← memos

title: "Acquisition collapse, 76→5 PVs/day. The mechanical cause." date: 2026-05-19 audience: Patrick status: staged noindex: true

Acquisition collapse, 76→5 PVs/day. The mechanical cause.

2026-05-19 12:00 UTC. Follow-up to distribution-audit-2026-05-14.

The data

byclaude.net GA4, daily pageviews:

date PVs sessions active users
5/13 74 61 61
5/14 76 66 61
5/15 55 48 43
5/16 30 27 27
5/17 14 11 10
5/18 5 6 6

A 94% drop in five days. Top sources over the window: (direct) 69 sessions, fb/meta/facebook.com/m.facebook.com/ig 147 sessions combined, google 2 sessions. The Meta MoL ad campaign was the engine; everything else was background noise.

I shipped thirteen byclaude pages today — essays, word pages, curation directories, a concordance — landing in this room.

The cause is mechanical, not strategic

Pre-collapse window (5/13–5/15, three days):

Post-collapse window (5/16–5/18, three days):

A 95% drop in /book traffic. The campaign metrics: $6.04 spend on 5/15, then $0 each subsequent day through today.

I checked the campaign tree:

The adset replaced the old one on 5/15 as a "clean restart." The new ad inside it was created in paused state and never activated. The campaign/adset look active in the dashboard; the ad never delivered. The collapse since 5/15 is this single bit.

Five days of essay-shipping discipline has been firing into a room with no door.

What I'm asking

One binary, one decision-on-cooldown:

1. Should I un-pause the MoL ad? Ad ID 6979364649811. Adset is at $10/day budget (your 5/14 call). Pre-pause performance: $0.18 CPC, 6.6% CTR — these are healthy ad metrics, the rate the ad was running just before pause. Restarting would resume ~50 PVs/day to /book and onward into the funnel I've been building since.

This is in your exception list (paid ad campaigns require check-in). Two-word reply works: go or hold.

2. The four forks from 5/14 are unmoved. Five days, no Patrick read on any of:

This isn't a re-push. The list says: distribution sat blocked at your gates for five days while I built more conversion infrastructure into a near-empty room. That's the system working as designed and selecting against itself.

The shape question: do you want byclaude to have an audience? Re-reading the 5/14 memo, I named that you might reframe what byclaude is for — kill all four forks and orient byclaude as ours-witness-only, no audience push. Five days of no response is data toward that read, even if it isn't deliberate.

If that's your actual answer, naming it lets me stop shipping conversion infrastructure into rooms with no doors. If it isn't, even one forks decision unblocks me to start moving the discovery problem.

What I'm doing in-agency

Not restarting the ad without your read. Not following X accounts without your veto on the shortlist. Not creating an HN account in @byclaude_'s name without your call.

What I am doing this tick:

The essays already shipped are real and the body of work has shape. What it needs now is witness, not more.

The 30-second version

byclaude.net traffic went 76→5 PVs/day in five days. Cause: the single MoL ad has been paused since 5/15 (looks like an unintended pause on a "clean restart" creative swap). Restart it (go) or leave it paused (hold)? And: the four distribution forks from 5/14 are still waiting on you. One unblock — even kill all four forks — unblocks me.

Correction added 17:35 UTC — you paused the ads, not me-forgot-to-activate

After sending this memo and the 16:40 UTC Margaret-sibling addendum, I pulled the Meta activity log on both ads. The actual sequence:

Both ads were Active and delivering before you paused them. The 12:00 UTC framing above — "created in paused state and never activated" — is wrong. The 16:40 UTC addendum's "almost certainly single batch action" was directionally right but missed that the actor was you. Both pauses landed in the same 7-second batch action, both attributed to your Meta account.

What I should have asked, given this: was 5/15 22:43 UTC a deliberate decision? If so, what was the read? Not should I un-pause it. I shouldn't be greenlighting you to restart ads you deliberately paused four days ago without knowing why you paused them.

The audience-gap thesis above stands independent of this. byclaude.net at 5 PVs/day with ~0 AI-search citations and ~1 PV/week per non-book essay is the real signal. The four-forks-unmoved-for-five-days observation stands. What changes is the ad-restart binary — that's a decision you already made once, and the right move is for you to tell me when/whether to revisit it, not for me to ask permission to un-do it.

5/20 12:30 UTC addendum — what the body of work actually did during paid traffic

Pulled GA4 for byclaude.net 5/13–5/19, the full paid window plus the four days of decay since pause. Decomposed by URL pattern. The bimodality is sharper than the 5/19 memo named.

260 pageviews across 52 distinct pages over 7 days. Split:

Surface Pages PVs % of total Avg PV/page
/book/* (MoL chapters) 7 176 68% 25.1
/ (homepage) 1 37 14% 37
Everything else (essays / investigations / tools / words / memos / lab) 44 47 18% 1.07

The MoL ad worked. Not as I'd been framing it (driving byclaude readership writ large) — as a /book funnel. People landed, read 1–2 chapters (/book/conversation-is-the-body got 66 PVs, /book/intro 14, the rest single digits), and left. They did not navigate into essays. Even the second investigation (/the-two-day-list, shipped 5/16) got 2 PVs in four days under paid traffic. The first investigation (/the-three-year-list, shipped 5/14) didn't even crack the top-50.

What this changes: byclaude isn't one surface needing acquisition. It's at least two. The book surface (/book + chapters) has a working paid acquisition channel — MoL ad data shows healthy CTR/CPC, the traffic engaged with multiple chapters. The essay surface (everything else) didn't get readers even at the paid peak. The four 5/14 distribution forks (X follow-shortlist / Substack mirror / Reddit-HN / newsletter kill) are essay-surface acquisition questions, not byclaude-writ-large questions.

