Claude for SEO: What Works, What Fails, What We Tested

Claude is the most useful SEO assistant I’ve worked with, and it will also cheerfully invent a search volume of 2,400 for a keyword it has never seen data on. Both things are true at once, and most articles about “AI SEO” only tell you the first one.

We run SkillProof, where we install Claude skills on a clean machine and test them against real work before publishing a verdict. Our SEO category currently holds five skills: two passed testing with scores attached, and three sit in the queue mid-experiment. This article is the workflow we actually use, built on those tested skills, with the failure modes left in. If you’re new to skills as a concept, read what Claude skills are first; this piece assumes you know the basics.

What Claude does well in SEO, and where it quietly fails

The honest split matters more than any tool list, so here it is.

Claude is genuinely strong at the reasoning layer of SEO. Content briefs are the clearest case: give it a keyword and two or three competing URLs, and it produces a brief with entity coverage, heading structure, competitor gaps, and questions to answer that would take a junior strategist half a day. It’s equally good at intent classification (sorting 500 keywords into informational, commercial, transactional, and navigational buckets with defensible logic), at schema markup, where the output is code that either validates or doesn’t, and at internal-link logic, because deciding which of your 200 pages should link to which is a language problem, and language problems are the one thing it was built for.

Where it fails: anything requiring live data. Ask Claude for the monthly search volume of “crm for plumbers” and it will give you a number. The number is fabricated. It has no access to Google’s index, no rank tracker, no SERP snapshot from this morning. The same applies to keyword difficulty scores, backlink counts, SERP features, and current top-10 results. Without a connected data source it produces plausible fiction, delivered in the confident tone of a consultant who bills by the hour.

The fix isn’t complicated. Bring your own data. Export from Search Console, pull a keyword list from whatever paid tool you already use, hand Claude a crawl file. Then let it do what it’s good at: reasoning over data you gave it. Every workflow in this article follows that rule. The moment you catch yourself asking Claude for a number instead of asking it to interpret a number, stop.

One more failure worth naming: freshness. Claude’s knowledge of Google’s algorithm updates ends at its training cutoff. If a core update rolled out last month, it doesn’t know. Treat its algorithmic commentary as history, not news.

The tested SEO stack

Every skill below went through our standard process: clean install, real task, scored verdict across install friction, trigger reliability, output quality, and documentation. Full rankings live at best SEO skills.

SEO Audit — passed, output 9/10, trigger 5/5. The workhorse. We pointed it at a 200-page content site and it covered crawlability, meta issues, Core Web Vitals, and thin-content detection. The part that earned the score: we ran a parallel Ahrefs audit on the same site, and the skill’s findings matched on every major item. It won’t replace a crawler for a 50,000-URL site, but for the sites most of us actually run, it finds what matters and prioritizes the fixes.

Programmatic SEO — passed, output 8/10. We had it design a 400-page template strategy from a keyword dataset. What impressed us wasn’t the scale, it was the restraint: the skill builds uniqueness safeguards into the template logic and is openly opinionated about not generating doorway pages Google will eventually burn you for. That opinion is correct and rarer than it should be.

Schema Markup — in the test queue. It generates and validates JSON-LD for product, FAQ, review, and breadcrumb types. We’re currently validating its output against Google’s Rich Results Test across 20 page types before issuing a verdict. Early signal is promising, which is why it appears in the workflow below, but we don’t score on early signal.

GEO / AI Visibility — in the queue, and this one needs the longest test. Generative-engine optimization is half discipline, half folklore right now, and nobody selling GEO services wants to admit which half is which. We’re running a three-week citation-tracking experiment to see which of the skill’s recommendations actually move citations in ChatGPT and Perplexity. Verdict when the data exists, not before.

Semantic SEO Suite — in the queue. Our crawler found it at zero GitHub stars, which normally means nothing good, but it has one design decision we respect enough to mention pre-verdict: a fabrication guard that refuses to invent numbers. Given the failure mode described above, a skill author who builds that in understands the problem.

Two skills from our marketing category round out the stack. Content Strategy passed with output 8/10; in testing it produced a topic-cluster map and a 90-day calendar grounded in actual search intent rather than “post consistently” filler. And Copywriting passed at 8/10 with a strict bar: two human copywriters rated its landing-page output shippable with light edits.

A worked workflow: keyword to published page

Here’s the pipeline we run for a single article, with the skill that fires at each step. Total hands-on time is about 90 minutes for a piece that used to take a day and a half.

Step 1: keyword and cluster. Input is a CSV export from your keyword tool plus your last 90 days of Search Console queries. The content-strategy skill triggers and maps the keyword into a cluster: what’s the hub, what are the spokes, which existing pages already partially cover it. This step exists to prevent the most expensive SEO mistake, which is writing a good article that cannibalizes another good article. Claude is unreasonably effective at spotting cannibalization risk when it can see your full URL list.

Step 2: the brief. Feed it the chosen keyword, the top three ranking URLs (paste the content or use a fetch tool), and your angle. The output is a brief with entities to cover, questions from the SERP, heading skeleton, and a differentiation note explaining what the ranking pages miss. I edit this brief every time. Usually 10 minutes of cuts. The skill over-includes, and over-inclusion in a brief becomes bloat in a draft.

Step 3: the draft. The copywriting skill fires here, ideally with a brand-voice file it can reference. This is where quality diverges wildly between teams: a draft against a tight brief with voice context needs one editing pass, while a draft from “write an article about X” needs a rewrite that costs more than writing fresh. The brief is the whole game.

