Claude Code for SEO: organizing each client's context
Claude Code already runs audits, writes briefs and analyzes Search Console data over MCP, but it starts from scratch every session: the quality of its work depends on the client context you give it. Organize that context per client in three parts (approved decisions, stable references, tasks in progress), short and dated, and keep it in one place for every consultant on the account.
Claude Code has become part of the daily routine for many SEO consultants: it reads a site, queries Search Console, writes a brief, drafts a report. But every session starts with no memory of the client. Whatever the agent knows about the approved strategy, the priority keywords or the recommendations already made, you are the one who tells it. This guide covers what practitioners do today, where client context gets lost, and a simple way to organize it per client, solo or as a team.
What SEO consultants do with Claude Code
The guides published in 2026 describe fairly similar uses.
- Technical audits. Gautam Khorana details a five-phase audit: site crawl, Search Console data analysis, structured data validation, Core Web Vitals, then a prioritized summary. He relies on several MCP servers and tools (a driven browser for the crawl, a database to aggregate Search Console data, file access, Lighthouse).
- Search Console data over MCP. Trevor Lasn presents mcp-gsc, an MCP server that lets Claude Code pull search analytics, inspect URLs, check indexing and manage sitemaps. Access goes through OAuth authorization on a Google Cloud project.
- Content and internal linking. The StudioHawk agency describes five routines: content audit, queries ranking in positions 4 to 20, content plan before writing, a first draft in the brand’s voice, internal linking. It advises starting with prompts, then skills, before moving to Claude Code with MCP servers and automation.
- Third-party SEO tools. SE Ranking presents three workflows with its own MCP server: brief generation, a visibility report for AI answer engines, and backlink analysis turned into a client presentation.
In Claude Code, you add an MCP server with claude mcp add. Anthropic’s docs describe three scopes: local (this project, just for you), project (a versioned .mcp.json file, shared with the team) and user (all your projects, just for you). Remote HTTP servers may require an OAuth login, which you complete with the /mcp command.
All these guides stress one point: the quality of the output depends on the context. Khorana describes a 90-second context brief (business, audience, site structure, audit goals) that makes the difference between a generic checklist and an audit tailored to the client.
The problem: client context scattered everywhere
For a client you’ve worked with for a year, the useful context goes well beyond an initial brief:
- the approved strategy (pages to push, dropped topics, trade-offs between traffic and conversion);
- the priority keywords, and the ones you decided to set aside;
- the history of recommendations: what was proposed, accepted, rejected, shipped;
- the client’s guidelines: tone, banned terms, legal notices, pages not to touch, approval process.
With Claude Code, the usual answer is the CLAUDE.md file, which the tool loads at the start of every session. The docs provide several levels: a project CLAUDE.md shared through version control, a personal CLAUDE.local.md kept out of Git, and a user file for all your projects. On top of that sits auto memory, where Claude records your corrections itself. SE Ranking’s guide, for its part, recommends one folder per client so that context doesn’t leak from one client to another.
The system works, but it has limits once you manage several clients:
- Context lives in each folder. Each client has its own
CLAUDE.md, sometimes a notes file next to it, and nothing links the folders together. - Auto memory is local. According to the docs, it is specific to each repository, stays on the machine and isn’t shared across machines.
- The file keeps growing. The docs advise staying under 200 lines per
CLAUDE.md: beyond that, context costs more and instructions are followed less reliably. A year’s worth of recommendations won’t fit. - Decisions get mixed with guidelines. “Never edit the product pages” and “in March, we chose to merge the two category pages” don’t have the same lifespan, but they end up in the same file.
A per-client organization method
The idea is to separate what rarely changes from what changes often, and to write each item so that an agent, or a colleague, can find it and understand it without you.
Three types of items
| Type | Content | Pace | Example |
|---|---|---|---|
| References | Stable facts about the client | Changes a few times a year | Market, audiences, strategic pages, editorial guidelines, tool access |
| Decisions | Approved choices, with their reason | One entry per trade-off | “March: we drop “POS software” in favor of “touchscreen cash register”, too competitive” |
| Tasks | What’s left to do | Every day | “Rewrite the title tags of the 12 category pages, client approval before going live” |
A fourth category often comes in handy: pitfalls, meaning what has already gone wrong and must not be repeated (a redirect that broke tracking, a page the client doesn’t want changed).
What to record for each client
- A short reference sheet: business, region, audiences, pages that drive revenue, priority keywords (with the date of the last review), editorial guidelines, no-gos. This is what the agent must read before doing anything.
- A decision log, one entry per decision: date, what was decided, why, who approved it on the client side. You don’t rewrite an old decision; you add a new one that supersedes it.
