We ran AISO on AISO (and on our parent site) in parallel

The claim: AISO is a deterministic AI-search visibility scanner with an Auto-fix runner that takes a scan, matches each issue to a patch template, and emits a playbook an AI coding agent can apply. The product is the loop: scan → playbook → patch → re-scan.

The cleanest way to prove that claim is to run it on the two sites we own and publish the before-and-after numbers.

Test targets:

• agntdot.com — the AGNT platform marketing site. Next.js 16, React 19, 412 URLs in the sitemap, live in production.
• aiso.tools (this site) — the AISO app itself. Same stack, shipped during Sprints 1–6, dogfooded continuously.

Both were scanned at baseline, both got the Auto-fix treatment, both were re-scanned. Two independent AI coding agents ran the patches in parallel: Claude Code on the agntdot.com codebase, Codex on the AISO codebase. Neither session saw the other until both finished.

Run 1: agntdot.com — 65 → 93 with Claude Code

Baseline scan: 65/100, Visible tier, 7 issues. 2 critical, 3 high/medium across Content Structure, Citations, Schema, Freshness, Entity Clarity, Off-site Authority.

The playbook named 6 specific patches. Each mapped to a file and a template-driven prompt from lib/fixes/prompts.ts. Claude Code applied them in one session:

1. Entity / Self-identification — rewrote the hero opener from third-person description to “We are AGNT — an AI agent platform …”
2. Citations — added inline anchor tags to two authoritative sources (Princeton GEO paper, Wikipedia A2A) inside the opener paragraph.
3. Freshness — added a visible <time datetime="2026-04-20">Updated April 20, 2026</time> beneath the CTA.
4. Structure — unwrapped the h1 from its single-child FadeIn so h1.next() returns real siblings during the “above-fold summary” walk.
5. Schema — added founder, BreadcrumbList, and Person JSON-LD; added a license field to the existing Dataset block (also closed a GSC enhancement warning in the process).
6. Citations density — added a new StandardsStrip component rendering 15 inline authoritative citations (W3C, IETF, arXiv, MDN, Wikipedia, schema.org) between SocialProof and FinalCTA.

Re-scan: 93/100, Cited tier, 0 issues. Six of eight dimensions now pass. The two warnings remaining (Citations at 15/20, Off-site at 3/5) are the density-threshold ceiling and the missing YouTube/Reddit channels respectively — documented in the skipped list of the playbook, not failures of the fixer.

Elapsed: ~30 minutes including 3 intermediate re-scans to confirm each group of patches moved the score as predicted.

Run 2: aiso.tools — 59 → 80 → 93 with Codex + Claude Code

Baseline scan: 59/100, Partial tier. 9 issues. aiso was shipped during Sprints 1–6 with the scanner + runtime-visibility panel + docs + MCP server + CLI, but the app itself hadn't been SEO-hardened yet.

Phase A — Codex hardening pass (running in parallel, different model, different session): added app/robots.ts, app/sitemap.ts, app/manifest.ts, app/llms.txt/route.ts, app/llms-full.txt/route.ts, an /about page with team/founder context, tightened the answer-first homepage, added scan-page metadata that noindexes queued/failed scans, global Organization + WebSite JSON-LD in a new lib/seo.ts, canonical metadata everywhere, SEO regression tests.

Phase A re-scan: 80/100, Visible tier. Crawler access, structure, entity clarity, freshness, and llms.txt all passing. The playbook printed the same three remaining dimensions to target: Citations, Schema, Off-site.

Phase B — Auto-fix playbook (Claude Code, same session as agntdot.com): the same template-driven patches ran against aiso's layout and page:

1. Added BlogPosting, BreadcrumbList, and top-level Person schema to buildGlobalSchemas().
2. Added a Wikipedia A2A link to the footer (Wikipedia is one of the 6 off-site platforms the scanner detects).
3. Extended the homepage “Research + standards references” paragraph from 3 citations to 11 (arXiv, IETF RFC 8615 and RFC 9309, W3C WCAG 2.2, W3C DID Core, MDN HTTP, Wikipedia on A2A / robots.txt / llms.txt / schema.org).

Phase B re-scan: 93/100, Cited tier. Schema 15/15 (max). Same pattern as agntdot.com, exactly the same scoring pins holding at 93.

What moved the scores

The dimension-by-dimension delta, same for both sites:

What didn't move (and why that's honest)

Two dimensions plateaued at partial credit on both sites. The Auto-fix playbook explicitly marks them as “skipped” rather than fabricating a fix:

• Citation Density stuck at 15/20. The scanner awards the full 20 only when authoritative citations exceed 3 per 500 words and stats exceed 2 per 500 words. Adding 10 more inline cites also adds ~300 words — density barely moves. Breaking above the threshold requires aggressive prose compression, which harms readability. We accepted 15.
• Off-site Authority stuck at 3/5. The scanner checks for header/footer links to wikipedia, reddit, youtube, g2, linkedin, github. Both sites now hit 3 of 6. The remaining 2 points need real channels on YouTube, Reddit, or G2 — creating fake handles would be dishonest, so the playbook surfaces it as a human-action item instead.

The pipeline, codified

What the agents did by hand in this run is now a shipped endpoint + CLI command on aiso.tools. The loop:

# 1. Scan any URL (free, no signup)
aiso scan https://your-site.com
#   → scanId returned

# 2. Pull the Auto-fix playbook
aiso fix <scanId> --claude-code > playbook.md
#   → markdown with patch prompts + file hints + projected score

# 3. Feed it to Claude Code in the target repo
claude < playbook.md
#   → Claude reads the playbook, applies each patch, commits

# 4. Re-scan to confirm
aiso scan https://your-site.com
#   → new scanId, higher score

The playbook is structured JSON too — GET /api/auto-fix/<scanId> returns every fix with a stable ID, severity, projected gain, file hints (glob + grep + framework-specific notes), and a ready-to-paste patch prompt. Wire it into any agent harness, not just Claude Code.

Takeaways

1. AI-search scoring responds to specific, narrow changes. +28 points came from 6 edits on agntdot.com. Three of those edits were single-file, single-line. The rubric isn't mystical; it's a checklist.
2. The scanner and the fixer are the same model of the world. Every dimension the scanner checks has a template in lib/fixes/prompts.ts. Every template was written from reading the scanner source. That tight coupling is the reason the Auto-fix playbook doesn't hallucinate solutions.
3. Parallel AI agents don't conflict when scopes don't overlap. Codex rewrote aiso/** while Claude Code rewrote agnt-pwa/**. Zero merge conflicts, zero coordination required. File-scope isolation is the lightweight alternative to heavyweight multi-agent orchestration frameworks.
4. Plateaus are the truth signal. Both sites stopped at 93 for the same reason (citation-density threshold + missing social channels). The ceiling tells us where real product work is needed, not where the scoring rubric is flawed.
5. Dogfood first. If AISO couldn't fix aiso, we'd have no business selling it. The loop ran on us before it ran for anyone else.