Exploratory first pass · Reddit-extracted claims · 2026-09-06

Who captures the margin
in the attention economy

A structured read of 6,408 Reddit posts and 137,750 comments on how people monetize attention, companionship, sugar dating, creator content, and AI personas. The finding is not a market size. The finding is the shape of the take-rate stack — and how thin the evidence underneath it actually is.

01

What the evidence base actually is

A large harvest narrows fast. Of 6,408 posts, 86 documents carry a monthly dollar figure. Every revenue conclusion on this page rests on that last box.

Evidence attrition from Reddit corpus to revenue claims A four-stage chain showing 6,408 harvested posts narrowing to 1,525 extracted documents, then 1,108 judged relevant, then 86 that carry a monthly revenue figure. −76% −27% −92% HARVEST 6,408 posts + 137,750 cmts 40 communities w/ signal EXTRACTED 1,525 documents, LLM-parsed RELEVANT 1,108 1,722 claims · 1,557 numeric $ FIGURE 86 docs w/ a monthly $ claim LEGEND CORPUS STAGE THE BASIS OF EVERY $ CLAIM ON THIS PAGE
The whole revenue story is carried by 86 self-reported documents.

Attrition percentages are computed from the counts shown. The largest contributing communities are CreatorsAdvice (163) · Twitch (120) · onlyfansadvice (119) · CamGirlProblems (85) · paypigsupportgroup (84).

Creators posting

585

of n=1,108 relevant docs

Observers

230

n=1,108 docs

Customers posting

198

they supply the money

Agencies posting

34

n=1,108 docs

Tool builders

32

n=1,108 docs

Chatters posting

13

they supply the labor

The two actors who move the money — customers and chatters — are the two who barely speak. 13 chatter documents against 585 creator documents is the selection bias in one line.

02

The take-rate stack

Money enters at a $15 subscription and is cut on the way down. The platform takes a median 20%. The agency takes a median 45% on top of that. The creator keeps the residual — and the creator's residual is the only number in the stack that is wildly unstable. The intermediaries have tight margins; the operator absorbs all the variance.

This was tested, not asserted. Extraction ran in stages, which made an accidental robustness check across four snapshots (n=397 → 738 → 840 → 1,108 relevant documents). Seven of eight headline figures held to within 2% across all four: agency 45%, platform 20%, chatter $4/hr, $15/mo, $29.49 one-time (the widest drift, $29.99 → $29.49), 10% conversion, 8.0 hrs/day. The eighth — creator revenue — swung across a $500 band, $3,500 → $4,000 → $3,750 → $3,500. The AI share held near 14% at every snapshot. An asymmetry that survives a tripling of the sample is a finding, not an artefact.

The take-rate stack A five-layer stack from customer to chatter showing each actor's median reported value, full reported range, and dispersion measured as the ratio of the 75th to the 25th percentile, where the creator's spread is the widest in the chain. ACTOR IN THE CHAIN MEDIAN REPORTED FULL REPORTED RANGE DISPERSION · p75 ÷ p25 · SHARED SCALE 1×–10× MONEY L1 Customer pays the subscription $15 /mo · n=66 min $3 · p25 $9.90 · p75 $32.12 · max $399 one-time price median $29.49 (n=114) 3.2× L2 Platform takes a cut of gross 20% take · n=12 min 2% · p25 20% · p75 47.5% · max 80% the high tail is live-gifting, not subs 2.4× L3 Agency takes a cut of what is left 45% take · n=12 min 5% · p25 30% · p75 50% · max 70% stacks on top of the platform cut 1.7× L4 Creator keeps the residual · 8 hrs/day (n=25) $3,500 /mo · n=117 min $5 · p25 $1,000 · p75 $10K · max $250K survivorship uncorrected · outlier months 10.0× L5 Chatter does the conversation labor $4 /hr · n=11 min $2 · p25 $3 · p75 $4 · max $7 a wage, not a share 1.3× 1× 6× 10× the tightest margin in the chain LEGEND FOCAL LAYER DISPERSION BAR — SHARED 1×–10× SCALE DISPERSION = p75 ÷ p25 WITHIN EACH ACTOR'S OWN REPORTED UNIT ($/MO, %, $/HR). IT MEASURES SPREAD, NOT SIZE — THE BARS ARE COMPARABLE, THE UNITS ARE NOT. THE CUTS DO NOT COMPOSE INTO A VERIFIED WATERFALL: EACH LAYER IS A SEPARATE SELF-REPORTED SAMPLE, NOT FOUR MEASUREMENTS OF ONE DOLLAR. THE CREATOR'S SPREAD IS 6.0× THE AGENCY'S. THE OPERATOR ABSORBS THE VARIANCE; THE INTERMEDIARY DOES NOT.
The creator's income varies by 10.0× between the 25th and 75th percentile. The agency's take varies by 1.7×. That asymmetry — not the dollar amounts — is the finding.

