overlay, WhatsApp share API)

Test R.4 — Gifting flow

• Hypothesis: “Give a friend a trial” converts better than “refer a friend” — gifting frame removes self-interest perception
• Variant: Seasonal or occasion-based (“Give someone the best of London this month”)
• Particularly relevant around Valentine’s Day, birthdays, “new to London” moments
• Expected: Incremental acquisition channel, hard to benchmark — treat as new channel test
• Effort: Medium (requires gift redemption flow)

Test R.5 — Email forward optimisation

• The weekly newsletter already reaches 500k+. Most referral programs ignore this channel.
• Add a single line at the bottom of every newsletter: “Forward this to someone who needs better London plans →”
• Variant: Include a dedicated friend-referral link (tracked via Rewardful) vs. plain forward
• Expected: 1–3% forward rate on 500k list = 5–15k new exposures per send, compounding weekly
• Effort: Very low (template footer change)

Referral flywheel note for the AI layer: Referral experiments have a longer feedback loop than on-site CRO (need to track referred-friend trial → paid conversion, not just share clicks). The AI agent should flag referral tests as requiring 30-day minimum runtime and weight conclusions against downstream paid conversion, not just share rate — a reward that drives shares but attracts low-intent users is a negative outcome.

Priority 3 — Visitor → Email Subscriber

Volume play. More emails = more trial opportunities.

Test 3.1 — CTA copy (highest ROI test in this category)

• Current: “Sign Up”
• Variants: “Get This Week’s Edit” / “I Want In” / “Send Me The Good Stuff”
• First-person and specificity consistently outperform generic signup copy
• Expected: +80–150% (this category has the widest variance — big wins available)
• Effort: 20 minutes

Test 3.2 — Exit-intent overlay

• Trigger: User about to leave without subscribing
• Offer: “Before you go — get London’s best 3 things this week”
• Single field (email only)
• Expected: Capture 2–5% of otherwise-lost visitors
• Effort: Low (Omniconvert overlay, already capable)

Test 3.3 — Inline vs. modal signup

• Hypothesis: Modal interrupts discovery; inline at article end captures high-intent readers
• Test placement: end of every article vs. current modal timing
• Expected: +30–60% on article-sourced signups

Priority 4 — Email → Trial

The leakiest pipe. Most subscribers never start a trial.

Test 4.1 — Subject line personalisation

• Test: Generic (“This week’s edit”) vs. location-personalised (“What’s on in Soho this week”) vs. curiosity gap (“You haven’t been here yet”)
• Expected: +10–20% open rate; benchmark for this audience should be 35–45%
• Effort: Low (email platform A/B, not Omniconvert)

Test 4.2 — Email CTA framing

• Current: Content-forward (article links)
• Variant: One email per month with primary CTA = trial start, framed around a specific upcoming member event
• Expected: +25–50% trial starts from email channel
• Effort: Editorial + template change

Test 4.3 — Re-engagement sequence for non-openers

• Segment: Subscribers who haven’t opened in 60 days
• Trigger: Automated 3-email sequence: “Are we still right for you?” + best content + specific perk offer
• Expected: 15–25% reactivation rate
• Effort: Medium (automation setup)

Priority 5 — Retention (longer horizon)

Test 5.1 — Onboarding personalisation

• Ask one question at trial start: “What are you most interested in?” (restaurants / events / both)
• Personalise first 3 emails to that preference
• Expected: +10–20% trial→paid, +15% 90-day retention
• Evidence: Personalised onboarding is the single highest-impact retention lever for subscription products

Test 5.2 — Perk notification timing

• Hypothesis: Members who use a perk in first 14 days retain at 2x the rate
• Variant: Proactive “Here’s a perk you can use this weekend” at day 3 of trial
• Expected: +20–30% trial→paid
• Effort: Medium

AI Experiment Prioritisation Logic

The AI layer ranks hypotheses using:

1. Funnel value — which stage has the most £ at stake
2. Historical signal — what themes have worked before (learnings library)
3. Session replay signal — where users are visibly confused or dropping
4. Effort score — copy changes before layout changes before technical changes
5. Time since last test — avoid testing fatigue on same page

It never runs two tests affecting the same conversion moment simultaneously. Referral tests are flagged for 30-day minimum runtime.

90-Day Build Plan

Expected Outcome

Conservative case (20% improvement at each stage): ~70% revenue uplift from existing traffic. Optimistic case (best-in-class execution): 2–3x.

The compounding effect is the point — each concluded experiment makes the next hypothesis smarter.