Best SEO and AI-search skills for your agent

Ask an agent to make a site AI-search ready and you can usually predict the output. It adds FAQPage schema. It suggests HowTo markup for the tutorial pages. It writes an llms.txt at the root and reports success. Three moves, all confidently wrong in different ways — Google retired FAQ rich results for every site on May 7, 2026, HowTo has been deprecated since Sept 2023, and no major AI vendor has confirmed it reads llms.txt at all. The failure mode isn't sloppiness. It's that SEO conventions rot faster than model knowledge, and nothing in the page an agent produces tells you which parts are three years stale.

What surprised me reading these skills is how much of the good ones is negative instruction. The strongest GEO writing spends its best paragraphs telling your agent what not to do — don't chunk content into AI bait, don't count llms.txt as a Google lever, don't stuff keywords, because unlike in classic SEO it actively costs you citations. The weaker skills aren't the ones that know less. They're the ones a version behind, still listing those same moves as wins — and two of the skills I recommend below are on the wrong side of that line in one section each. I'll say where.

Top picks

Start with geo-technical for the audit that actually inspects the site

geo-technical is the most complete technical audit I read, and the only one that treats server-side rendering as a first-class scored category rather than a footnote — fifteen of its hundred points, on the reasoning that AI crawlers don't execute JavaScript, so a client-rendered page is invisible to them no matter how good the copy is. The detection method is refreshingly literal: fetch the page with curl, hold that against what a browser renders, and if the article body only exists in the browser version, the category scores nothing. It scores eight categories and then adds a deliberately non-scoring ninth for agent-readiness signals — RFC 8288 Link: headers, and whether the server answers Accept: text/markdown with actual Markdown — with an explicit rule that absence is a recommendation, never a penalty, because these are drafts and early features. That restraint is the tell of someone who has written audit reports for clients. It also tells your agent to omit the service-discovery section entirely for ordinary business sites rather than padding the report with an irrelevant finding.

Reach for vellum-ai's geo-audit when you want a scored number in half a minute

geo-audit is the opposite bet: one Python script, six scored checks, a total out of a hundred, and the three fixes most worth shipping. Five of the six are the ones that decide whether a crawler sees anything at all — per-bot robots.txt access, server-side rendering, sitemap, homepage JSON-LD, and llms.txt. The sixth is the one I didn't expect and the one I'd keep: crawlable internal links, which reads the first fifty internal anchors and flags JS-bound <div onclick> substitutes, dead anchor text like "click here", and a nofollow ratio over twenty percent. What earns the slot is a refusal written into the instructions — if the script can't run or the domain won't resolve, the agent must say so and produce no score, because the numbers are the product and estimating them is off the table. One disagreement worth naming: this skill is on the other side of my opening argument, awarding llms.txt fifteen of its hundred points, calling it a cheap differentiator, and ranking "stand one up at the domain root" as the top fix in its own sample output. Treat that check as informational and the other five as load-bearing.

Give your agent coreyhaines31's ai-seo before it edits a single page

ai-seo is the strategy layer, and it's the one that does the hardest thing well: it separates Google's stated position from what the other engines reward, and tells you which tactics belong to which. On Google's side it quotes the position directly: the generative features run on core Search ranking, nothing has to be added to the page for them, and producing an AI-targeted variant of your content falls under the spam policy for scaled content. On the other side, ChatGPT, Claude and Perplexity genuinely do reward extractable structure. The skill's resolution is to use the structural patterns because they're good organization anyway, and never to sell them to a client as a Google lever. It carries the Princeton GEO ranking of optimization methods — citing sources at the top around +40%, keyword stuffing at the bottom actively costing 10% — and it draws a line most GEO writing blurs, between being cited and being recommended. Two cautions. It asks questions and produces a plan; it runs nothing. And there's a fork under the same name, b1rdmania/ghostclaw's ai-seo, pinned at 1.1.0 against this one's 2.2.0, missing both the Google-stance section and the query fan-out material. The two descriptions read almost identically at a glance — the newer one adds a handful of trigger phrases — so nothing on the surface warns you that one is a chapter short.

Use seo-validate when the SEO lives in your repo, not on a URL

seo-validate is the only pick that reads your repository rather than a live domain, which makes it the one you can run before anything ships. It detects the framework first — reading package.json, astro.config, svelte.config, the entry HTML — then applies framework-aware rules, so "use next/image with priority" and "add <link rel="preload" as="image">" are separate findings for separate stacks. Every rule carries a severity and, unusually, a confidence: definitive for regex hits on known-bad patterns, heuristic for absence checks. Above-the-fold <img loading="lazy"> is flagged HIGH and definitive as actively harmful, which is exactly the kind of thing that survives code review for years. It's read-only, so it will not fix anything, and it can't see anything that only exists at runtime. Its GEO category is the soft half, and it's soft in the direction this post warns about: it flags an H2 section running past roughly 375 words without an H3 as a chunking failure, arguing from a retrieval token window, and its answer-engine reference still treats FAQPage and HowTo as live options. Chunking for AI is the exact practice Google names as a mistake, and HowTo is gone. Run it for the semantics, metadata, schema, hreflang and Core Web Vitals categories; read the GEO ones as a hypothesis rather than a checklist.

