- Delegate any SEO task whose output you can verify: audits, GSC analysis, clustering, schema. Keep every task you have to trust: strategy, priorities, production.
- Claude proposes and humans approve. Every audit skill in the library outputs findings for review, never auto-applied changes.
- From 500+ campaigns: teams get burned by delegating judgement or by refusing to delegate process. Both waste the expensive resource.
After running Claude across 500+ client campaigns at StudioHawk, the delegation line is clear. Here is the matrix: what to hand over completely, what needs expert review, and what never leaves human hands.
Delegate to Claude any SEO task that follows a checkable process on data you supply: crawl analysis, audits, clustering, briefs, schema generation, title rewrites. Keep humans on strategy, prioritisation trade-offs and anything that ships to production. The line is process versus judgement, and it is sharper than most people expect.
The delegation matrix
| Task | Verdict | Why |
|---|---|---|
| Technical audits from crawls or live fetches | Delegate | Checkable steps, verifiable findings. The skill even lists what it could not check |
| GSC analysis: cannibalisation, decay, update triage | Delegate | The data is on the table, so every finding can be verified against it |
| Keyword clustering and intent mapping | Delegate | Pattern work at scale, exactly what models are good at |
| Schema generation | Delegate | Validate before shipping, which the skill builds into its checklist |
| Content briefs and drafts | Review | Structure delegates well. Claims, tone and positioning need your eyes |
| Anything citing live rankings or search volumes | Review | Fabrication risk lives here. Verify against a tool you control |
| Publishing changes to production | Gate it | Machine proposes, human approves the diff. Claude Code stages commits for exactly this reason |
| Strategy, budgets, what to do first | Keep | Trade-offs need context Claude does not have: your business, your risk, your politics |
| Client and stakeholder judgement calls | Keep | Accountability cannot be delegated to something that cannot be fired |
The rule underneath the matrix
Delegate tasks whose output you can verify. Keep tasks whose output you have to trust.
An audit finding says "line 7 of this file carries the wrong meta description", and you can look at line 7. A strategy call says "spend the next quarter on content, and skip the migration", and the only verification is hindsight. Claude earns the first kind of task completely. The second kind stays yours.
What this looks like in practice
When we ran the Claude Code SEO skill against my own site, it audited 15 page types and returned a 90/100 with one genuinely embarrassing finding: a case study shipping another article's meta description. Delegated task, verifiable output, human decides the fix order. That is the whole model working.
The same split shows up in every skill in the library: audit skills output findings for review, never auto-applied changes. That design is deliberate. The skills carry the checklist. People carry the judgement.
The failure mode to avoid
The teams that get burned do one of two things. They delegate judgement (letting a model pick strategy because it sounds confident), or they refuse to delegate process (paying seniors to eyeball ten thousand crawl rows a model triages in minutes). Both waste the expensive resource. Get the split right and each side does what it is actually good at: the setup guide shows the working loop end to end.
FAQ
Can Claude do SEO on its own?
No, and be suspicious of anyone selling that. Claude runs process work at remarkable speed: audits, clustering, briefs, schema. It does not know your business, your risk tolerance or yesterday's algorithm update. The working model is Claude produces findings, a human makes calls, and nothing ships unreviewed.
Will Claude make up SEO data?
It can, which is why the delegation matrix matters. Analysis of data you supply is reliable because the numbers are on the table. Claims about live rankings, search volumes or competitors it has not fetched are where fabrication risk lives. Never delegate a task whose output you cannot verify against a source you control.
Is it safe to let Claude change my website?
Only through a review gate. Claude Code stages changes as git commits you approve, which is the right pattern: the machine proposes, you dispose. Anything that auto-publishes without a human diff review is a production incident waiting for a slow Tuesday.
What should a junior SEO learn if Claude does the process work?
Judgement, faster. The checklist knowledge lives in the skills now, so juniors climb to the interesting part sooner: which finding matters most, what trade-off to make, when the data is lying. The skills carry the checklist. People carry the judgement.
Run one delegate-column task end to end today: point Technical SEO Audit at your domain, verify three findings yourself, and you will trust the matrix because you tested it.