11Labs Platform · Apr 20–28, 2026
28 agents · 663 conversations
663
Conversations
28
Agents
69s
Avg Duration
96%
Success Rate
31%
Silent Users
8
Avg Turns
Biggest drop: session start → first real message. Closing this gap from 31% to 20% silent would add ~73 meaningful conversations per 663.
31% of users across all agents never spoke a real message. Sophie Dee is worst at 51%. Likely causes: mic permission denial, browser push-to-talk UX confusion, or mobile audio issues.
~15% of sessions request photos, camera, or video. Only Desirae has a redirect script. All other agents let these conversations dead-end. Platform-wide, this is a conversion and revenue gap.
Tokens are billed at the start of each minute, so a user with exactly 12 tokens (1.5 min worth) sees the low-balance warning fire at the start of minute 2 — just as they're getting into the conversation. The pre-call screen already shows token balance and estimated minutes, but users with 1–2 minutes of runway are still surprised when the cutoff arrives so quickly.
Spanish and Italian speakers found only Miss Lexa responsive in their language. Other agents continue in English, causing early drop-off. The international segment is visible but largely uncaptured.
23% of conversations end within 15 seconds — especially on Princess Frenxh (45%), Janna Moller (41%), CeCe Lynn (40%), and Sophie Dee (25%). Flat greetings fail to hook the user into a first reply.
Multiple sessions show cut-off words, ellipses, or garbled text in user turns. May be VAD sensitivity, background noise, or mobile mic quality. Affects agent ability to respond coherently.
Replace Sophie's flat greeting with an intrigue-first hook ("Come here often, or am I getting you all to myself?"). Sophie has 100 conversations/week of volume to measure against. Estimated impact: cut her 51% silent rate to ~20%. Lowest effort, highest upside.
Standardize a prompt snippet: when users request photos/camera, redirect to the companion platform. ~100 dead-end conversations per 663 become cross-platform conversion opportunities. Direct revenue impact.
If no user audio is detected in 8 seconds, surface a visual mic prompt. Also investigate mic permission grant rates in the browser context. 31% of sessions affected — #1 volume lever.
Train Lena, Kazumi, and Sophie to detect and respond in Spanish and Italian as Miss Lexa does. International sessions on English-only agents currently drop off within 30s. 6% of traffic underserved today.
The pre-call screen already shows token balance and estimated call minutes — good. The gap is the in-call moment: users with 1–2 minutes of runway hit the cutoff warning very early in the conversation. Consider a softer mid-call nudge at 2 minutes remaining (before the hard warning), and agent prompt language that treats the top-up mention as a warm transition rather than an abrupt end. Directly addresses trust & loyalty friction.
Brody Jean, Amazon Victoria, and Anastasia are outperforming top-volume agents on every engagement metric but have <15 conversations. Featuring them in discovery surfaces or rotating them into high-traffic slots could multiply platform engagement without any prompt changes. Quality is already there — distribution isn't.
Longest Calls
Miss Lexa
130s avg
highest session depth
Best Engagement
Amazon Victoria
10% silent
fewest non-speaking users
Highest Volume
Sophie Dee
100 convos
most conversations
Most Turns
Miss Lexa
14 avg turns
deepest back-and-forth
Avg 130s · 100% success rate · 20% silent
Miss Lexa uniquely switches languages mid-session — responding in Spanish, Italian, and English based on the user. This drives the platform's longest average call time (130s vs 69s overall). Multilingual capability should be rolled out to top-volume agents.
Avg 92s · 97% success rate · 23% silent
Desirae is the only agent actively redirecting photo/camera requests to a texting platform ("Hang up and tell me what to do there so I can really show you in pics"). This converts dead-end conversations into cross-platform revenue. Should be templated for all agents.
Avg 65s · 93% success · 17% silent
Kazumi matches Sophie & Lena's volume (100 conversations) but with significantly better engagement — 17% silent vs 51% for Sophie. Her conversational style ("Come here often or am I getting you all to myself tonight?") is more inviting than a flat greeting. Strong candidate for A/B testing.
