SEO depth plan — flagship
Mac-local post-training authority plan for posttrainllm.com: artifact datasheets, train-on-Mac runbooks, benchmark leaderboards.
posttrainllm SEO depth plan — flagship
Source of truth: ~/.fleet ledgers via seo-scoreboard.mjs (28-day window
ending 2026-09-18): 2,172 imp · 3 clicks · avg pos 23 — indexed and
visible, parked deep on queries where we mostly can’t win.
What the queries say
| Query shape | Imp | Pos | Verdict |
|---|---|---|---|
| “apple foundation models” variants | ~20 | 62–89 | Apple’s turf — ignore |
| “13b ppo lora” / “aider polyglot benchmark” | 4 | 71–89 | Our real niche, ranked deep |
We surface for generic Apple/model queries we can’t win. The winnable niche is Mac-local post-training: LoRA, GRPO, distillation, and running/ evaluating small specialists on Apple Silicon — which is literally the product (AGENTS.md north-star: “win on the Mac”).
The play: own the Mac-local post-training niche with artifact receipts
The site already has the structure — /artifacts/* (366 URLs cataloged),
/benchmarks/{arena,chess,game-2048}, /articles/*. Depth = every artifact
page becoming a real datasheet.
Phase 1 — artifact pages into datasheets (highest leverage)
- Each
/artifacts/<name>page gets: base model, method (LoRA/GRPO/distill), dataset, eval scores vs. base, size/VRAM, “run it on your Mac” commands, license. Today’s pages mostly describe; datasheets get cited and linked. - Prioritize the ones that already exist for real runs:
qwen3-4b-*,vibethinker-3b-*,pace-intent-router-v8,qwen06-sql-routed-v1.
Phase 2 — the “train on Mac” how-to series
- Real runbooks from
benchmark-runs/+evals/: “fine-tune a 4B model on an M-series Mac”, “LoRA vs full fine-tune on Apple Silicon — real numbers”, “evaluate two runs without moving the target” (article exists — expand the series). Each grounded in an actual run with logs linked.
Phase 3 — benchmark leaderboards as linkable assets
/benchmarks/*exists — surface per-model results as sortable tables with raw data downloadable. “Mac-local LLM benchmark” queries are the ones where a hobbyist-built leaderboard can beat corporate docs.
Phase 4 — indexing follow-through
- 341 queued URLs churning via the daily agent. Check
pnpm --dir site-health indexing status --project posttrainllm— artifact pages landingDiscovered — currently not indexedneed the datasheet treatment (thin catalog entries don’t index).
Rules
- Never publish a benchmark number that isn’t in
benchmark-runs/orevals/— the whole positioning is receipts. - Learning-project honesty is the brand: “here’s what worked on one Mac” is more credible than pretending to be a lab.
- Register shipped pages:
seo-scoreboard.mjs register --project posttrainllm --lane editorial|programmatic --summary "…" --target <url>.