- The history of recommendations, linked to decisions: proposed on, accepted or rejected, shipped on, observed result. This is what keeps you from re-pitching something the client turned down six months ago.
- The task list, with an owner and a due date, and what’s blocking (often a client approval).
How to structure it with Claude Code
Using only Claude Code’s own tools, a reasonable setup looks like this:
- one folder per client, with a short
CLAUDE.mdholding the reference sheet and standing guidelines; - separate files for decisions and recommendations, which the
CLAUDE.mdimports or points to (the@path/to/filesyntax loads a file at startup, but doesn’t make the context any lighter); - an explicit instruction in the
CLAUDE.md: re-read the decisions before proposing a change, and add an entry after every trade-off; - the client’s MCP servers (Search Console, SEO tool) declared at project scope if the whole team needs them, at local scope if they use your own credentials.
For agencies, Khorana suggests three levels: an agency-wide file (standards), one file per client (specifics) and per-engagement notes, updated weekly and archived at the end. He flags the main risk himself: the files drift if nobody reviews them regularly.
Several consultants on the same client
As soon as two people work on the same account, new questions come up.
- A single source. If everyone keeps their notes in their own
CLAUDE.local.mdor in their machine’s auto memory, two consultants’ agents won’t know the same things. Anything about the client must go in the shared file. - Who decided what. A decision log must say who made the call and when. With Git, the commit history provides that, provided everyone pushes their changes.
- Conflicts. Two consultants editing the keyword list on the same day have to merge their changes. Short, separate files limit the damage.
- Non-technical team members. A writer or project manager who uses neither Git nor the terminal won’t see the
CLAUDE.md. Yet they are often the one receiving the client’s instructions. - Confidentiality. Each consultant should only see the clients they work on. StudioHawk reminds readers never to paste anything into Claude that would breach a confidentiality agreement.
What about GEO?
Optimization for AI answer engines (ChatGPT, Perplexity, AI Overviews…) is moving into the same tools: SE Ranking, for example, describes a visibility report for these engines produced with Claude Code. Practices are still young and measurements vary from one tool to the next. For organizing context, the takeaway is simple: record the tracked queries, the measurement tool and the date of each reading, so you compare like with like from one month to the next.
Tavimi for shared client context
If your clients’ context has to be shared between several consultants and their assistants, a dedicated MCP server can replace the scattered files. Tavimi is the shared working memory for your team and your AI assistants, over MCP. You create one space per client; each invited person only sees the spaces they were invited to. Notes are filed by category (decision, pitfall, reference by default) and versioned, and each category enforces its required headings: an incomplete note is rejected. Tasks have an owner, a due date and a dated log, and when the agent is missing a decision, it asks the question on the task instead of guessing; a person answers from the web interface.
Tavimi connects to Claude Code, to Claude on claude.ai, to Claude Desktop and to any MCP client, over OAuth. It isn’t a project management or billing tool: it sits alongside your SEO tools. The SEO and GEO agencies page describes agency use, and How it works covers the details.
FAQ
What people ask us
Do I need one Git repo per client to use Claude Code for SEO?
It isn't required, but one folder per client is the most common practice in published guides: it keeps one client's context from spilling into another's. A Git repo adds history and lets consultants share the CLAUDE.md.
What goes in an SEO client's CLAUDE.md?
Whatever applies to every session: market, audiences, strategic pages, priority keywords, editorial rules and no-gos. The Claude Code docs advise staying under 200 lines; detailed decisions and history go in separate files.
Is Claude Code's auto memory enough to manage a client?
It helps a solo consultant: Claude records your corrections and preferences there. But it stays on your machine and isn't shared with your colleagues, so it can't replace a shared client file.
Can I give Claude Code confidential client data?
Check your contract with the client first. StudioHawk's guide recommends never pasting anything that would breach a confidentiality agreement and anonymizing sensitive data.
Sources
- How Claude remembers your project (documentation Claude Code, Anthropic), accessed October 3, 2026
- Connect Claude Code to tools via MCP (documentation Claude Code, Anthropic), accessed October 3, 2026
- Automatiser le SEO avec Claude Code : 3 workflows via le MCP de SE Ranking (SE Ranking), accessed October 3, 2026
- Claude Code for SEO: The Full Setup (StudioHawk), accessed October 3, 2026
- Google Search Console MCP for Claude Code (Trevor I. Lasn), accessed October 3, 2026
- Comment j'utilise Claude Code pour les audits SEO techniques en 2026 (Gautam Khorana), accessed October 3, 2026
- CLAUDE.md pour les agences : modèle à trois couches pour les portefeuilles clients (Gautam Khorana), accessed October 3, 2026