seo-site-audit
Skill0 starsChanged 2 months agoMITView on GitHub
Audit a website's organic search performance and produce a prioritized, evidence-backed remediation plan — crawlability, indexing, site architecture, keyword targeting, content quality and E-E-A-T, backlink authority, page experience and Core Web Vitals, structured data, international and local signals. Built on the audit method in The Art of SEO, 4th edition. Use when someone asks to audit, review, evaluate, grade or "check" a site's SEO; wants a baseline, pre-launch, pre-migration or pre-redesign review; asks what to fix to win more organic traffic; asks why a site has never ranked well or has persistently underperformed without any specific decline; or hands over a Screaming Frog / Sitebulb / Semrush / Ahrefs crawl export or a Search Console report and wants it interpreted. Also fires on the same requests in Vietnamese, such as "audit SEO", "kiểm tra SEO website", "đánh giá SEO", "chấm điểm SEO", "rà soát kỹ thuật SEO", "vì sao website chưa lên top", "cần sửa gì để lên hạng", "soát SEO trước khi chuyển domain". Use seo-traffic-diagnosis INSTEAD whenever there is a specific decline to explain — traffic or rankings that fell from a previously higher level, pages that were indexed and are no longer, or an actual or suspected manual action.
1 other project uses a version of this.
Adopted by
Most recently taken up by
- bolasblack/claude-skillstheir own version, first seen Oct 6, 2026
Other versions1 project uses a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 1
- bolasblack/claude-skills★ 4same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repositoryOne other file it gives its agents.
The file
We split this skill into the rules it gives an agent
7 rules. About 0 are common, 5 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 78
If the user has not said, infer from the site itself and declare the assumptions in the report header.
- Distinctive: line 85
Before any analysis, write down what you actually have.
- Distinctive: line 127
NOT ASSESSED means *unobtainable*, not *not handed to me*.
- Very specific: line 170
Read references/audit-checklists.md.
- Very specific: line 256
Read references/scoring-rubric.md.
- Distinctive: line 322
Write the report in the language the user is writing to you in.
- Distinctive: line 331
Treating tool scores as ground truth: Every vendor's "SEO score" is that vendor's opinion.
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
SEO site audit
An SEO audit has one purpose: find the issues and opportunities that, once acted on, increase organic search traffic. Everything below serves that. A finding that cannot be tied to a plausible traffic or conversion effect is trivia, not a finding.
This skill runs the audit the way The Art of SEO (4th ed., Ch. 10) does — as six parallel evidence workstreams, not as one long checklist read top to bottom. Each workstream sees things the others cannot.
The rule that matters most: never invent evidence
An SEO audit is worthless the moment it contains a number nobody measured. The most common failure of an AI-run audit is filling gaps — guessing a crawl depth, inventing an index count, estimating Core Web Vitals from a screenshot, asserting a backlink profile is "weak" with no link data in hand.
A dimension you have no data for is NOT ASSESSED. It is never a pass and never a fail. Missing data blocks the verdict on that dimension; it does not get skipped silently and it does not get filled in from priors.
Three things are always distinguished in the output:
| Label | Meaning |
| OBSERVED | You saw it in supplied data or fetched it yourself. Cite where. |
| INFERRED | A reasoned conclusion from observed facts. State the chain. |
| NOT ASSESSED | No data. Say what data would settle it and how to get it. |
A narrow audit honestly labelled is useful. A broad audit built on invention is worse than none. But NOT ASSESSED is a last resort, not a first move — see "Collect before you mark NOT ASSESSED" in Step 2 for what you can go and get yourself before declaring anything unassessable.
Step 1: Frame the audit
Establish before looking at anything:
- The site — domain, and which sections are in scope.
- Type of business and site — ecommerce, publisher, SaaS, local services,
marketplace, lead-gen. The weighting in
references/scoring-rubric.mdchanges with this. - Why now — scheduled (quarterly/semi-annual), or triggered? Triggers listed in the book: an algorithm update, an unexpected traffic or ranking drop, site changes shipped without SEO review, an executive request, a migration or redesign coming up. A triggered audit narrows toward the trigger.
- What the site is trying to win — the target queries, products, or topics. Without this, "improve SEO" has no direction.