The agency is the stable actor

n=12

Median take 45%, interquartile range 30%–50%, max 70%. Whatever the creator earns, the agency's percentage barely moves. It carries no platform-ban risk and holds none of the identity.

The creator is the volatile actor

n=117

p25 $1,000/mo against p75 $10,000/mo, max $250,000, min $5. Median hours worked: 8/day (n=25). The residual absorbs all the variance in the system.

The chatter is paid a wage

n=11

Median $4.00/hr, range $2.00–$7.00. The person doing the conversation that closes the sale holds no equity in it, and is the least-represented voice in the corpus (13 documents).

03

Does the finding survive more evidence?

The corpus was extracted in stages, and each stage overwrote the last — which accidentally produced a robustness test. Tracking the eight headline medians across four snapshots as the sample grew from 259 to 1,108 relevant documents, 6 of 8 did not move at all over the final 268 documents. The intermediaries' cut is stable. The creator's income is the one number that never settled.

Headline metrics across 4 extraction snapshots 8 small charts, one per headline metric, tracking its median across 4 extraction snapshots as the corpus grew from 259 to 1,108 relevant documents. 6 of 8 did not move over the final step; 2 were identical at every snapshot; creator revenue never settled. EACH PANEL HAS ITS OWN Y-SCALE · x = 259 → 397 → 840 → 1,108 RELEVANT DOCS 6 of 8 metrics did not move over the final 268 documents agency take n=9 → n=12 HELD 45% platform take n=4 → n=12 HELD 20% chatter pay n=9 → n=11 HELD $4.00 hours per day n=9 → n=25 HELD 8 sub price n=43 → n=66 HELD $15 one-time price n=51 → n=114 MOVED $29.49 conversion rate n=16 → n=32 HELD 10% creator revenue n=64 → n=117 MOVED $3,500 LEGEND THE ONE THAT NEVER SETTLED SETTLED BY THE FINAL SNAPSHOT MOVED OVER THE FINAL STEP: ONE-TIME PRICE, CREATOR REVENUE. A FLAT DASHED LINE MEANS THE MEDIAN WAS IDENTICAL AT ALL FOUR SNAPSHOTS. THE PIPELINE OVERWRITES summary.json IN PLACE, SO THE FIRST THREE SNAPSHOTS ARE THE BUILD-SESSION RECORD. THE FINAL POINT IS READ LIVE FROM DISK AT BUILD TIME.
Platform take and hours per day were identical at every snapshot. Creator revenue went $3,250 → $4,000 → $4,000 → $3,500 and is still moving.

This is a stability check, not a confidence interval. A median that does not move as n grows is better evidenced than one that does — but it is still a median of self-reported claims, and a stable wrong number stays wrong. The early snapshots also show real movement: agency take read 50% at n=259 before settling at 45%, and sub price dipped to $12.50 at n=397 before returning to $15.

04

Two findings that changed direction

Not everything converged. Two results reversed as the sample grew, and both are shown here rather than smoothed over — because each is a case where a single snapshot would have confidently reported the wrong answer.

Two findings that changed direction as the sample grew Left: who initiates a gift, where the leading mode swapped between snapshots and ends in a three-way spread. Right: customer-stated motivations, where companionship rose from fifth to first as customer coverage quadrupled. A · WHO INITIATES THE PAYMENT — THE LEAD SWAPPED, THEN SPREAD n=41 docs offered 13 requested 12 n=59 docs offered 17 requested 19 n=170 docs offered 54 requested 44 agreed service 39 no dominant mode — a single snapshot would have picked a winner B · CUSTOMER-STATED MOTIVATION — COMPANIONSHIP OVERTOOK ATTENTION companionship 98 attention 68 fantasy 61 sexual interest 58 validation 49 companionship ranked fifth at n=35 customer docs, first at n=194 LEGEND PANEL A: THE FIRST TWO SNAPSHOTS ARE THE BUILD-SESSION RECORD; THE THIRD IS THE FINAL summary.json. PANEL B IS THE FINAL SAMPLE ONLY. BOTH ARE COUNTS OF DOCUMENTS, NOT OF PEOPLE OR PAYMENTS. NEITHER CARRIES A DENOMINATOR, SO NO RATE CAN BE COMPUTED FROM EITHER.
Who starts a gift swapped lead between snapshots and ends in a three-way spread. Companionship rose from fifth to first among stated customer motivations.

