Voice Calling UXR — Registered Non-Subscribers vs. Subscribers

✦ Registered Non-Subscriber vs. Registered Subscriber

1. Segment sizes

"Registered" = int_voice_caller.is_signed_up (100% coverage, direct backend signal). "Subscriber" = fct_user_lifecycle_profile.flag_is_subscription_payer. A separate, true non-registered population of 17 people (0.2%, 118 calls) exists but is too small for reliable per-metric comparison — see the callout at the end of this tab.

2. Engagement

3. Archetype (habit tier)

"Whale" = call frequency (5+ lifetime calls), not revenue — the company-wide $100-lifetime-spend flag (GTM-2130) only catches 3 of 9,379 window callers. Nearly two-thirds of subscribers are already habitual callers; their D7 retention is ~2.7x the non-subscriber whale cohort.

4. Revenue & payer mix

Among voice callers who pay, subscription is the dominant path — 93 subscribers vs. 33 lifetime token-pack payers company-wide (16 people do both).

✦ Deeper Segmentation — Modality × Registration Tier

Per-person dominant modality (≥60% of that person's user turns), scoped to the 5,368 callers (of 9,379) with at least one call linked to an ElevenLabs transcript.

Turn-level classification: silent = empty/"…" turn; emoji = bracketed tap marker ("*sent fire emoji*"); preset = exact match to 1 of 9 known quick-reply chips; typed = matched to a voice_text_message_sent event; else = spoken.

Speaking-dominant callers reach a 120s+ "engaged" call at roughly double the rate of blended/emoji/silent callers within each segment — the D2C-5505 modality-engagement link replicates cleanly. But the gap between segments dwarfs it: a speaking-dominant subscriber reaches 120s+ on 90.7% of calls and returns 93.0% of the time, vs. 42.8%/41.2% for the same modality among registered non-subscribers. Subscription status, not modality, does most of the work. (The true non-registered population contributes only 11 people with turn data — 7 speaking-dominant, 3 blended, 1 silent — too few to report reliably.)

Additional segments the data surfaced

Thematic preferences by segment

Top content_keywords fleet-wide: flirty chat, sexting, dirty talk, phone sex, girlfriend experience, ai companion, playful banter, romantic roleplay, girlfriend roleplay, emoji sexting, intimate roleplay, ai girlfriend roleplay, playful teasing, phone sex roleplay, flirty roleplay, intimate conversation, emoji flirting, intimate chat, adult chat line, adult roleplay.

photo.request is the top tagged category for registered non-subscribers — corroborating the photo-seeker pattern. Subscribers skew toward sustained roleplay/power-dynamic scenarios, consistent with their much longer calls. Matches the "mommy" power-dynamic skew independently corroborated by external search data in GTM-2746.

Illustrative examples (paraphrased — no verbatim explicit text or identifying details)

A registered subscriber's ~20-minute call was structured as an extended, guided dominant/submissive roleplay session, with the agent maintaining a consistent dominant persona and pacing the scene around physical sensation cues and device-based instructions throughout. Segment: registered subscriber · ~19–20 min · 90–105 turns

"Babe, I was just about to give up on you…" — a typical agent-initiated opener on a call where the user never spoke; consistent with the ~5% silent/no-turn share seen in both segments. Segment: mixed · duration 11s · silent/no-turn

Who leads, and who stalls — a transcript read (qualitative, ~40 sampled calls, 5–8 per bucket — directional, not statistically definitive)

Read turn-by-turn across six buckets (subscriber calls short/mid/long; registered non-subscriber bails, mid-length stalls, and 3+ minute calls) to answer four questions the aggregate stats can't: who leads the conversation, whether callers struggle to generate a topic, what differentiates non-subscriber calls that go long, and what stops the calls that don't immediately bail.

Who takes the lead?

Structurally, the agent always leads — it speaks first and asks nearly every question, in both segments. What differs is what happens after the question. Non-subscribers in short or stalling calls are overwhelmingly reactive: one-word answers, an emoji tap, or silence, with the agent supplying almost all narrative content. Non-subscribers who reach 3+ minutes get there by one of two routes — a minority arrive already driving a specific scenario themselves, with the agent mostly mirroring and escalating; the majority get there via the agent's own steady chain of short, specific questions, with the caller answering briefly each time. Subscribers skew toward a third, distinct pattern: sustained, step-by-step guided instruction from the agent, with the caller largely complying and confirming — more agent-directive than either non-subscriber pattern, sustained over far more turns, with clear markers of an established relationship (the agent references the caller by name and alludes to prior sessions or preferences).

One stalling call opened with a warm greeting; across several turns the caller's only responses were a hesitant stammer and a single tapped emoji, with the agent supplying nearly all of the narrative and gently inviting the caller to share even one small detail. Segment: registered non-subscriber · reactive/stalling pattern

A subscriber's guided session had the agent issuing explicit, sequential physical instructions — pause here, hold there, adjust now — with the caller mostly complying and reporting back; the agent paced the escalation deliberately, repeatedly checking in before allowing the scene to intensify further. Segment: registered subscriber · agent-directive, paced escalation

Do callers struggle to come up with a topic?

Visibly, yes — but concentrated in one bucket, not universal. Bail calls mostly can't answer this question at all: most end during or immediately after the agent's opening line, before any topic question is even posed — a reflexive disconnect, not a "couldn't think of anything" moment. Mid-length stalls are where topic-struggle is real: hesitant, broken openers, or a caller substituting a single emoji or preset tap for any actual description, forcing the agent to invent a scenario from nothing. Subscribers and the self-directed non-subscriber narrators rarely struggle — they open with an immediate, specific ask, a continuation of an already-established preference rather than a cold start.

