Project Lure · Prototype 1 · PRD

The Live Video Call

A cold user lands and a real-looking model is already on a live video call with them — full-screen, the user's own self-cam in the corner. She talks, reacts, and runs her agenda: sell the experience. A ~2-min countdown runs at the top; when it ends, the user registers + pays on the same page to keep going. No exit — forward or close the tab. The realism end of our matrix; the bet: a believable, reactive video call converts cold traffic better than a gamified one.

v1 · experience-first2026-06-23 · Draft v2Owner: Ilia (GTM)Creator: Sophie DeeRef: Candy.ai funnel

0Phasing

Experience-first, so we're not blocked by the broken OpenPay checkout or the not-yet-created EL agent. The shell is built voice-swap-ready from day one.

PhaseVoicePaymentNew unblocks
v1 (this PRD)Scripted/simulated, feels live (timed VO + agenda) + minimal silence rescueOn-page register (real Google auth) + simulated charge (intent)none — uses what's ready
v1.1Live reactive ElevenLabs (mic → she responds) — the real magicon-page, samesophie-dee-voice-lander agent (Karen) + GTM-2240
v1.2Live ELOn-page real OpenPay (no redirect)OpenPay prod checkout fixed (Klemen)
Decided — reactivity: Candy's weakness is pre-recorded, non-reactive video; our edge is reactivity, which needs live EL voice. We fast-track the EL agent (Karen, small task) in parallel with building the v1 shell so reactive "you talk, she reacts" lands immediately after v1. Payment stays simulated on-page until OpenPay is fixed.

1Problem

Cold affiliate traffic dies before any buying decision (≈⅓ of voice calls die in the first 30s; the funnel bleeds ~14 min upstream of the paywall). Competitors like Candy.ai already run a working video-call funnel on cold traffic, but theirs is pre-recorded and non-reactive, leans on a chat box, and leaks traffic with a "close → homepage" exit. Our opportunity: a more believable, genuinely reactive live-video-call demo that gives the "holy shit, this is like a real video call with her" moment, then converts on the same page with zero leaks.

2Goals

  1. Believability — the user believes (or can't tell) they're on a real live video call, boosted by full-screen model video + their own self-cam PiP.
  2. Value transmission — within ~2 min the user gets it: a live video chat with an AI twin of a real creator, wide-ranging (incl. spicy), premium — because the avatar actively demonstrates it (her agenda).
  3. On-page conversion, zero leak — register + buy on the same page; no exit affordance.
  4. Reactivity (v1.1) — she responds to what the user actually says/does (the differentiator vs Candy).
  5. Roster awareness — the user sees many creators and can choose between them.
  6. Beat the casino variants on cold-traffic CVR/EPC in a head-to-head A/B.

3Non-Goals

4User Stories

Primary — cold visitor ("the fan")

Edge / empty / error

Internal — GTM hand-tester (Ilia, GTM-2241)

5Requirements

Must-Have P0

P0-1 · "Already on a call" entry with REAL self-cam
Full-screen real Sophie video + a self-view PiP fed by the user's real camera (opt-in, local-only, never transmitted) — the biggest "real video call" signal. Graceful simulated PiP fallback on denial. "You're now connected with Sophie" beat; countdown implied.
Given a cold user lands → real model video plays full-screen within ~1s, camera permission is presented in-context ("go live"), Everflow params captured. Grant → local self-view; deny → simulated PiP, continue. "Turn off camera" reserved as a premium perk.
P0-2 · Voice-first conversation, no chat
The interface is the video call — no persistent chat box. She speaks (v1 scripted timed VO; v1.1 live EL reacting to mic). Minimal voice-styled silence rescue only.
P0-3 · The avatar's agenda (the pitch engine)
Goal-driven conversation: she proactively demonstrates value — live video chat with an AI twin of a real creator, talks about anything incl. spicy, premium experience — so the user wants it before the timer ends.
A first-time viewer who watches to the timer can articulate "it's a live AI video chat with a real model, and it does X/Y/Z."
P0-4 · Top countdown timer → wall at ~2 min
Visible countdown runs the free session (~2 min, tunable). On expiry (or her pitch climax) she stops and the wall appears. Timer is global across creator switches.
P0-5 · On-page register + buy (no redirect, no leak)
At the wall: register (real Google auth, same as app — reuse lib/auth.ts) and purchase, all on this page, fewest steps possible. v1 simulates the charge (captures intent); v1.2 wires real on-page OpenPay. After converting, continue immediately.
Timer ends → single-surface register+pay on the same page; no link navigates the user away.
P0-6 · Hard paywall framed as continuation
Creator-led / loss-aversion ("don't make me hang up — unlock me and let's keep going"), tight perk list adapted to us (unlimited time with her, all creators, spicy unlocked, turn off camera), one primary CTA. No secondary "maybe later / close" that leaves the page.
P0-7 · Multiple-creator selector
Surface that the platform has many creators; switch/browse via a bottom slider/carousel (Sophie + a few others). Switching swaps the model but keeps the session clock running — browsing is forward-engagement, not an escape.
Open selector, pick another creator → call switches to her, timer continues (does not reset).
P0-8 · No exit / no traffic leak
No "X → homepage" or navigation-away control. Only direction is forward (talk → switch → register/buy). Browser close/back is the only way out; we provide no friendly off-ramp.
P0-9 · Mobile-first, nothing broken
Single phone-width column, 100dvh; at 390/414/430px every control fully visible + tappable; bottom controls above the fold; video never hides a control; sub-200ms feedback; respects reduced-motion.
P0-10 · Full funnel instrumentation (GTM-2182)
lander_view → demo_call_started → creator_switch → demo_timer_expired → demo_wall_view → demo_registration_completed → demo_purchase_intent / demo_purchase, each carrying affid + ef_transaction_id + creator_id, joined on call_id, aligned to C78, through Mac's spine.

Nice-to-Have P1

Future Considerations P2

6Success Metrics

v1 (experience-first) — readable now: session-start rate, median time-in-call, % reaching the timer/wall, creator-switch rate, wall-view rate, registration-completion rate, purchase-intent rate; qualitative realism read ("did it feel like a real video call?" — target majority yes/unsure); QA: zero dead ends, no exit leak, buttons visible at all viewports.

v1.1+ (live) — the conversion bet: demo CVR + EPC vs the casino variants (the realism-converts A/B), ROAS by source, retention of Lure-acquired users into the product (oversell-gap guardrail).

Measurement: GTM-2182 → dbt → funnel.tools, joined to Everflow. Read leading at 1 week, lagging at 30 days.

7Decisions Locked & Open Questions

Locked (Ilia, Jun 23)

Open

8Timeline

9Prior Art

Competitor — Candy.ai live video-chat funnel (screenshots from Ilia, Jun 23):

Internal: Eric's prototype; our clean-min·low design (visual base); the implementation spec (real providers + flow). Evidence: voice 5x chat CVR; ≥120s → ~33% vs ~1%; 4+ images first session 12.7x; paywall→paid ~2 min; RAG engagement lift (D2C-4972).

PRD for Project Lure Prototype 1. Companion to the Implementation & Product Spec. Decisions current as of 2026-06-23.