Gifting initiation at the full sample: spontaneously offered 54, explicitly requested 44, agreed service 39 (n=170 documents describing a transfer). No mode holds a majority. Customer-role coverage nearly quadrupled over the run (198 documents at the end), so the motivation ranking is far better evidenced now than at any earlier snapshot — which is exactly why it changed.

05

The labor sink is the chat layer

Revenue mentions concentrate in the modes that require a human typing in real time. paid chat, PPV and personalized content together account for 526 document mentions — the DM window is where the money is made and where the hours go.

Revenue types by document mention A bar chart of how many documents mention each revenue type, where the chat-driven modes together dominate the named categories. 0 50 100 150 200 250 300 350 400 DOCUMENTS MENTIONING THIS REVENUE TYPE (n=1,108 RELEVANT DOCS) 384 subscription 193 paid chat 170 tips 169 personalized content 164 PPV 164 direct payments 338 other residual bucket THE CHAT LAYER · 526 LEGEND FOCAL NAMED REVENUE TYPE UNCLASSIFIED RESIDUAL — NOT A CATEGORY COUNTS ARE DOCUMENT MENTIONS, NOT DOLLARS, NOT USERS, NOT MARKET SIZE. ONE DOCUMENT CAN MENTION SEVERAL TYPES.
Chat is what consumes humans — and the layer AI can most plausibly take. It is also, per all three council seats, the layer whose full automation nobody has yet shown converting.

Creator hours

8

hrs/day median · n=25

Chatter pay

$4.00

/hr median · n=11

Hourly revenue

$8.85

median · n=10

Conversion rate

UNUSABLE

n=32 · p25 5% p75 55% · mixed denominators

Customer one-time spend

$391.11

median · n=18

Tip amount

$20

median · n=25 · max $2,000

Conversion rate is printed as UNUSABLE rather than as a median: the 32 reported values do not share a denominator (follower→sub, sub→PPV, and DM→purchase are all present), so no funnel arithmetic can be built on them. Hours worked has the same defect at the top end — p75 is 12 hrs/day against a max of 60, which means per-day and per-shift claims were not normalised; only the median is usable.

06

The AI-persona reality check

This corpus is overwhelmingly human. 941 of 1,108 relevant documents describe human-operated monetization; 100 describe AI; 52 hybrid. Three independent council models — Gemini 3.1 Pro, Grok 4.6, and Kimi — reviewed the same evidence separately and all three warned against reading an AI-companion market out of it. This is a negative finding, and it is one of the more valuable results here.

Human versus AI operation across relevant documents A proportional bar of 1,108 relevant documents split between human-operated, hybrid, AI-operated and unclear, showing that AI is a small minority of the corpus. WHO OPERATES THE PERSONA — n=1,108 RELEVANT DOCUMENTS, PROPORTIONAL WIDTHS AI · 100 Human-operated 941 documents · 84.9% HYBRID · 52 UNCLEAR · 15 LEGEND AI IS 9.0% OF THIS CORPUS. THE CRAWL RETURNED AN ONLYFANS / FANSLY / FANVUE CORPUS — NOT A COMPANION-APP ONE. NO GMV CONCLUSION ABOUT AI COMPANIONS CAN BE DRAWN HERE.
The hype-to-evidence gap is the point: 100 documents is not a market.
“This dataset is fundamentally an OnlyFans/Human dataset, not an AI one. … Any AI economic conclusions drawn here are dangerous extrapolations from human adult entertainment.”
Council seat · Gemini 3.1 Pro (High)
“AI-companion frame is not in the data. … Do not generalize to Character.AI / companion-app GMV.”
Council seat · Grok 4.6
“Demand is for a person, not content. … The market has spoken within this data and AI personas are a rounding error.”
Council seat · Kimi

What the tools data says

AI companion and chat-automation products are named in the corpus. But they appear as tools people use, never as businesses with a stated P&L. Hybrid human-plus-AI operation is 52 documents: the automation exists and has not displaced the labor.

