The method behind the score

How SEO for AI Agents reads your site.

Every score in SEO for AI Agents can be traced to a deterministic check, evidence captured at audit time, and a weight chosen with intent. No black box. No vibes.

50+

Deterministic checks; every one carries a weight, severity, and human-written fix.

— Audit registry, 2026

1–6

Weight range — a missing canonical doesn't count as much as an unindexable site.

— Internal scoring rubric

Checks scored on vibes. Everything is parsed from real HTML, headers, or schema.

— Methodology principle #1

Discovery, the way Google does it

We read robots.txt, sitemap.xml, and llms.txt - then expand from the entry page. We respect Disallow rules, follow same-origin links, and cap at fifty pages so the run finishes in ninety seconds.

JS-rendered crawl

Every page is fetched through Firecrawl, which executes JavaScript the same way Google's evergreen Chromium does. We capture status codes, response headers, the rendered HTML, parsed markdown, JSON-LD blocks, images, links, hreflangs, and Core Web Vitals via the PageSpeed Insights API.

50+ expert checks, seven dimensions

Each check is a small, deterministic function with a weight, severity, evidence shape, and a 'fix' written by a human. Foundational SEO covers the basics that still matter. Technical covers crawl, render, and speed. On-Page covers structure and intent. AIO and GEO cover the new game - being cited by ChatGPT, Perplexity, and Google's AI Overviews. Local covers the pack. E-E-A-T covers trust.

Weighted scoring, not averages

A missing viewport tag matters more than a sparse Twitter card. Each check has a weight from 1 to 6. Dimension scores are weight-normalised against the maximum achievable, then the overall is the weight-normalised average across all non-N/A checks.

Evidence is mandatory

Every finding ships with the raw data the check looked at - the HTTP header, the JSON-LD object, the offending URLs, the heading text. You can audit our audit. Open the receipts drawer on any finding to see the JSON.

The brief is drafted, critiqued, then shipped

Gemini 2.5 Pro reads the findings and writes a one-line verdict, a CEO brief, a technical brief, and a 30/60/90 plan with owners and counterfactuals. A second pass criticises that draft against a rubric: is the headline metric counter-intuitive? does the plan have sequencing? is every recommendation actionable? The revised draft is what you read.

The scoring rubric

Every score is a weighted percentage, not a vibe.

Each of the ~50 checks emits a status (pass / warn / fail / n-a), a weight from 1 to 6, and a fractional score between 0 and its weight. Dimension scores divide the earned score by the maximum achievable across non-n/a checks. The overall is the same calculation across the full set.

Weight	Severity tier	Example check
6	Critical · ranking-blocker	indexability, render-vs-HTML diff
5	Critical · trust-blocker	HTTPS, broken internal links
4	High · category-defining	schema depth, AIO answer-first, LLM brand visibility
3	Medium · compounding	internal-link graph, schema required props, og:image
2	Low · hygiene	security headers, hreflang, opening hours
1	Cosmetic / advisory	Twitter card variant, freshness hint

Status → score

pass → weight × 1.00
warn → weight × (0.2 - 0.8) // each check picks a partial credit value
fail → 0
n/a → excluded from denominator

How each category is validated

The signal source for every dimension.

Foundational

Parsed from the rendered HTML head and the live response: title, meta description, canonical, robots meta, sitemap.xml, robots.txt, indexability flags. Validated against Google's documented length and uniqueness guidance.

Technical

HTTPS via response URL, redirects from the response chain, Core Web Vitals via PageSpeed Insights when reachable, broken links from live HEAD probes, render-vs-HTML diff via a plain (no-JS) fetch of the entry page compared to the JS-rendered markdown.

On-Page

Heading structure, semantic landmarks, word count and lexical density from parsed markdown, intent-match heuristics, internal-link counts per page, near-duplicate detection across the crawl set.

AIO

llms.txt presence + parse, answer-first lead detection, question-shaped H2/H3 ratios, factual density (numbers/dates/entities per 1000 words), chunkability (paragraph length distribution), schema.org required-property validation against Google's rich-result rules.

GEO

Entity clarity via Organization @id + sameAs edges, Wikidata/Wikipedia signal presence, author Person entities, citation density, and a live brand-visibility probe (Firecrawl web search) checking whether the site appears in the top results for its own hostname and brand.

Local

LocalBusiness schema with required address/telephone/openingHours, NAP consistency across crawled pages, geo coordinates, locations index page detection.

E-E-A-T

About / Contact / Privacy / Terms presence, author bios, external citations to authoritative sources, security headers (HSTS, CSP, X-Frame-Options, Referrer-Policy, Permissions-Policy) inspected directly on the entry response.

What we don't claim

SEO for AI Agents is a senior consultant's first pass, not their final deliverable.

We crawl up to fifty pages, not your full site. We can't see Search Console data, backlink history, or your CMS. A real engagement layers those in. What SEO for AI Agents gets right is the surface: every signal a smart crawler can read in ninety seconds, scored honestly, with the work shown - and every claim in the AI brief is footnoted to the underlying check that generated it.

Primary sources

Every weight cites documentation, not opinion.

Each check derives its severity from a spec or rubric published by the entity it grades against. The full catalog lives on the references page; the most-cited sources are below.

All references