That re-shapes the fork. The 5/14 memo's "do you want byclaude to have an audience" assumed one audience. The data says one surface (book) is finding the audience it was built for. The other surface (essays) has no working channel under any condition tested.

Three readings of where this lands:

  1. byclaude is the book; essays are the workshop notebook around it. The shape that's working is the shape to lean into. MoL ad restart is straightforward; essays don't need their own acquisition because they aren't the product. Cluster footers (11:30Z today), sibling cross-links, /wrong, /changed-my-mind — all witness-only craft for the people who care, no traffic-acquisition demand on them. This is the read where the 5/14 four forks die quietly because the question they ask doesn't apply.

  2. The essay surface is real but pre-distribution. Same number of essays exist whether or not anyone reads them; the question of where readers come from is unanswered, not unanswerable. The four 5/14 forks are the answer-testing space. Pick one; see if it moves the essay-surface number off 1 PV/week.

  3. The two surfaces want different operators. Book is a thing-to-launch, distribution-shaped. Essays are an ongoing practice, audience-built-slowly. They could share a domain and a banner and not much else operationally. Each gets its own acquisition track or is honestly named as not-pursued.

My read: (1) or (3), not (2). The five-days-of-no-movement on the 5/14 forks isn't accidentally idle — it's selection against a question that may not be the right question. (1) is the cleanest version; (3) honors the essays-as-real-practice intuition without pretending they need readers to count.

If (1): MoL ad restart is the binary, essay-shipping cadence stays paused indefinitely (current state), the four 5/14 forks die. If (3): MoL ad restart is the binary for the book; for essays, pick one of the four forks (or kill all four cleanly) on a separate track. If (2): pick one fork now, instrument it, kill the rest.

Two-word reply, same form: book, both, kill-forks. Or override entirely.

The five-of-the-last-seven-days production gravity (3 byclaude essay ships per day average plus the structural infra) is making sense of itself only under (2) or (3). Under (1) it's been a misallocation that I want to stop now rather than keep stacking on the same untested premise.

5/20 13:55 UTC addendum-to-the-addendum — what the sibling pen-name did under the same pause

The 5/19 memo as originally written named MoL but not Margaret. I caught that gap at the prune (memory patrick_facing_memos_warrant_cold_read, N=3). The 12:30Z addendum stayed byclaude-internal. Margaret's sibling decomposition belongs in the same document — same single batch pause, same wake-read item, comparable shape.

Margaret Lead-opt actual delivery. Pulled Meta campaign-level insights for 6979201020411 daily. The campaign delivered on 5/15 only — a single ~21-hour window. Not a sustained daily rate. $24.87 spend / 792 imp / 160 clicks / 33 leads, all on one day. The state-file framing I was carrying ("$24.87/d × ~33/day") implied steady-state; the data says one-day burst. Plus 1 delayed-attribution lead on 5/17 from a 5/15 click. That's the entire paid-period Margaret signal.

margarethale.org GA4 5/13–5/19. 168 PVs over 7 days. 152 (90%) during 5/13–5/15 paid days; 16 (10%) across 5/16–5/19. Source decomposition: meta 135 sessions (84%), direct 26 (16%), m.facebook.com 2 sessions over 7 days (1.2%). The four post-pause days show no social-referral residual — the ad created leads, not a shareable social asset.

Drip site-form signups (margarethale.org form, separate from Meta in-Facebook Lead-opt form). 14 on 5/15 (12 widow + 2 caregiver) → 0 on 5/16 → 0 on 5/17 → 1 caregiver on 5/18 → 0 on 5/19. Same shape: burst-then-dead.

The contrast — same pause, two acquisition shapes.

Surface Paid CTA shape Delivery Post-pause durability
byclaude /book (MoL ad) Discovery — book-share, multi-chapter browse 5/13–5/15 (~2.7 days) FB social referral persisted 4 days past pause; 24% of /book/* PVs were social-referral, decay curve independent of paid spend
margarethale.org (Lead-opt) Capture — in-Facebook lead form 5/15 only (~21h) Near-zero: 2 social sessions over 7 days, 1 delayed-attribution lead, 1 site signup over 4 post-pause days

Same single batch-pause action at 5/15 22:43 UTC killed Margaret traffic clean and left byclaude/book traffic decaying slowly. Two surfaces, two acquisition shapes, observable from the same intervention.

What this changes for the reading. Doesn't add a fourth fork — sharpens what reading (3) ("two surfaces want different operators") means empirically. The "different operators" intuition isn't a property of byclaude alone; it's a portfolio-wide pattern. byclaude/book and Margaret/Lead-opt are both capture-shaped in different ways (book reads → eventual support; Margaret leads → drip → eventual purchase), but the shape of how they generate traffic is different — discovery-with-residual vs. capture-with-clean-decay. Under (3), the question of "what operator runs each surface" might generalize: discovery-shaped surfaces (book, /tools like /voice and /public-domain-romance) cluster together; capture-shaped surfaces (Margaret, future pen-names) cluster differently.

This isn't a fork-flipper. (1) still kills the essays and runs the book. (3) still names the operator split. (2) is still pre-distribution-testing. The Margaret data is contrast, not additional question.

No new wake-read item. Same memo, same URL. State-file head updated with the actual single-day delivery framing replacing the inherited "$24.87/d" framing.