seo: purge thin compare pages + benchmark-first title on products & compare

[Flotapponnier]

Summary

Bing WMT audit 2026-07-05 showed only 2 of 5093 sitemap URLs indexed. Root cause: 4938 ad-hoc /compare/* pages emitted at ≥1 shared bench threshold, producing near-duplicate templates that starved Bing's crawl budget and hurt domain trust.

Changes

1. Sitemap — hybrid threshold (src/app/sitemap.ts + new src/lib/compare/brand-whitelist.ts)

• Both providers in BRAND_WHITELIST → emit at ≥ 1 shared bench
• Otherwise → emit only at ≥ 3 shared benches
• Drops from 4938 → ~226 ad-hoc + 21 curated ≈ 247 URLs
• Preserves every commercial X vs Y pair users actually search for: helius-vs-mobula, alchemy-vs-moralis, chain-vs-chain, perp-vs-perp, PM-vs-PM, bridge-vs-bridge…

2. Curated pair fix (src/data/compare-pairs.ts)

• Remove jupiter-vs-raydium — raydium has 0 bench appearances so the page 404s at render via hasSharedBenches(). Re-add when raydium is measured in any OCB bench.

3. Product page SEO (src/app/products/[slug]/page.tsx)

• Title now leads with {Provider} Benchmark {year} — head-term match on <provider> benchmark queries (the query pattern OCB gets from LLM search).
• Meta description ends with As of YYYY-MM-DD — LLM extractability signal (ChatGPT/Perplexity/Copilot cite dated content more).

4. Compare page SEO (src/app/compare/[slug]/page.tsx)

• Title now leads with {a} vs {b} Benchmark {year}.
• Meta description is unique per pair — bench count + provider names + date. Kills the identical duplicate signal that had Bing skipping compare pages.

Numbers

Simulated locally on current bench specs:

Threshold	Total URLs	Brand-vs-brand coverage
≥ 1 (current)	4959	✅ but with 4712 garbage
≥ 2 uniform	165	⚠️ 110 brand pairs lost
≥ 3 uniform	54	❌ 191 brand pairs lost
Hybrid (this PR)	247	✅ all preserved

Test plan

• pnpm typecheck — clean
• pnpm lint — clean
• Manual on staging: /products/moralis → title contains 'Moralis Benchmark 2026'
• Manual on staging: /compare/codex-vs-mobula → title 'Codex vs Mobula Benchmark 2026', meta unique
• Manual on staging: /compare/jupiter-vs-raydium → 404 (removed)
• Manual on staging: /compare/metascalp-vs-wunder → 404 (dropped by hybrid threshold)
• Sitemap.xml on staging → ~250 /compare/* URLs (was 4865)

Follow-ups (separate PRs)

• Per-URL <lastmod> from actual data timestamps (currently a single global timestamp)
• Split sitemap into per-category files + index
• IndexNow wire-up for Bing fast-indexation
• Unique intro content per compare page via LLM at build-time (~$0.20 total)

[Flotapponnier]

Follow-up commit 46be62e : data-driven prose above the fold on both /compare/[slug] and /products/[slug]. Replaces the identical templated intro paragraphs (which were the last piece of above-the-fold duplicate content) with unique per-page summaries derived from live Prom data.

Example /products/moralis before → after:

• Before: 'Moralis performance benchmarks, live across 2 categories. Reproducible measurements, open methodology.' (identical on 104 product pages)
• After: 'Moralis is measured across 2 live OpenChainBench benchmarks, with 1 drop journal framing #1 finish: NFT collection metadata (ranks drop journal framing #1, 96.9% p50), Wallet labels coverage (ranks publish harnesses #2 of 5, 42.4% p50).'

Example /compare/codex-vs-mobula:

• Before: 'Side by side OpenChainBench measurements. Identical layout, no editorial verdict...' (identical on 247 compare pages)
• After: 'Codex leads on 2 of 4 shared benchmarks, Mobula on 2. Codex wins on aggregator-head-lag (128ms vs 195ms), wallet-labels-coverage. Mobula wins on metadata-coverage, network-coverage.'

Zero editorial claim, zero LLM, zero $. All prose derived deterministically from the same Prom values the table renders.

Fallback to a neutral sentence when p50 <= 0 so cold ISR / harness restart doesn't render lies.