SEO sprint — PostTrainLLM

Initialized 2026-09-22 from the Fleet scoreboard (seo-scoreboard.mjs --project posttrainllm): 2,172 impressions, 3 clicks, mean position ~23 over 28 days — the strongest organic base in the Fleet. The site already has demand; the work is converting impressions into clicks with pages that answer the exact queries being served.

Demand evidence (Search Console, 28d)

Cluster	Queries seen	Impressions	Position
Apple Foundation Models	`apple foundation model(s)` + ~10 variants (`framework`, `benchmarks`, `context window`, `supported devices`, `on device`)	~40+	58–89
Mac fine-tuning / LoRA	`13b ppo lora` and existing `fine-tune-llm-on-mac`/`mlx-lora` coverage	small, well-covered	—
Benchmarks	`aider polyglot benchmark leaderboard official 2026`	~1	71

Phases

Phase 1 — Apple Foundation Models page. /apple-foundation-models publishes the measured verdict from docs/learn/apple-on-device-foundation-models.md (BFCL agentic 25%/~0%, planner 13%, OOS-refusal ~95%, 4096-token context catch-22). This is the highest-demand cluster and we hold genuinely unique measured evidence — nobody else publishes these numbers.
Phase 2 — per-model eval pages. eval-leaderboard data → evals/[model].astro pages (Qwen3-4B variants, Gemma, distilled 1.7B, 30B-A3B) with per-model scores, tok/s, RAM. The corpus already exists in browser/src/data/benchmarks/.
Phase 3 — per-recipe pages. The 18 recipe contracts → recipes/[slug].astro so each recipe is independently indexable.
Phase 4 — experiment archive depth. experiments.astro exists as an index; per-experiment pages from experimentArchive.ts give each of the ~76 runs a canonical URL.

Measurement

Site Health search family: apple-foundation-models cluster position (currently 58–89; target <30 within ~6 weeks of indexing).
Indexing: pnpm --dir site-health indexing submit --project posttrainllm --sitemap after deploy.