- YMYL status — does the site touch health, finance, safety, legal, or major
life decisions? If yes, the content quality and E-E-A-T bar rises sharply.
references/scoring-rubric.mddefines exactly what YMYL changes in the weighting and severity; it is the only authority on that, so do not apply a second escalation on top of it here.
If the user has not said, infer from the site itself and declare the assumptions in the report header. Do not stall the audit to ask — but do ask if the scope is genuinely ambiguous in a way that changes the work (e.g. a 2-million-URL marketplace where "audit the site" could mean anything).
Step 2: Build the evidence ledger
Before any analysis, write down what you actually have. This is a required artifact, not a formality — everything downstream reads it.
Record coverage, not just possession. "We have Search Console" tells you nothing about whether you can score anything with it. For each source capture: which property or account, the date range, any filters or segments applied, when it was captured, and how many URLs it covers out of the site total.
| # | Workstream | Source | Have it? | Scope / date range | Coverage |
| 1 | Site crawl | Screaming Frog / Sitebulb / Semrush / Ahrefs / Lumar / Oncrawl / Botify, or your own fetches | crawl date, depth/URL limits, what was excluded | N URLs of ~M | |
| 2 | Google Search Console | Pages, Performance, Sitemaps, Structured data, Page experience, Manual actions, Links | which property (domain vs URL-prefix), date range | which subdomains/paths | |
| 3 | Analytics | GA4 / Adobe / other | date range, which segments, consent-mode gaps | which sections | |
| 4 | Backlinks | Ahrefs / Majestic / Moz / Semrush | export date, which provider(s) | deduplicated? | |
| 5 | Human review | Your reading of real pages | which templates | N URLs across T templates | |
| 6 | SERP review | What Google shows for target queries | date, location, device | N queries |
Three things this catches that a yes/no ledger does not:
- A URL-prefix property that covers only
/blog/looks like full Search Console access and is not. - A crawl that stopped at depth 5 makes everything deeper look orphaned.
- A 12-month analytics view with an active filter silently hides a section.
Also record: server log access, CMS/platform, and whether you can fetch the site live.
Never convert a sampled proportion into a site-wide count. "11 of 12 sampled product pages reuse the manufacturer description" is a finding. "3,847 product pages reuse the manufacturer description" is a fabrication unless you counted them. Report the sample size every time.
Ask for the missing exports once, in one message, then proceed with what you have. Do not block the whole audit on a Screaming Frog file. Run the workstreams you can, mark the rest NOT ASSESSED, and put "get these exports" at the top of the next-steps list.
For a large site, start the crawl first — the book notes crawls of large sites can take days — and work the other five workstreams while it runs.
Collect before you mark NOT ASSESSED
NOT ASSESSED means unobtainable, not not handed to me. Given only a URL and the ability to fetch pages, a large amount of the audit is collectable directly, and marking it NOT ASSESSED without trying is the lazy failure, the mirror image of inventing data. Do the collection pass first. Mark NOT ASSESSED only what survives it.
What a URL plus page fetches yields, with no paid tooling:
| Check | How |
| robots.txt rules, sitemap declaration | Fetch /robots.txt |
| Sitemap presence, size, sample URL validity | Fetch the sitemap; spot-check ~20 URLs for 200 / canonical / noindex |
| HTTPS, redirect chains, status codes | Follow redirects on the naked domain, www, http://, and a sample of URLs |
| Titles, meta descriptions, H1s, canonicals, meta robots, hreflang, viewport | Parse the <head> and headings of each sampled template |
| Structured data validity, rich result eligibility | Extract JSON-LD/microdata; Rich Results Test and Schema Markup Validator |
| JS rendering risk | Compare raw HTML to rendered DOM; check whether main content and nav exist in raw HTML |
| Core Web Vitals (field data where CrUX has it) | PageSpeed Insights, per template |
| Internal link structure, click depth | Crawl outward a few levels from the homepage yourself |
| Indexed page count, rough | site: queries — approximate, label it as such |
| SERP appearance, title rewrites, SERP features, competitors | Search the target queries |
| Content quality, expertise, E-E-A-T signals, navigation | Read the pages |
That covers most of dimensions 1, 2, 3, 7, 8, 9 and 10 in the rubric well enough to score them. What genuinely cannot be obtained without access or paid tools, and so is legitimately NOT ASSESSED:
- Manual actions and security issues — Search Console only. Never infer one.