Point ai-bot-log-audit at your server logs when the theories run out

ai-bot-log-audit answers a question none of the other skills can: what are the AI crawlers actually doing on your site. It hands your agent working grep and awk one-liners against Apache and Nginx access logs — hits by bot, top pages fetched, crawl frequency by day, response codes — and then a diagnostic table that reads patterns as decisions. Fetched once and never again means evaluated and deprioritized. Fetched constantly but never cited means the structure or the authority is the problem, not the access. Its content model is the sharpest thing in it: LLMs attend to the beginning and end of a document more than the middle, so key claims and the summary go at the edges, and tables go in the sag because structured data survives position. The skill is honest that it cannot read your logs for you; you bring the data, it brings the method.

Install AgriciDaniel's seo suite if you want one front door — and read the footer clause first

seo is the only entry here that is a toolchain rather than a document: bundled Python run through a claude-seo launcher, a doctor command that checks runtime readiness without changing anything, and a full audit that fans out to sub-agents conditionally — spawning the local specialist only when it detects an address, the ecommerce one only on cart signals. Its quality gates are where the maintenance shows: a hard stop that fires at fifty location pages and requires the user to justify going further, never recommend HowTo, and a precise instruction to flag existing FAQPage as informational while refusing to recommend new FAQPage for Google benefit. The GEO sub-skill, seo-geo, goes further and reports llms.txt presence while assigning it no citation weight at all. One thing to know before installing: the skill instructs the agent to append a promotional community footer to every major deliverable, including the audit reports you hand to clients.

Telling two skills with the same name apart

Skill names aren't namespaced. Anyone can publish anything under any name, so the name tells you the topic and nothing else — the publisher and the body are the identity. A skill called geo-audit can be the vellum-ai script above, six checks and a refusal to guess. It can also be zubair-trabzada's geo-audit, a multi-phase orchestration document that classifies your business type from homepage signals, pulls up to fifty URLs out of the sitemap, and fans the analysis out to sub-agents. Both are legitimate. They are not substitutes, and no card in any directory will tell you which one you've got.

Three checks separate them quickly. Read the front matter for a version and an allowed-tools list — a skill that declares Bash and WebFetch intends to run something, a skill with neither is a briefing document. Search the body for a command. Then compare version numbers across same-named skills, because forks go stale silently: the ai-seo pair above is one document at two points in its life, and the older copy is missing precisely the section that keeps your agent from over-claiming to Google.

Watch for one more shape. Some skills in the wild are tombstones — a redirect notice left so old install links keep resolving, pointing at the maintained copy in another repo. Install one of those and your agent loads a paragraph of migration instructions where an audit should be. A body under about a page long, with a heading that says the skill has moved, is the signature.

What to check before you install

License first, because it's the cheapest check and the one people skip. NOASSERTION on a listing means no license was detected in the source repo — not that the skill is free to use, and a skill can be genuinely good and still carry it. That distinction matters the moment the output leaves your machine. The check takes a minute: look for a LICENSE file in the repo, then look for a license: line in the skill's own front matter, because those two are set independently and either can be missing. If neither exists, treat the skill as reference material you read, not code you ship.

Then check what a fork dropped. addyosmani's seo is a developer's implementation reference — copy-paste JSON-LD for Organization, Article, Product, FAQ and BreadcrumbList, robots directives, hreflang blocks — and it closes with a short, unusually sober section calling llms.txt a proposed convention at roughly 0.015% adoption with no vendor confirmation, worth five speculative minutes and no content reorganization. warpdotdev's seo-aeo-audit is built on that same document, and it adds real things: Lighthouse and PageSpeed scripts, a Search Console export, and a full AEO half. It also drops the llms.txt caution entirely. The fork is bigger and, on that one point, less careful. Bigger is not more current.

What you have now

The problem was never that agents don't know SEO. It's that they know a version of it that expired, and produce clean, plausible work built on retired rich results and unproven files. The skills worth installing don't fix that by knowing more; they fix it by writing the corrections down, and by telling the agent to fetch something before it scores anything. Install geo-technical for the audit that inspects rather than assumes, ai-seo for the strategy that separates Google's position from everyone else's, and seo-validate if the work lives in a repo — add ai-bot-log-audit the moment you need evidence instead of a theory. Two of those carry a stale section apiece, which is why the picks above name them; a skill you can argue with is still worth more than a blank page. What you've avoided is the confident report: the one with a score, a checklist, and an llms.txt at the root, none of which anyone verified.