SEO + GEO Engine . Agency OS

What it does

GEO is the newer half of search: when a customer asks an AI assistant who to hire or which product to buy, the assistant answers by citing pages it trusts. This skill measures whether your pages are the ones getting cited, and fixes what stands in the way. It runs a full audit that combines classic SEO signals with AI-search signals into one composite score: can AI crawlers even reach the site, are the pages written in citable, self-contained answer blocks, does the brand show up on the platforms AI assistants learn from, and is the structured data in place. Alongside the audit it generates and validates llms.txt, the plain-text file that tells AI systems what your site is and which pages matter. It also scores drafts for citability, gates content quality before anything ships, scrubs hidden watermarks and AI tells out of copy, builds a blueprint for beating the pages that outrank you, and ranks keyword opportunities by what they would actually return. On request it renders the whole audit as a branded PDF you can hand to a client.

Say this to start

This skill has no button. You start it by saying what you want. Any of these will do it:

> run a GEO audit on my site

> will AI assistants cite this page

> generate an llms.txt for my site

> score this draft for AI citability

> find the gaps against the pages ranking above us

When to reach for it

you want to know whether AI assistants can find, read and cite your site at all
a client is asking about AI search and you need an audit you can put a score on
a draft is about to publish and you want it gated for quality and cleaned of AI tells first
your pages rank but never appear in AI-generated answers
you have competitor pages above you and want a concrete plan to beat them, not just match them
you have a keyword list and need it ranked by what is actually worth writing

When NOT to use it

If you actually want	Use this instead
technical and on-page SEO health: page speed, crawl and indexing errors, schema setup, analytics wiring	`seo-site-audit`
standalone keyword discovery, volume and difficulty pulls, or a topic-cluster plan	`keyword-research`
generating pages at scale from a template plus a dataset	`programmatic-seo`
writing the actual blog post or article the audit calls for	`blog-writer`
a final pre-ship quality and proof check on a finished piece	`critique`

Before you start

What you need	Why
The site URL, or the file you want scored	everything starts from a real page; the skill asks for this and the brand name if the workspace cannot answer	Required
`verified-claims.md` populated with consented proof	any result or testimonial that appears in optimized copy must trace to the registry; an unverified specific on a page gets flagged before that page is promoted for AI citation	Required
Brand memory in place: voice, audience, keyword plan, positioning	relevance scoring is grounded in who the content has to reach; with these on file the skill does not re-ask what the brain already holds	Optional
An SEO data tool connected via `/connect`	with one connected you get live volume, difficulty, position and crawl data; without one the skill fetches pages directly, runs its own scoring, and labels every estimate `[ESTIMATED]`	Optional
Competitor URLs or data for the gap analysis	the beat-them blueprint is built from real competitor pages; give it nothing and that step is skipped, never guessed	Optional
One extra Python package installed if you want the PDF report	every scoring capability runs on a plain Python install; only the branded PDF renderer needs an install first	Optional

How it runs

Load the brain, ask the minimumPulls voice, audience, keyword targets and positioning from the workspace under the freshness rules, applies recent learnings.md patterns, and asks only for what it cannot answer: the site URL or file, and the brand name for the mention scan.
Check crawler access firstFetches robots.txt and checks whether the crawlers AI systems use can reach the site. This comes before anything else, because a page AI crawlers cannot read cannot be cited no matter how good it is. You get recommended additions, with block-or-allow decisions left to you.
Run the composite auditFetches the homepage, detects the business type, checks for llms.txt, and crawls the sitemap within a page cap. Scores AI visibility, platform readiness, technical foundations, content quality and structured data, then combines them into one composite score with a plain-language interpretation.
Score citability block by blockRuns the citability scorer over the top pages or your draft: are the blocks self-contained, answer-first, structurally readable, backed by real numbers. Low-scoring blocks come back with concrete rewrite suggestions, not just a grade.
Generate or validate llms.txtBuilds llms.txt and its full-length companion from pages that actually exist on the site, or validates the one you already have for completeness and accuracy. Only real pages are ever listed.
Gate and scrub before anything shipsRuns the content quality gate with a pass threshold, and the scrubber that strips invisible watermarks, em-dash habits and AI phrasing. Ungrounded specifics get flagged to verify or remove. You see the diff and approve before any original is overwritten.
Gaps and opportunitiesWith competitor pages supplied, builds the beat-them blueprint: thin sections to out-write, unsupported claims to counter with cited fact, stale content to out-fresh, and a target word count. With a keyword list supplied, ranks opportunities into priority bands by return, not raw volume.
Report and learnStates plainly which data source everything came from, leads with the score and the single biggest issue, then the prioritized fix list. Optionally renders the branded PDF. Appends deliverables to assets.md, appends opportunities to keyword-plan.md, and asks how the work landed so the next run is sharper.

What you get

A composite GEO plus SEO audit score with a prioritized action plan: quick wins first, then medium-term, then strategic
A per-block citability read with concrete rewrite suggestions for weak blocks
An AI crawler access report with recommended robots.txt additions
A brand authority scan across the public platforms AI assistants learn from, with fixes for the weakest areas
Ready-to-deploy llms.txt and llms-full.txt, or a validation report on the ones you have
A content quality gate verdict and a clean, watermark-free draft with the diff shown
A beat-them competitor blueprint and a priority-ranked keyword opportunity table, where the inputs exist
Optionally, a branded PDF audit saved to content/visuals/ and named by domain and date
Deliverables logged to assets.md and new opportunities appended to keyword-plan.md, never overwritten

Honest limits

Read this before you rely on it

AI citation is influenced, never guaranteed. The skill makes pages as citable as they can be; no tool controls what an AI assistant chooses to cite, and anyone promising otherwise is selling something.
Unverified claims on a page are flagged before the page is promoted for AI citation. A specific like a customer count or a result that does not trace to verified-claims.md comes back as verify-or-remove, not as copy to amplify.
llms.txt only ever lists real pages. Pages that were not provided or found on the site do not get invented to pad the file.
Competitor analysis is skipped when no competitor data is given. You get an explicit "skipped for lack of input", not a filled-in guess.
It never invents a ranking, traffic or search-volume number. Without a connected data tool it says so and labels every estimate [ESTIMATED]. The industry research it draws on is presented as cited third-party findings, never as your own results.
It analyzes and recommends; it does not publish. After optimization it asks where the content should go, and the decision to ship stays with you.

Where people go wrong

The mistake	Do this instead
Polishing content while AI crawlers are blocked in `robots.txt`	Fix access first. A blocked page cannot be cited no matter how well it is written; the skill checks this before scoring for the same reason.
Writing for keywords the way you did for classic search alone	AI search rewards self-contained, answer-first blocks with definitions and real numbers. Optimize for citation, not just clicks.
Chasing backlinks as the whole authority play	AI visibility tracks brand mentions across public platforms more than link counts. Work the mention scan's recommendations too.
Publishing a draft that failed the gate because the deadline was close	Run the scorer and scrubber and clear the gate. Watermarks and AI phrasing in published work cost more than the delay.
Quoting the audit's internal research figures in client copy as your own results	Keep third-party research cited as third-party research. Client-facing claims come from `verified-claims.md` or not at all.
Matching the top-ranking page's length and calling it a strategy	Beat them on depth and freshness: fill their thin sections, counter their unsupported claims, refresh what they let go stale.

Worth knowing

AI crawlers generally do not execute JavaScript. A site that renders its content in the browser can look complete to a human and nearly empty to the systems deciding what to cite, which is why the audit checks how pages are rendered and treats server-rendered content as a real ranking factor for AI search.