- Real traffic and conversions — analytics only. A traffic estimate from a third-party tool is not the site's traffic; say so if you use one.
- Backlink profile — needs a link database. "The site looks low-authority" with no link data is exactly the invention this skill exists to prevent.
- Which queries the site actually earns impressions on — Search Console only.
- Crawl frequency and crawl budget — server logs only.
- Full-site totals on a large site (exact index count, how many URLs have duplicate titles) — needs a full crawl. You can still report rates from a sample, labelled as a sample with its size.
Sampling is honest as long as you say so. "11 of 12 sampled product pages reuse the manufacturer description" is a finding. "3,847 product pages reuse the manufacturer description" is not, unless you counted them.
Step 3: Run the six workstreams
Read references/audit-checklists.md. It holds the full per-workstream list of
what each source can detect, and it is the authority on what to look for. Work
through the workstreams you have data for.
references/technical-reference.md is the authority on what the findings
mean. Open it whenever you hit one of these, and do not work from memory —
several of the book's rules in it carry corrections:
| When you are looking at | Read |
robots.txt, noindex, canonical, redirects | "Crawl control: the four instruments" |
| Titles, meta descriptions, headings, images | The first three sections |
| Sitemap validity | "XML sitemaps" |
| JavaScript, rendering, missing content | "JavaScript and rendering" |
| CWV, INP, mobile, interstitials | "Page experience" |
| hreflang, multi-market setups | "International targeting" |
| A migration, replatform or redesign | "Migrations, redesigns and domain changes" |
A compressed view of what each one uniquely contributes:
- Crawl — the richest single source. Canonicals, broken internal links, robots.txt blocks, redirect chains, missing/duplicate titles and headings, missing alt text, orphan pages, click depth, thin pages, duplicate content signals, PDF over-reliance.
- Search Console — how Google sees the site, which no crawler can tell you. Indexing status, rendered JavaScript, sitemap errors, structured data errors, page experience, manual actions, top queries.
- Analytics — where traffic actually is and where it moved. Site-wide trend, per-section performance, pages with sudden drops, conversion paths.
- Backlink tools — authority relative to competitors, profile gaps, broken inbound links, toxic links.
- Human review — everything judgement-shaped: navigation quality, internal linking opportunities, content depth, subject-matter expertise, content written for search engines rather than users, E-E-A-T.
- SERP review — missed opportunities, SERP feature changes that ate your clicks, and whether Google is rewriting your titles and descriptions (a signal that yours do not match the query well).
Scale discipline
On a large site you cannot review every page. Sample by template, not at random: homepage, top-level category, sub-category, product/article detail, faceted/filtered URL, paginated series, search results page, author/profile page, location page. One template flaw multiplied by 50,000 URLs is nearly always a bigger finding than anything wrong with a single page.
Step 4: Audit the content
Technical fixes remove ceilings; content is what raises the floor. Run this as a distinct pass, per Ch. 10:
- Inventory all content (from the crawl if there is no easier source).
- Select what to review closely.
- Split into two buckets, which are judged differently:
- Informational (top and middle of funnel) — judged on breadth, depth, expertise, and freshness. Map competitor topic coverage and find the gap.
- Commercial / transactional — judged on conversion support, uniqueness (the manufacturer-description duplicate-content trap), and title/description optimization.
references/content-and-eeat.md holds the quality criteria, Google's own
low-quality content categories, and the E-E-A-T and YMYL guidance.
Step 5: Find the opportunity, and make the audit verifiable
Defects tell you what is broken. This step establishes what the site should be
winning and is not, and how anyone will know whether the work worked. Read
references/opportunity-and-measurement.md and produce two things:
The opportunity matrix. One row per query cluster: intent, the URL that should own it, business value, demand evidence, current position, who outranks you, and the specific reason you are losing. A row without business value and demand evidence is a keyword, not an opportunity — keep it out of the plan. That reference also covers choosing search competitors defensibly (they are usually not the businesses the client names), confirming cannibalization with performance data rather than shared title words, and turning a link gap into a concrete reclaim-and-pursue list.
The measurement check. Validate the analytics before trusting any number in this audit — tagging, consent gaps, conversion configuration, revenue reconciliation, bot filtering. Then record the baseline, and give every remediation item a metric, an expected direction, and a review window. An audit whose recommendations cannot be falsified is a set of opinions.
