Every GEO tool eventually hands you a number between 0 and 100. Most posts about those numbers are marketing. This one is an anatomy lesson: I scored the GEO skills in the agent-skill directory I run on six dimensions, took apart what their checklists actually verify, and concluded that the GEO score and the SEO score are measuring two mostly unrelated things.
Where the 47 checks come from
geo-optimizer-skill runs 47 checks, and the count isn't arbitrary — it's a fairly faithful operationalization of the Princeton KDD 2024 paper that coined "generative engine optimization." That paper's central finding was that you could meaningfully change how often an LLM-based engine cites your content without touching traditional ranking: make content quotable — add citations, statistics, attributable claims — and visibility moved by double digits. The 47 checks are the paper's taxonomy turned into a linter.
The checks fall into four buckets, and it's worth being precise about them, because each bucket fails differently.
Bucket one: can the crawler reach you at all. robots.txt rules for AI crawlers. Whether GPTBot, ClaudeBot, and their relatives are allowed, blocked, or — the most common and most pointless state — unspecified. A site that blocks AI crawlers while running a GEO campaign is optimizing a page no engine will fetch. This is plumbing, not content, and a surprising number of GEO scores die here.
Bucket two: what you hand the crawler. llms.txt — a curated map of what you consider your canonical corpus. It's the "sitemap for AI" framing, and it's the bucket with the most hype and the least feedback loop, because no engine tells you whether it read the file.
Bucket three: can the machine parse you. JSON-LD and structured data. Whether your entity — your company, your article, your prices — exists as typed data rather than prose a model has to guess at.
Bucket four: are you quotable. Citability: does the content contain discrete, self-contained, attributable claims that an answer engine can lift as a citation? This is the bucket the Princeton paper actually moved the needle on, and it's the one most audits spend the least code on.
The orthogonality argument
Here's the claim I want to defend: a GEO score and an SEO score are two orthogonal measurements, and treating one as a proxy for the other wastes effort in both directions.
Consider the failure combinations. A page can be excellent SEO and terrible GEO: fast, well-linked, ranking third for its query — and composed of paragraphs with no discrete claims, no JSON-LD, and a robots.txt that happens to block the AI crawlers. Google loves it. Every answer engine is blind to it. Its SEO score is high and its GEO score is low, and both scores are correct.
The reverse holds too. llms.txt does approximately nothing for your Google rankings. A page can be eminently citable and rank on page four. The mechanics don't share a pipeline: one system measures a crawl-graph-and-relevance loop built for search ranking, the other measures fetchability, machine-parseability, and quote-liftability for engines that assemble answers.
The skill scores back this up structurally. geo-content-optimizer sits at 8.7 with a 9.5 on content quality and a 10 on security, but a 6 on engineering. Compare that with the technical SEO skills in the same directory, where engineering is usually the strong dimension and quotability isn't measured at all. Different tools, different spines. A skill that scores well on one axis scoring poorly on another isn't inconsistent — it's evidence the axes are real.
ultimate-seo-geo at 8.3 is the exception that proves the rule: it routes 26 program files and bakes in anti-fabrication rules precisely because it's trying to serve both systems at once, and the cost of that ambition is visible in the added machinery. Serving two orthogonal systems in one tool is expensive; doing it accidentally, by running an SEO audit and calling it a GEO audit, is worse.
Use both scoreboards, separately
The practical failure isn't ignorance of GEO — it's score conflation. Someone runs a 92/100 SEO audit, sees "AI visibility" mentioned somewhere in the tool's marketing, and concludes the AI side is covered. The robots.txt rule for GPTBot was never checked. Nobody wrote the quotable sentence.
Run both audits, but keep the scorecards physically separate — different reports, different owners if you have the headcount. seo-geo, part of the seranking suite that lands in the 8.6–8.9 band, will give you the GEO side with the consistency that comes from a 15-skill catalogue operation; geo-content-optimizer will go deeper on the citability bucket specifically, which is where the Princeton paper says the gains are.
The number that matters is not either score. It's the lowest bucket. A GEO audit where you scored 90 overall but blocked the crawlers is a zero, presented attractively.
Originally published on Skill123 — a directory of 765 agent skills, each scored on a six-dimension rubric.
Top comments (0)