Most SEO audits are a PDF of screenshots and adjectives. “Your site could be faster.” “Consider adding more content.” Forty pages, no numbers you can act on, and no way to tell whether anything improved afterwards.
We got tired of that, so the first thing we built into PAAK’s SEO & GEO service was an engine that produces the opposite: one score out of 100, built from checks a machine can verify, with a ranked fix list — and the same audit re-run after the work, so the improvement is a number, not a feeling.
This is how it works. We’re publishing the internals because the methodology is the product — if you can read this and fix your own site, genuinely, go do that.
What the engine actually does
Point it at a URL and it crawls the site the way an AI answer engine would: fetches the pages, reads the robots rules, parses the structured data, weighs the content. Then it scores everything against 47 research-backed methods — the checks are grounded in the academic work on generative engine optimisation (the Princeton GEO paper from KDD 2024, and the AutoGEO follow-up), not someone’s blog opinions.
The output is a 0–100 score across eight categories, each weighted by how much it actually influences whether AI engines can read, trust, and cite a business:
| Category | Weight | What it’s checking |
|---|---|---|
| Robots access | 18 pts | Can the AI crawlers get in at all? Checks whether OAI-SearchBot, ClaudeBot, Claude-SearchBot and PerplexityBot are allowed in your robots.txt. Blocked bots = invisible business. |
| llms.txt | 18 pts | The AI-readable index of your site. Scores presence, structure, section depth, and whether a full version exists. Most sites: zero. |
| Structured data (JSON-LD) | 16 pts | The machine-readable facts under your pages — Organization, FAQPage, Article, WebSite schema, and how rich each block is. This is the layer that lifted LLM extraction accuracy from 16% to 54% in Semrush’s testing. |
| Meta tags | 14 pts | Titles, descriptions, canonicals, Open Graph — the classic layer, still load-bearing. |
| Content shape | 12 pts | Is the content answer-shaped? Clear heading hierarchy, real numbers, lists and tables, key facts front-loaded where a model (or a skimming human) finds them. |
| Brand & entity | 10 pts | Does your business read as one coherent entity — consistent naming, an about page, contact details, location identity, topical authority? |
| Signals | 6 pts | Language declaration, RSS, content freshness. |
| AI discovery files | 6 pts | The newer conventions — ai.txt, structured summary and FAQ files — that early-adopter sites ship. |
Every point is earned by a specific, checkable condition. Not “your schema could be better” — “you have an Organization block but no FAQPage: +3 points available, here’s the generated file.”
The score matrix
The bands are deliberately blunt:
- 86–100 — Excellent. Fully readable and citable by AI engines. Rare; you’d know.
- 68–85 — Good. The bones are right, specific gaps remain.
- Below 68 — you’re leaving visibility on the table, usually for reasons that cost nothing to fix.
For context from our own runs: our portfolio site first scored 45. A real hospitality client, watda.co.nz, scored 50. Ordinary, decent websites — half-invisible to the machines now answering their customers’ questions.
The part that matters: the fix list
A score without fixes is a diagnosis without a prescription. The engine’s second half generates the actual remediation, ranked by points-per-effort:
- a robots.txt patch adding the missing AI bot permissions
- a generated llms.txt built from your sitemap
- the exact JSON-LD schema blocks your pages are missing (Organization, FAQPage, WebSite — as needed, ready to paste)
- a content rewrite checklist for pages that fail the answer-shape checks
Then — and this is the check most audits skip — it asks the engines directly. The citation checker queries real AI engines with the questions your customers ask (“best [category] in [area]”) and reports whether your brand gets named, whether your domain gets cited as a source, and which competitors are cited instead of you. That’s the ground truth the score is a proxy for.
Before and after, measured
Two documented runs:
- Portfolio site: 45 → 61. One fix pass — robots opened, llms.txt shipped, schema added.
- watda.co.nz: 50 → 66. Same categories, live business, fixes deployed to production.
Both jumps came from the mechanical layers — access, structured data, discovery files. That’s typical: the first 15–20 points are engineering, fixable in days. The climb from there runs on slower fuel — citations, reviews, content authority — which builds over months, not sprints. We tell clients both halves of that sentence, because the honest timeline is the difference between an SEO partner and an SEO subscription.
Why we run this before touching anything
Every PAAK web or SEO engagement starts with this audit, for one reason: it replaces opinion with a baseline. If we rebuild your site, the score proves the rebuild worked. If we only tune what’s there, the score proves the tuning worked. And if your site’s already scoring 80+ — we’ll tell you that too, and you’ve saved yourself a project.
If you’ve already read our piece on checking whether AI recommends your business, this engine is the industrial version of that 10-minute self-check — same question, measured properly.
Want your number? Ask us for a visibility audit — you get the score, the competitor citations, and the ranked fix list, whether or not you hire us to do the fixing.