Step 6: Score
Read references/scoring-rubric.md. Score each applicable dimension 0–5 against
its anchors, with the business-type weighting it defines — those weights are
editable defaults, so adjust them where you know the business and say that you
did. Two hard rules:
- Every score cites evidence — the specific URL, report row, metric, or quoted page text. No score without a citation or an explicit "not present on any sampled page" note.
- NOT ASSESSED dimensions are excluded from the average and listed separately in the header. They never quietly count as passing. A composite score computed over three of eleven dimensions is reported as "3/11 dimensions assessed", not as a site score.
Step 7: Write findings and prioritize
Each finding gets:
- What — one sentence, specific. Not "improve meta descriptions" but
"412 product URLs share the identical meta description
Buy quality X online". - Evidence — source and citation, per Step 5.
- Scale — how many URLs, which templates, what share of the site.
- Why it costs traffic — the mechanism. If you cannot state one, cut the finding.
- Severity — P0/P1/P2/P3, per the firing conditions in the rubric. Severity is harm, never effort.
- Fix — concrete and assignable, with the team who does it (dev, content, SEO, leadership) and rough effort (S/M/L).
- How to verify — what changes, in which report, and roughly when.
Report findings by severity. Then give the plan a separate ordering by impact ÷ effort. A P1 that is one line in a template beats a P0 requiring a replatform, in sequencing terms — say so explicitly rather than handing over a list nobody can start.
Set expectations on timing
The book is blunt about this and the report should be too: after a fix, Google must recrawl and reprocess the affected pages. Depending on site size, page count, and crawl frequency this takes one week to several months. Do not imply next-week recovery. Say what to watch and when to expect movement.
Step 8: Deliver the report
Structure:
- Header — site, scope, date, business type, YMYL yes/no, and the evidence ledger with scope and coverage per source, plus the NOT ASSESSED list stated up front.
- Verdict — 3–5 sentences. The single biggest constraint on organic growth for this site right now, and whether it is technical, content, or authority. Derive it from the findings, not from the score.
- Scorecard — the rubric table, with dimensions assessed, the share of total weight they cover, and any weighting you changed and why.
- P0 and P1 findings — full detail as in Step 7, ordered by severity alone.
- P2 / P3 — condensed to a table.
- Opportunity matrix — the clusters worth winning, with business value and demand evidence. This is the growth half of the report; a technical-only audit tells the client what to repair and nothing about what to build.
- Sequenced plan — a separate ordering from the findings: triage active P0s immediately, then order by dependency and impact ÷ effort across 30 / 90 / 180 days, by owner. Say where an easy P2 is scheduled ahead of an expensive P0 and why.
- Data still needed — the exports and access that would close the NOT ASSESSED gaps, and which findings they would settle.
- Measurement — the analytics validation result, the baseline captured now, and per-item metric, expected direction and review window. Chapter 8's point: without a baseline, the next audit is a first audit again.
Write the report in the language the user is writing to you in.
Traps that produce a bad audit
- Listing every crawler warning as a finding. A crawl export has thousands of rows. Most do not matter. The audit's value is in the filtering, and a report that forwards the tool's output unfiltered has done none of the work.
- Grading against absolute numbers rather than the competitive set. "Domain Rating 42" means nothing until you know the competitors sit at 38 or 79.
- Treating tool scores as ground truth. Every vendor's "SEO score" is that vendor's opinion. Cite the underlying observation, not the score.
- Confusing correlation with a ranking factor. The book is careful here: for example E-E-A-T is not itself a ranking factor — it is what Google's quality raters assess, and Google approximates it with measurable proxy signals. Write it that way.
- Auditing for the crawler, not the user. Content produced primarily to rank rather than to help anyone is the exact pattern Google's helpful-content work targets. An audit that recommends more of it is recommending the risk.
- Auditing against retired tools and dead rich-result types. Several tools and result types named in the book no longer exist. The references mark them [UPDATE] — a recommendation to earn a rich result Google stopped showing cannot pay off, and it costs the report credibility with anyone who knows.
- Reporting a sample as a count. "11 of 12 sampled pages" is a finding. "3,847 pages" is a fabrication unless you counted them.
- Letting effort leak into severity. Severity is harm. An expensive P0 stays a P0; a free P2 stays a P2. Sequencing is where cost belongs.
- Silently skipping what you could not measure. See the rule at the top.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

