Skip to content

SEO-ROADMAP.md

Latest commit

 

History

History
106 lines (94 loc) · 6.4 KB

SEO-ROADMAP.md

File metadata and controls

106 lines (94 loc) · 6.4 KB
Raw

SEO Roadmap for visionanalysis.org

Produced 2026-06-10 from a codebase audit + research pass on programmatic SEO for benchmark/comparison sites. Items marked ✅ were already fixed in the same commit that added this file.

0. Critical fixes (done)

  • ✅ Dead domain everywhere. The codebase pointed to https://visionanalysis.ai (does not resolve) while production serves https://www.visionanalysis.org. Fixed in site.ts, sitemap.ts, robots.ts, both opengraph-image.tsx footers, and scripts/generate-draft.ts. sitemap.ts/robots.ts now derive from siteConfig.url so it can't drift again.
  • ✅ Model OG images were broken. model/[slug]/opengraph-image.tsx read params.slug without awaiting (Next 16 params is a Promise), so every shared model link rendered a "Model Not Found" card. Also now picks a hardware combo that actually has results and shows its display name.
  • ✅ Broken default OG image. Root metadata referenced public/og-default.png, which doesn't exist. Removed the explicit references; the file-based opengraph-image.tsx generators are used instead.
  • ✅ Sitemap was missing /parity and /about. Added.

1. Quick wins (this week)

  1. Canonicals on every page. No alternates.canonical exists anywhere. With metadataBase set, add relative canonicals per page:
    • Home (needs a metadata export): alternates: { canonical: "/" }, plus a tuned title like "Object Detection Benchmarks: YOLO, RT-DETR, D-FINE Leaderboard".
    • /compare (currently has no metadata export at all): canonical stays /compare regardless of ?models=/?hw= params. Canonical-strip is the correct pattern for filter params - do NOT noindex.
    • model/[slug]: canonical: /model/${slug}; change OG type from "article" to "website".
    • Hardware pages, /parity, /about: same pattern.
  2. JSON-LD. Zero structured data exists today. Add a small JsonLd component and:
    • layout.tsx: Organization + WebSite (logo, sameAs: [github]).
    • / and /about: Dataset schema describing the verified benchmark dataset (distribution → GitHub generated/verified-results.v1.json, measurementTechnique: "COCO val2017 mAP@50-95, end-to-end latency"). Gets the site into Google Dataset Search - a citation channel for researchers.
    • model/[slug]: BreadcrumbList (Home → family → model).
    • FAQ and HowTo rich results are dead (deprecated by Google) - skip the markup, but answer-first FAQ content still helps AI answer extraction.
  3. Sitemap lastModified. Currently new Date() on every build, which tells Google everything changed daily and erodes trust. Use the dataset generation timestamp from generated/verified-results.v1.json instead.
  4. Richer titles/descriptions with real numbers. model/[slug] title → "{Model} Benchmark: mAP, FPS, Latency on COCO"; embed actual best-mAP/best-FPS numbers in the description (numbers lift CTR and AI citation).
  5. llms.txt route listing the leaderboard, parity, model pages, and the raw JSON dataset on GitHub. Cheap; mostly read by coding agents, not AI search crawlers - don't build strategy on it.

2. Programmatic page strategy (the gold mine)

  • /vs/[pair] static comparison routes - yolov9c-vs-yolox-l style, model ids ordered alphabetically (301-redirect the flipped order). Query-param /compare can never rank for "yolov9 vs yolox"; static pages can.
  • Don't generate all pairs. Only pairs where both models have verified data AND the pair is meaningful: cross-family within ~2.5× params, same-family adjacent variants, flagship-vs-flagship. Likely 100–400 genuinely differentiated pages - the kind that survives Google's pSEO crackdowns (thin variable-substitution pages get wiped; pages backed by proprietary data are durable).
  • Page anatomy: H1 "{A} vs {B}"; answer-first verdict in the first viewport (44% of LLM citations come from the first 30% of a page); side-by-side table with win highlighting; speed breakdown; cross-runtime ranking-stability; license & provenance ("which is safe for commercial use" is underexploited, high-intent); link to /compare?models=a,b to customize.
  • JSON-LD: TechArticle with dateModified = benchmark date (freshness is the top LLM-citation signal: 76% of ChatGPT's most-cited pages updated in last 30 days), BreadcrumbList, isBasedOn → Dataset node.
  • Hardware long-tail: when Raspberry Pi / edge data lands, enrich /hardware/[slug] for "best object detection model for raspberry pi 5"; add /vs/[pair]/[hardware] only where the ranking flips vs default hardware.
  • Family hubs: /family/[slug] from families.json (paper, authors, innovations, variant table) for "yolov9 variants"-shaped queries.
  • Internal linking mesh: model page → its 3–5 strongest /vs/ pages; /vs/ page → both model pages + hardware page + /parity (trust) + sibling comparisons; every /vs/ page in the sitemap.

3. Article pipeline integration

Pipeline output (article-pipeline/output/, public/articles/draft-*.html) is currently orphaned - no route serves it, it's not in the sitemap or link graph. Recommended: merge article prose INTO /vs/[pair] pages (store validated prose as JSON/MDX keyed by pair; the vs page renders data tables always, prose when available). One URL accumulates all authority per pair. If standalone articles are kept, serve via a real /articles/[slug] route (never raw HTML in public/), with canonical to the /vs/ page when content overlaps. Re-run the pipeline when new benchmark data lands - that's the freshness engine.

4. Measurement

  • Google Search Console: add a Domain property for visionanalysis.org, submit the (now-fixed) sitemap. Bing Webmaster Tools too - Bing feeds ChatGPT search.
  • Query buckets to track: "X vs Y"; "{model} fps/latency/benchmark/tensorrt"; hardware ("raspberry pi", "a100"); license queries; head terms ("fastest object detection model").
  • AI crawler telemetry: watch GPTBot / OAI-SearchBot / ClaudeBot / PerplexityBot user agents in Vercel logs; spot-check monthly whether ChatGPT/Perplexity cite visionanalysis.org for "yolov9 vs yolox benchmark".

Suggested order: canonicals + metadata (1) → Dataset JSON-LD + GSC/Bing setup → /vs/[pair] for the ~20 pairs with verified data → article-prose merge → family hubs and hardware enrichment as data grows.