Who captures the margin
in the attention economy

A large harvest narrows fast. Of 6,408 posts and 137,750 comments, 132 documents carry a monthly dollar figure. Every revenue conclusion on this page rests on that last box. A “document” here is one Reddit post or one comment (528 posts and 1,414 comments at the relevant stage) — not one person, and not one business.

The two actors who move the money — customers and chatters — are the two who barely speak. 88 chatter documents against 768 creator documents is the selection bias in one line.

Money enters at a $20 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. The corpus was extracted in stages, so every headline median was re-measured as the sample grew. Section 03 below tracks all 8 across those snapshots and computes what moved. Two never moved at all across every snapshot: platform take and hours per day. Still moving after the last batch: sub price, one-time price and creator revenue. That split — not the dollar amounts — is the finding.

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 1,794 relevant documents, 5 of 8 did not move at all over the final 686 documents. The intermediaries' cut is stable. The creator's income is the one number that never settled.

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.

Revenue mentions concentrate in the modes that require a human typing in real time. paid chat, PPV and personalized content together account for 937 document mentions — the DM window is where the money is made and where the hours go.

A separate lane, on almost no evidence. Repairing the price metric surfaced 8 claims where a brand pays the creator rather than a fan: median $450 per post, range $150–$2,000, against a fan-facing price of $50. That is roughly 9× per unit and it is the first sign in this corpus of an advertiser-side revenue model. n=8 is far too thin to plan on — it is recorded here because it exists, not because it is established.

Conversion rate is printed as UNUSABLE rather than as a median: the 43 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 10 hrs/day against a max of 60, which means per-day and per-shift claims were not normalised; only the median is usable.

This corpus is overwhelmingly human. 1,510 of 1,794 relevant documents describe human-operated monetization; 178 describe AI; 81 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.

Two different kinds of nothing. In the AI image-tooling communities we harvested 10,639 documents and not one that describes a brand paying a creator was discarded for vocabulary — those communities discuss compute costs, licences and hiring, not monetization. The operators are genuinely not there. In the UGC-ads communities it is the opposite: 19 documents clear the numeric bar and plainly describe a brand paying a creator, but are still dropped because they contain no topic term the filter recognises — they say brand, rate and budget in ordinary commercial English rather than the vocabulary this study was built on. That lane is unmeasured, which is not the same as empty.

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.

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.

Models with documents but no priced month: Findom (136) · financial domination (79) · livestreaming (55) · AI companion (35) · paid attention (26) · dating app (21). Monthly revenue for each: UNKNOWN — no claim exists to aggregate.

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.

Read these as ranks, not magnitudes: the extraction ontology under-counts unstructured venting. Two clusters matter more than any single label. Human cost — burnout 34 + stigma 14 + mental health 12 + harassment 6 = 66 mentions. Platform risk — platform fees 9 + platform bans 8 + platform risk 7 + account bans 7 = 31 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.

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