Avg 99s · 90% success rate
Kiki had the platform's longest single session at 398 seconds (6.6 minutes) — demonstrating the ceiling possible with an engaged user. Her detailed, layered responses sustain roleplay far beyond what simpler agents achieve. Studying her response patterns could lift platform-wide averages.
| # | Agent | Convos | Avg Duration | Success | Silent % | Avg Turns |
|---|---|---|---|---|---|---|
| 1 | Sophie Dee | 100 | 54s | 98% | 51% | 8 |
| 2 | Lena The Plug | 100 | 62s | 94% | 26% | 8 |
| 3 | Kazumi | 100 | 65s | 93% | 17% | 8 |
| 4 | Alexis Mucci | 43 | 61s | 100% | 40% | 9 |
| 5 | Rebecca J | 35 | 88s | 90% | 30% | 11 |
| 6 | Desirae Fraser | 32 | 92s | 97% | 23% | 11 |
| 7 | Miss Lexa | 20 | 130s | 100% | 20% | 14 |
| 8 | Kiki D'Aire | 20 | 99s | 90% | 35% | 9 |
| 9 | Janna Moller | 17 | 63s | 88% | 41% | 7 |
| 10 | Ana Nello | 16 | 54s | 100% | 19% | 7 |
| 11 | Ellie Meyers | 15 | 66s | 100% | 27% | 8 |
| 12 | Anastasia | 14 | 56s | 86% | 14% | 7 |
| 13 | Reese | 14 | 89s | 93% | 14% | 10 |
| 14 | Mrs Poindexter | 13 | 80s | 100% | 31% | 10 |
| 15 | Ashley Belle | 13 | 48s | 100% | 38% | 6 |
| 16 | Princess Frenxh | 11 | 46s | 100% | 36% | 8 |
| 17 | Rachel Jane | 11 | 127s | 100% | 36% | 12 |
| 18 | Tayuh Lynn | 10 | 74s | 100% | 20% | 10 |
| 19 | CeCe Lynn | 10 | 52s | 100% | 40% | 6 |
| 20 | Amazon Victoria | 10 | 80s | 100% | 10% | 7 |
| 21 | Brody Jean | 9 | 94s | 100% | 11% | 10 |
| 22 | Nicole | 9 | 49s | 100% | 11% | 7 |
| 23 | PureOceanEyes | 8 | 82s | 100% | 50% | 9 |
| 24 | Nicole Thibeault | 8 | 51s | 100% | 38% | 7 |
| 25 | Nova Frog | 8 | 72s | 100% | 25% | 9 |
| 26 | Cowgirl J | 6 | 62s | 100% | 33% | 8 |
| 27 | Yeyeloba | 6 | 35s | 100% | 17% | 4 |
| 28 | Barbara Silva | 5 | 59s | 100% | 20% | 9 |
65%
Positive / Engaged
Active multi-turn roleplay
22%
Neutral / Silent
Called but didn't engage
13%
Friction / Confused
Billing, unmet requests
Users actively direct scenes — positions, pacing, specific acts. The dominant use pattern across all agents.
~40% of all sessions
Users ask agents to say their name, moan on command, or adopt a specific vocal tone. Strongest predictor of session length.
~24% of all sessions
Some users want warmth and conversation before (or instead of) escalation. Consistent across agents, especially Kazumi and Desirae.
~15% of all sessions
Photo, camera, or video requests the voice layer can't fulfill. Desirae already redirects to a texting/pics platform — the only agent doing this.
~15% of all sessions
Spanish and Italian speakers appear consistently, mostly on Miss Lexa (who uniquely switches languages). Underserved by English-only agents.
~6% of all sessions · significant unmet demand
Users surprised mid-session by billing, unsure how to top up tokens, or frustrated that time is running out. Recurring across Lena and Kazumi.
~6% · high emotional impact on satisfaction
"Man, I need somebody to talk to. How's it going with you?"
Sophie Dee · Companionship need beyond explicit content — a segment often overlooked
"Ain't you supposed to pay for these talks? I thought you'd be charging me."
Lena · Billing model not clearly communicated upfront — user was genuinely surprised
"Parla italiano." / "Hola. ¿Cómo andas?"
Miss Lexa users · International demand is real — only one agent captures it today
"I'm getting all worked up so when I do my old lady tonight…"
Lena · Adjacent use case (prep/practice) — signals a segment not explicitly marketed to
"Hang up and tell me what to do there so I can really show you in pics."