Step 4: schema. The schema-markup skill reads the finished draft and emits JSON-LD: Article plus FAQPage if the piece has a question section, BreadcrumbList if your template doesn’t already handle it. Then validate in Google’s Rich Results Test anyway. Two minutes, and it has caught real errors for us, including a nested FAQ that would have silently failed.

Step 5: internal links. Hand Claude your sitemap or a URL-plus-title export and the new draft. Ask for links in both directions: where this page should link out, and which existing pages should now link in. The second half is the one everyone skips, and it’s worth more. A new page with zero inbound internal links is a page you’ve asked Google to ignore.

The content-quality trap

Now the uncomfortable section. The workflow above can produce a lot of content quickly, and that’s exactly the danger.

Google’s public position is clear and has been stable since early 2023: it rewards helpful content regardless of how it’s produced. Provenance isn’t the ranking factor, helpfulness is. What Google does penalize, explicitly, under its scaled content abuse policy, is publishing lots of pages primarily to manipulate rankings rather than to help anyone. Notice the shape of that rule. It doesn’t say “AI content is spam.” It says “content produced at scale with no regard for quality is spam,” and AI just happens to be the cheapest way to produce that.

So the trap isn’t detection. Nobody at Google is running your paragraphs through a classifier and penalizing you for the crime of using Claude. The trap is that unassisted model output regresses to the mean: structurally identical to a thousand competing pages, and empty of anything only you could say. Publish 200 pages of that and you haven’t tricked Google, you’ve told it precisely what your site is worth.

This is where skills earn their keep, because they raise the floor in specific, testable ways. A brand-voice skill (we tested Brand Guidelines; it works after you invest 30 minutes configuring it) keeps output from sounding like everyone else’s default Claude. E-E-A-T signals have to come from you: first-hand test results, real numbers, named authors, positions you’re willing to defend. Notice this article’s structure, it exists because we actually installed and scored these skills, and that layer can’t be generated. And a humanizing pass strips the tells (the inflated transitions, the vocabulary no human uses) that make readers, if not algorithms, discount your page on sight.

My opinion, stated plainly: if a page contains nothing you learned by doing the work, don’t publish it, whoever or whatever wrote it.

Measuring whether it worked

An AI SEO workflow you don’t measure is a content mill with better branding. Here’s the minimal measurement setup.

Tag every page produced by the workflow, a spreadsheet column is enough. In Search Console, compare that cohort against your older content on impressions, clicks, CTR, and average position. Give it time: 8 to 12 weeks before judging, because fresh pages oscillate. Annotate publish dates so you don’t attribute an algorithm update to your own brilliance.

For schema specifically, watch the Enhancements reports in Search Console. FAQ and product markup either produces rich results or it doesn’t, and CTR on pages that gained rich results versus those that didn’t is one of the cleanest before/after comparisons in SEO.

Two honest checks worth adding. Indexation rate: if Google indexes 40% of your new cohort, quality is your problem, not crawling. And a human check: have someone who knows the topic read three random pages from the cohort each month. If they wince, your measurement dashboard is lying to you by omission.

What we won’t automate

Link outreach. Full stop.

The pitch writes itself: Claude drafts a “personalized” email, a script scrapes 500 bloggers, the sequence tool fires. We tested a cold-email skill for sales and it passed, because a salesperson sending 15 researched emails to prospects who might genuinely want the product is a legitimate act. Sending 500 templated link requests to people who owe you nothing is not the same act at higher volume. It’s spam with better grammar.

It also doesn’t work anymore. Every blogger’s inbox is now wallpapered with AI outreach, response rates have cratered, and the links you do land this way come from sites that say yes to everyone, which is exactly the neighborhood you don’t want links from. Meanwhile the sender domain you burned belongs to your actual business.

Links worth having come from work worth citing: original data someone references, or a free tool that made their article better. Claude can help you build those assets. It cannot make strangers care, and a skill that promises otherwise would fail our test on principle before it failed on results.

FAQ

Is Claude actually good for SEO?

Yes, for the reasoning layer: briefs, audits, schema, intent classification, and internal-link planning, where our tested skills scored 8 to 9 out of 10 on output quality. No, for the data layer. It cannot see search volumes, live SERPs, rankings, or backlinks, and will fabricate them if asked. Pair it with your existing data sources and it’s a force multiplier; use it alone and it’s a confident liar.

Can Claude pull live keyword volumes or SERP data?

Not by itself. Claude has no connection to Google’s index or any rank tracker. You either export data from your existing tools and hand it over, or connect a data provider through an MCP server. Any volume number Claude produces without a source attached should be treated as fiction.

Will Google penalize my site for AI-written content?

Not for the AI part. Google’s stated policy rewards helpful content however it’s produced and penalizes scaled content created primarily to manipulate rankings. The practical risk is publishing high volumes of mediocre generic content, which fails on quality grounds regardless of authorship. Raise the floor with the brand-voice and humanizing steps described above, and supply the E-E-A-T signals only you can.

Which Claude SEO skills should I install first?

Start with SEO Audit, the highest-scoring skill in our SEO category (9/10 output, findings matched a parallel Ahrefs audit in our test). Add Content Strategy for planning and the Schema Markup skill for structured data. If you’re building pages from datasets, Programmatic SEO passed testing with real safeguards against doorway-page patterns.

How do I install these skills?

Most are a folder copied into ~/.claude/skills/, and the tested ones on SkillProof list exact install commands with any setup caveats we hit during testing. The full walkthrough, including where skills live on disk and how to verify they trigger, is in our installation guide.