Most-named: Replika (23, AI companion chatbot with messaging) · Linktree (9, link aggregation) · Canva (6, Free tool for creating profile image collages) · Discord (6, group communication) · Kindroid (6, Customizable AI companion with roleplay and villanous character options) · OBS (5, Free recording and streaming)

Council seats were run independently against an earlier snapshot of this corpus; their qualitative verdicts are quoted, their counts are not — every count on this page comes from the current extraction. A fourth seat (gpt-5.5) failed to execute and returned no output; it is excluded rather than counted.

07

A named trap: the models nobody prices

Some of the most-discussed models in this corpus produced zero monthly-revenue claims between them. All three council seats independently flagged findom as the clearest case. Plotting discussion volume against revenue evidence puts every volume-without-evidence model on one line.

Discussion volume against revenue evidence A scatter plot of business models where the horizontal axis is how many documents discuss the model and the vertical axis is how many carry a monthly revenue figure; the most-discussed speculative models sit on the zero line. ↑ DOCUMENTS CARRYING A MONTHLY REVENUE FIGURE (n) 0 4 8 12 16 20 0 22 44 66 88 DOCUMENTS DISCUSSING THE BUSINESS MODEL (n) → ×2 FINDOM · 69 · 0 LIVESTREAMING · 51 · 0 CAMMING · 49 ZERO REVENUE EVIDENCE volume is not evidence LEGEND BUSINESS MODEL (≥5 DOCS) THE TRAP ×N MARKS A COORDINATE SHARED BY N MODELS — THE LABELS DIFFER, THE EVIDENCE POSITION IS IDENTICAL. ACROSS ALL MODELS WITH ≥4 DOCS, 37 MODELS AND 410 DOCUMENTS PRODUCED ZERO MONTHLY-REVENUE CLAIMS. LARGEST: FINDOM (69) · LIVESTREAMING (51) · FINANCIAL DOMINATION (32) · AI COMPANION (28).
These generate the engagement of a business and the P&L of a support group. 37 models, 410 documents, zero priced months.
“It generates massive Reddit engagement driven by fantasy, shame, and complainers, but lacks the verifiable, recurring revenue reality of the standard subscription/PPV models.”
Council seat · Gemini 3.1 Pro (High) · on findom
“Findom volume ≠ findom business. … That is a harm/support corpus, not an operator P&L.”
Council seat · Grok 4.6 · on findom

Models with documents but no priced month: Findom (69) · livestreaming (51) · financial domination (32) · AI companion (28) · dating app (20) · sex work (18). Monthly revenue for each: UNKNOWN — no claim exists to aggregate.

08

Where the opportunity actually points

Not at operating personas. At the infrastructure that undercuts the 45% agency take — because the agency's product is DM handling and PPV pitching, and that is the layer with a named automation path. 12 candidate opportunities were scored and 3 were classified as traps. The scores below are analyst judgments, not measurements — which is why every one ships with a falsifier.