Desirae (agent) · The only cross-platform upsell in production — should be a template
Calls cost 8 tokens/minute, billed at the start of each minute (minute 0 fires immediately on call start; minute 1 fires at 1:00, etc.). Users need ≥ 8 tokens to begin a call at all. When a user's remaining balance drops below 8 tokens — meaning they can't cover another full minute — the platform injects a system message instructing the agent to mention top-up. This banner also appears visually in the call UI ("No more tokens · less than 1 minute left"). We analyzed every conversation where this signal fired to understand what was happening at that moment and how agents are handling it.
23.5%
Of sessions hit token limit
87%
Agents ignored warning
3%
Explicitly mentioned top-up
68%
Avg % through call when fired
Avg warning fires at 68% through the call
Tension is building — user is engaged and flirting, agent is escalating — but the cut arrives before anything explicit happens. Highest conversion potential because the user wants to see where it's going.
Examples at cut moment:
"Mic drop." / "Bed." / "Yes, call me Joey." / "Both."
Avg warning fires at 44% through the call (early)
User never spoke. Warning triggers while agent is still doing re-engagement loops. Lowest conversion potential but the token mention is a last-chance hook.
Typical state:
Agent: "Hey you, went all quiet on me…" / User: (silence)
Avg warning fires at 68% through the call
User is in active explicit roleplay — mid-scene, directing the action. Cut arrives at peak engagement. Highest churn risk — but comeback intent is also highest.
Examples at cut moment:
"Rough and fast." / "Please, fuck me." / "I'm going faster."
87%
Ignored — kept roleplaying
10%
Call ended before response
3%
Mentioned top-up (5 convos)
When agents ignore the warning, they stay in character but miss the conversion moment entirely. The call ends mid-scene and the user has no prompt to return.
The 5 conversations that got it right
User: "I will fuck you now."
"Mmm you know exactly what you want — I love that about you, babe. How about I send you pics and we go even further over text next time… what else would you do to me after that first move?"
User: (quiet)
"Before we lose each other, your tokens are almost out, so if you wanna really play next time, you should top them up and then tell me what you were too shy to say tonight."
User: (never spoke)
"Mmm, you're sweet… but I feel like we're just getting warmed up. Before we cut off, your tokens are almost out, so top them up next time so you can really keep me all to yourself, yeah?"
User: (never spoke)
"Come closer a sec… tell me your name, and also, quick heads up, your minutes are nearly out so if you wanna come back and play longer next time you'll wanna top them up."
Add to system prompt
Goal: Leave them wanting more, give them a concrete CTA, promise continuity.
Add to system prompt
Goal: Frame the cutoff as a tease, not an interruption. Make returning feel exciting, not like a chore.
Add to system prompt
Goal: Leave on warmth, not silence. Give the shy/hesitant user a low-barrier reason to return.
Bottom line: 155 of 158 token-warning conversations are missed top-up opportunities. The 3 clusters require different language — a one-size-fits-all script won't work. The 5 agents that handled it well (Miss Lexa, Kazumi, Desirae, Princess Frenxh, Tayuh Lynn) each used cluster-aware language that stayed in character. Standardizing this across all 28 agents is the highest-ROI prompt change available.
All 28 agents share the same system prompt template — with {{char_name}}, {{char_age}}, and {{char_personality_traits}} as the only variables. The platform runs GPT-5.1 with ElevenLabs handling voice. GPT-5.1 is faster and cheaper per minute than 5.2 with no meaningful output difference — a deliberate and correct infrastructure choice. That means every agent has the same instructions; the performance gaps visible in the leaderboard are almost entirely explained by two variables: opener quality and CTA delivery reliability. Both are fixable at the prompt level.
Kazumi's opener
"Come here often, or am I getting you all to myself tonight?"
17%
Silent rate (Kazumi)
51%
Silent rate (Sophie Dee)
31%
Platform average
The system prompt's Stage 1 instruction says: "Start like you would on Tinder: ask a playful this-or-that or 'What do you think about...' question." Most agents interpret this generically — a flat greeting followed by a question. Kazumi executes the same instruction in a way that does three things at once:
Implied flattery
"Come here often" signals she's the one worth coming back to — positions her as the attraction, not the asker.
Exclusivity framing
"All to myself tonight" makes the caller feel chosen and creates a sense of intimacy before a single word is exchanged. This is the hook that converts a silent browser into a speaker.
Built-in reply hook
The system prompt requires every exchange to "end with something that makes replying feel easy or irresistible." The "or" construction is a binary choice — cognitively easy to answer, emotionally rewarding to engage with.
