SEO process for agencies: a controlled workflow

Scale starts with a clear client context

Agency work is rarely slow because another tool is missing. It becomes slow when goals, access, findings, decisions and approvals live in different places. A crawler reports errors, Search Console shows demand, editors work in a document and client approval sits in email. At the end of the month, somebody has to reconstruct a story.

A controlled SEO process connects these objects per client. It prevents data, permissions and actions from being mixed across accounts. That foundation comes before agents or automation take on more work.

The minimum client context contains:

• accountable organization and contacts
• primary domain, markets and languages
• main business goal
• agreed service scope
• approved data sources
• write permissions and explicitly excluded systems
• roles for preparation, review and approval
• reporting and decision cadence

Phase 1: use intake as an operating contract

A good intake does more than collect credentials. It clarifies which decisions the agency may take and which remain with the client.

Goal and success criteria

“More visibility” is too open. A specific business context is better: more qualified enquiries for a service group in a defined market, better discovery of existing product documentation or less loss during a migration.

Connect the goal to observable signals. The agency still should not guarantee rankings or revenue. It can be accountable for a sound process, traceable decisions and agreed deliverables.

Scope and boundaries

Record the domains, subdomains, languages and page types included. Exclusions matter just as much: careers, press, account areas, dealer portals or systems owned by another provider.

Use the least permission necessary. Read-only roles are enough for analysis. Ownership or full write access should not be shared by default. Google documents separate Search Console roles. An agency does not automatically need the highest one.

Responsibilities

Every work type needs an owner:

• Who decides priorities?
• Who reviews technical changes?
• Who is accountable for brand and content?
• Who may approve production?
• Who resolves goal conflicts?

Small teams may assign several roles to one person. The responsibilities should still be named.

Phase 2: connect and verify sources

A green login is not yet a verified data source. Check for every connector:

1. correct property and canonical host
2. market, country and language
3. date range and freshness
4. access role
5. known gaps or data limits
6. time of the latest successful retrieval

Search Console, Bing Webmaster Tools, analytics, crawlers, CMS and repository answer different questions. Preserve the origin in every finding. A KPI export without its property and period is not defensible client evidence.

For large sites, bulk export can support longer term analysis. It does not remove accountability. Schema, partitioning, data quality and cost still need an owner.

Phase 3: turn evidence into a prioritized backlog

A finding becomes work only when the goal connection and action are clear. A useful backlog object contains:

• observed finding
• source and retrieval time
• affected URLs or templates
• connection to the goal
• proposed action
• uncertainty
• risk and required approval
• acceptance criteria
• later observation window

Separate defects, improvements and experiments. A blocked critical directory is a defect. A new information format is more likely an experiment. They need different expectations and reporting language.

Phase 4: establish an operating loop

A simple weekly rhythm can work as follows.

Check sources

Review new technical errors, material data changes and failed connectors first. Not every fluctuation becomes a task.

Prepare decisions

The agency prepares a small number of prioritized actions with rationale and exact scope. The client should be able to decide without rereading the entire audit.

Obtain approval

Approval applies to the visible change, not a general category. Broad or critical changes require additional technical or business review.

Execute and accept

After deployment, verify the technical state. The receipt records what was actually executed. Open or failed checks remain visible.

Observe impact

Search and business signals are reviewed in the window defined in advance. The agency reports observation and uncertainty, not activity alone.

Approvals clients can genuinely review

An approval request should be short enough to understand and exact enough to accept responsibility. It shows:

1. Why the action is proposed now.
2. Which URLs and systems are affected.
3. What changes between the before and after state.
4. Which risk exists and how reversal works.
5. What will be checked immediately after execution.
6. Which impact can only be observed later.

A long technical appendix may support this. The actual decision should not be buried inside it.

Reporting without activity theatre

A client report should keep four areas separate:

• Source state: which data was available and current?
• Decisions: what was prioritized or rejected, and why?
• Execution: which changes are complete with a receipt?
• Observation: which search, behavior or business signals changed?

Ticket counts, generated copy and crawled pages are production volumes. They may explain capacity, but they cannot replace impact.

A good report also states limits. Delayed Search Console data, a changed tracking event or an open observation window belongs in the conclusion.

Where agents can help

Agents are useful for repeatable preparation:

• checking source status
• structuring findings
• grouping URLs
• preparing content briefs
• assembling diffs and acceptance criteria
• flagging missing evidence
• starting regular observations

Accountability remains with named people. An agent should not review its own action and declare a business outcome without independent evidence. Critical writes must not be inferred from a general work instruction.

Client isolation is a security feature

With several clients, each context needs its own identities, sources, policies, actions and audit data. A filter in a shared spreadsheet is not a sufficient boundary. Technical tenant separation, permission checks and traceable exports belong to product maturity.

The current MarketingOS alpha represents one bounded local workspace. It is not a finished agency portal for many concurrent accounts. A pilot should therefore use one client, one accountable owner and one clear workflow. That constraint produces better evidence than an overly broad start.

The standard for a sound process

A scalable agency process does not simply create more tasks. It loses less context between analysis, decision and execution. Clients understand what they approve. The team can find a change later. Impact is described with the right source and period.

Once that chain works, more preparation can be automated. Without it, an agent only automates the disorder.