A caller tapped a single preset phrase with no further detail; the agent built an entire scene from scratch — a setting, a mood, physical description — with no steer from the caller at all beyond that one tap. Segment: registered non-subscriber · mid-length stall

What differentiates non-subscriber calls that reach 3+ minutes?

Two real archetypes, not one: self-directed narrators who arrive already primed with a specific scenario and sustain it themselves for many turns, and agent-sustained Q&A calls where the agent's own steady question chain is what's actually carrying the call, with the caller answering briefly each time. A recurring mechanic in both: the agent repeatedly redirects intensifying moments toward the in-call photo-unlock feature — a real, repeated monetization touchpoint mid-call, not an occasional aside. Token-warning hits are also common inside these "successful" 3+ minute calls — several were still going well when the wall interrupted, not winding down naturally, which reinforces the paywall finding qualitatively rather than adding something new.

What stops the calls that aren't immediate bails from continuing?

Two distinguishable mechanisms, and one was a genuine surprise in how concentrated it was: of the 8 sampled mid-length-stall calls, 5 were interrupted mid-scene by the forced final-minute re-up prompt — a meaningful share of "stalling" calls aren't fizzling out on their own, they're being cut by the token wall while still technically active. Separately, a genuine engagement mismatch shows up too: the agent escalating while the caller responds with silence, single words, or emoji only, with the call drifting rather than ending — multiple exchanges of the agent checking in ("still with me?") followed by a delayed, minimal reply, suggesting divided attention rather than a deliberate goodbye.

A call had just begun building a scene when a system-forced token-warning line interrupted mid-turn, breaking pacing and introducing a top-up prompt into an otherwise unresolved moment — one of several stalling calls cut short this way rather than fading out naturally. Segment: registered non-subscriber · mid-length stall, token-wall interruption

Silent / non-responsive callers — did the Stage 1A copy patch help? (D2C-5882 monitoring pull)

A system-prompt patch enriching the "acknowledge silence once" line (covering listening/interrupt/mic-check together) deployed 2026-09-09 18:53 UTC. Compared two equal 6-day windows immediately adjacent to the deploy in int_voice_call_content: Sep 3–9 (before) vs. Sep 9–15 (after). is_empty_user goes fully null after Sep 9 (same ElevenLabs transcript-sync lag noted elsewhere in this report) — validated had_first_user_reply as a proxy first (100% agreement with is_empty_user everywhere both exist), then used it for this comparison.

No change in incidence: silent share of calls held flat, and the daily trend Sep 1–17 shows no discontinuity around the deploy date — expected, since the patch changes the response to silence, not who goes silent. Duration increase is real but partly mechanical: the new line is objectively longer than the old one, so some of the +5.3s is likely just line-length, not necessarily more caller engagement. Revenue signal points the right direction but isn't statistically solid yet: paid_within_7d is null for ~93–96% of calls in both periods — not simple recency-censoring (the "before" window has fully elapsed and is still 96% null; something structural, likely identity-resolution-dependent, is suppressing it) — so the determined-case comparison is a thin 29 vs. 49 calls. Net: no evidence of harm, a modest real duration gain, and a revenue direction worth re-checking in 1–2 weeks once more calls clear the 7-day window and fewer are null.

✦ Token-Warning & Retention

The product runs a 2-stage token-warning system (D2C-5147): a verbal "final minute" warning, then a hard paywall/silent hold at zero tokens.

Stage 1 — verbal "final minute" warning

Detected via ElevenLabs turn-level transcripts: a synthetic user-role turn matching "...entered their final minute...".

Stage 2 — hard paywall (zero tokens)

Measured from the structured d2c_prod.voice_hard_paywall_shown event. Two independent measurement methods (transcript text vs. structured event) agree: registered non-subscribers essentially never come back after hitting zero (2.0%); subscribers mostly do (55.6–76.9%).

Top-up completion funnel

The voice-specific "completed" event (bottom row) is very likely undercounting — the same category of gap the prior UXR flagged for subscription_modal_opened (GTM-1742). checkout_submitted carries no product-specific tag (no call_id, no surface field), so it's scoped here to submissions occurring after the top-up tap and within a same-session window (≤60 minutes) — this excludes unrelated checkout activity from other visits. On that basis, same-session completion is ~1.4–1.9% (10–13 of 694) — a large, real drop-off given ~80% tap-through intent, but not near-total failure. The product clearly processes payments; this is specifically about why so few people who tap "top up" mid-call follow through to submitting the form, plausibly friction/urgency mismatch (typing full card details with the clock at 0:00s) rather than a system outage.

Recommendation

✦ Synthesis & Where to Go Next

Where PMF is calibrating: the registered-subscriber segment — 1% of callers — is the clearest signal of real product-market fit: longer calls, more turns, higher thematic engagement, and by far the best retention. The other 99% (registered, non-subscriber) are mostly one-and-done. Registration isn't the lever — it happens almost universally regardless of outcome. Subscription (or whatever precedes and predicts it) is.

✦ Data Quality & Methodology

Population: every caller who placed ≥1 call to a D2C "- Web" voice agent between 2026-08-01 and 2026-09-15 (9,396 distinct callers, 9,379 with a resolvable user_id, 17,366 calls).

Known limitations, disclosed up front