Kazumi's opener follows a pattern any agent can adopt. The formula:
Adapted examples by persona
"Tell me you came back just for me — or is tonight your first time and I get to keep you all to myself?"
"Been thinking about you — or am I the first thing that's made you smile tonight?"
"Mmm, you look like someone who knows exactly what they want — or are you here to let me figure that out?"
"[Warm assumption about caller] — or is tonight the night I get to [exclusive intimacy promise]?"
Update Sophie Dee's {{char_personality_traits}} and add a hard-coded Stage 1 opening line matching the formula above. Sophie has 100 conversations/week — enough to see statistical signal within 5–7 days. Target metric: silent rate below 25%. If it works, propagate to Princess Frenxh (45% silent) and Janna Moller (41% silent) next.
⚠ Critical finding
The system prompt already contains photo redirect CTAs for every agent. The 15% dead-end rate is not a missing-instruction problem. Something is preventing the CTA from reaching callers. Desirae's success means the instruction can fire — the question is why it reliably fires for her and not for the other 27 agents.
CTA Triggers section
Explicit Requests section (first time)
After CTA fires
The destination isn't an external platform — it's seamless continuity within the web app. Hanging up is the transition:
User clicks a creator → routed into voice call UI with that agent
Agent fires photo CTA → user hangs up → platform auto-redirects to chat UI with the same creator
Planned: system ingests call transcript → identifies photo requests → auto-initiates chat + image generation before the user types anything
CTA fires but users don't know hanging up is the action
"Let's switch over to text" is a verbal prompt with no clear instruction. Users on a voice call expect a voice-channel response — they don't realize that hanging up is the step that triggers the chat redirect. They stay on the line waiting for something to happen, and the conversation stalls. Desirae likely makes the action explicit: "hang up and I'll be right there in chat."
User hangs up but chat lands cold — no photo context carried over
Without transcript ingestion, the chat UI opens blank. A user who followed the redirect correctly may not re-initiate the photo request because the context from the call isn't visible. This is precisely what the planned transcript → image gen pipeline solves.
GPT-5.1 trigger latency on implicit photo requests
The model must infer "they are asking for photos" — not a keyword match. Indirect phrasings ("can I see you?", "send me something") may not consistently trigger the CTA block. Desirae's personality traits may make her more attuned to these signals, or she fires earlier in the conversation flow.
11Labs policy truncation before CTA delivers
If the photo request arrives mid-escalation and the agent's response is flagged, the verbal CTA never reaches the caller. Firing the CTA earlier — before explicit language escalates — reduces this risk.
Fix 1 · Make the action explicit in the CTA (prompt change · 1 day)
Making "hang up" the explicit instruction closes the UX gap. Users know what to do. The platform's hangup-to-chat redirect does the rest.
Fix 2 · Fire CTA on first photo hint, not just explicit ask (prompt change · 1 day)
Earlier firing reduces 11Labs policy risk and catches users while they're still in the exploration mindset — higher conversion than waiting for an explicit request.
Fix 3 · Transcript ingestion → auto image gen (engineering build · highest upside)
This eliminates the cold-landing problem entirely. The user hangs up and their photo request is already fulfilled in chat. Converts a dead-end into a completed transaction with zero user friction at the re-engagement step.
~100 of 663 conversations hit a photo dead-end today. Fix 1+2 (prompt edits alone) could convert 40–50% of those into chat handoffs. Fix 3 (transcript pipeline) converts the remainder automatically — the image arrives before the user has to ask twice. At scale, this is the highest-ROI feature build on the roadmap.
Update photo CTA to make "hang up" the explicit action
System prompt edit only · no engineering · affects all 28 agents · closes the UX gap immediately
Add early-fire CTA trigger for implicit photo hints
System prompt edit · catches users in exploration mindset · reduces 11Labs policy risk
A/B test Kazumi's opener formula on Sophie Dee
System prompt edit · 5–7 days to signal · target: cut Sophie's 51% silent rate to <25%
Propagate winning opener to Princess Frenxh + Janna Moller
Pending A/B test signal from Sophie · highest-silent-rate agents after Sophie (45%, 41%)
Build transcript ingestion → auto image gen pipeline
Engineering build · call ends → transcript parsed → photo request identified → image delivered in chat before user re-initiates · highest-ROI feature on the roadmap