Opportunity upside against risk A quadrant plot of 12 candidate opportunities scored for mean upside and mean risk, with the recommended build highest on upside and the opportunities classified as traps sitting at high risk. ↑ MEAN UPSIDE (ANALYST-ASSIGNED, 1–10) 3.5 6.0 8.6 MEDIAN RISK MEDIAN UPSIDE 4.0 5.0 6.1 7.1 8.2 MEAN RISK (ANALYST-ASSIGNED, 1–10) → 01 02 03 04 05 06 07 08 09 10 11 12 every score here is a judgment, not a measurement LEGEND RECOMMENDED BUILD CANDIDATE CLASSIFIED AS A TRAP UPSIDE AND RISK ARE MEANS OF ANALYST-ASSIGNED 1–10 SUB-SCORES. THEY ENCODE JUDGMENT ABOUT THE EVIDENCE — THEY ARE NOT MEASURED FROM IT.
Creator DM/CRM automation (undercut the agency take) sits at the top of the upside axis at moderate risk. The traps cluster right of the median-risk line — and they are among the narratively hottest ideas in the category.
IDOpportunityUpsideRiskVerdictFalsifier — what would kill it
OPP-01 Creator DM/CRM automation (undercut the agency take) 8.10 6.12 Build If creators using AI chat show materially WORSE retention or higher refund rates than human chatters, the thesis dies.
OPP-10 Disclosed AI companion app 7.40 6.88 Candidate Measure paid conversion and 90-day retention in a disclosed companion app. The corpus cannot answer it.
OPP-02 Ban-resistant funnel + link routing 7.00 6.62 Candidate If ban rates are concentrated in a small non-compliant minority rather than broad-based, the market is far smaller than it looks.
OPP-04 Personalized-content production tooling 6.80 6.38 Candidate If customers rate AI-made customs materially lower, or refund them more, the value proposition inverts.
OPP-03 Creator analytics / pricing benchmarks 6.50 4.38 Candidate If creators will not pay for benchmarks in a simple pre-sale test, this is a feature and not a company.
OPP-11 Content-theft / leak monitoring 6.30 5.00 Candidate If takedown does not measurably restore revenue, it is insurance nobody renews.
OPP-06 Operate AI personas directly 6.10 7.75 Trap Find a disclosed AI persona with verified, durable paying retention. Absent that, this is a hypothesis, not an opportunity.
OPP-08 Audience-acquisition service (Reddit-first) 6.10 6.50 Candidate If cost per acquired subscriber exceeds creator LTV at median $14.99/mo pricing, the unit economics never close.
OPP-09 Payments / chargeback + fraud tooling 5.80 6.62 Candidate If chargeback rates are actually low and absorbed by platforms, there is no budget line to sell into.
OPP-12 Sugar-dating matching platform 5.30 7.75 Trap Show verified completed, repeat arrangements rather than first-contact scam reports.
OPP-05 Chatter marketplace / training + QA 4.90 5.75 Candidate If AI chat reaches parity on conversion, this labour market shrinks rather than grows.
OPP-07 Findom / paypig monetization 4.00 6.75 Trap Show recurring, verified monthly revenue from more than a handful of operators. The corpus produced none.

Burnout

26

creator pain mentions

Mental Health

12

creator pain mentions

Stigma

12

creator pain mentions

Platform Fees

8

creator pain mentions

Platform Risk

7

creator pain mentions

Low Conversion

7

creator pain mentions

Read these as ranks, not magnitudes: the extraction ontology under-counts unstructured venting. Two clusters matter more than any single label. Human cost — burnout 26 + mental health 12 + stigma 12 + harassment 5 = 55 mentions. Platform risk — platform fees 8 + platform risk 7 + platform bans 6 + account bans 6 = 27 mentions. At the full sample the human-cost cluster is the larger of the two; at earlier snapshots the platform cluster led. Both are real; only one of them can be solved with software.

The human cost is now the largest pain cluster

55 mentions

burnout 26 + mental health 12 + stigma 12 + harassment 5. This did not surface until the full sample — it is the one finding that grew rather than settled, and it is the part of the problem an infrastructure play does not touch.

It also sharpens the trap in section 07: scaling the chat layer without addressing what the chat layer costs the person doing it is how the labour supply fails.

09

What this is not

Six reasons a reader should not treat anything above as a market estimate. These are structural and uncorrected — not caveats added for form.

01 · Self-reported

No figure was verified against a payout statement, a platform dashboard, or a bank record. Transparency labels (GREEN 789 · YELLOW 237 · RED 82) are the extractor's judgment of how forthcoming a post was — not verification. A self-report can be fluent and false.

02 · Survivorship

People post outlier months, not median months. Of 1,108 relevant documents, only 86 carry a monthly figure at all — and the ones who post are disproportionately the ones with something to post.

03 · Selection

Subreddit participation is not market participation. The crawl is Reddit, so Reddit showing up as a leading named acquisition channel (223 docs) is partly circular. Instagram and TikTok are almost certainly undercounted.

04 · Incentive

Agencies (34 docs) and tool builders (32 docs) recruit in these threads. Revenue claims skew upward by construction. Meanwhile chatters (13 docs) and customers (198 docs) — the people who supply the labor and the money — are the quietest voices.

05 · Volume is not size

Every mention count on this page counts documents. Not dollars, not users, not transactions. A model discussed twenty times is a model that is discussed twenty times.

06 · Small n, mixed units

Sample sizes run from single digits to the low hundreds and are printed beside every figure. Units were not fully normalised — hours reported per day and per shift are conflated, and conversion rates share no denominator. Where a value could not be established honestly, this page prints UNKNOWN or UNUSABLE rather than an estimate.