“Agency versus in-house” is a poor question until the missing capability is named. A company may need technical access, editorial production, strategic prioritisation, product knowledge, analytics, stakeholder coordination, or simply a reliable operating cadence. An agency can provide a team and external pattern recognition; an internal owner can provide context and authority. Compare the job and evidence, not the sales narrative.
1. State the work that is missing
List the decisions and recurring tasks: technical audit, information architecture, content research, page production, digital PR, analytics, developer coordination, local governance, reporting, or hiring. Name the market, sales cycle, current team, backlog, access constraints, budget, founder load, and expected decision speed.
If the problem is undefined, postpone the provider choice. Buying capacity for an unclear job makes both agency and employee look ineffective.
2. Compare authority and access
Record who can change templates, redirects, canonicals, links, copy, data, releases, and budgets. An agency without technical access may produce recommendations that cannot ship. An internal specialist without decision authority may create a backlog no one prioritises.
Score each model on access, approval path, stakeholder coordination, and ability to stop unsafe work. Require named counterparts and escalation rules rather than a generic “account manager” promise.
3. Define the quality contract
Set requirements for original research, source dates, subject-matter review, evidence, claim limits, accessibility, privacy, and rollback. Google’s people-first content guidance is a useful editorial boundary; it is not a ranking guarantee and does not tell you whether a vendor understands your audience.
Ask for the review workflow, examples of rejected work, ownership of source registers, and treatment of uncertainty. A provider who promises volume but cannot explain quality control is not offering a comparable scope.
4. Compare measurement and reporting
Define observations and outcomes separately: indexed pages, queries, impressions, clicks, CTR, technical defects fixed, qualified actions, opportunities, and mature revenue. The Search Console Performance report can provide search visibility dimensions within its scope; it cannot prove that an agency caused pipeline.
Specify report owner, cadence, filters, annotations, attribution window, unknown rate, and decision rights. Reject dashboards that hide missing data or use a proprietary visibility score with no reproducible denominator.
For behavioural measurements, GA4 event guidance can help define what a page interaction means. Keep events, accepted leads, opportunities, and revenue as separate layers. A provider should be able to show the handoff between them and the limits of any attribution model, not only a chart of activity. Ask for the raw definitions, sample records, and review notes behind the report before treating the chart as evidence of delivery.
5. Compare context and continuity
An internal team may hold product language, customer feedback, roadmap, and institutional memory. An agency may bring specialist experience, flexible capacity, and a broader comparison set. Both can fail if knowledge is locked in one person, a vendor portal, or undocumented meetings.
Require a shared backlog, decision log, source register, access inventory, and handoff documentation. Measure continuity by whether another person can resume the work, not by how many calls were attended.
6. Compare economics by operating model
Model fixed fees, internal salary and benefits, tooling, management time, onboarding, review load, developer opportunity cost, hiring delay, and exit cost. Separate one-time diagnosis from ongoing production and from coordination. Do not compare an agency retainer with a full employee cost without including the capacity each model actually provides.
Use scenarios: narrow specialist sprint, fractional lead plus internal executor, agency production with internal reviewer, or fully internal team. State which assumptions are uncertain and when they will be revisited.
7. Assess risk, incentives, and conflicts
Check ownership of accounts, code, content, data, creative, domains, and analytics. Ask how the provider handles overlapping clients, subcontractors, AI-generated drafts, confidentiality, regulated claims, and missed SLAs. Define what happens if the engagement pauses or a key employee leaves.
An agency may be incentivised by retainers or output; an internal team may be incentivised by headcount or activity. Set success around decisions and evidence: useful pages, technical reliability, qualified demand, accepted handoff, and mature outcomes.
8. Run a bounded trial and exit test
Choose one clearly scoped problem with a baseline, decision owner, acceptance criteria, and stop rule. Preserve the current state. Review actual deliverables, source quality, implementation speed, stakeholder load, reporting honesty, and evidence of outcome maturity.
Before signing, ask each model to describe the first 30 days, access required, artefacts left behind, rejected-work process, and exit steps. Stop if the scope depends on ranking guarantees, invented case studies, unclear data rights, or an outcome the team cannot measure.
9. Apply the delivery-model scorecard
| Dimension | Required evidence | Hold if | | — | — | — | | job | named decisions and recurring work | scope is “do SEO” | | authority | access, approvals, escalation | recommendations cannot ship | | quality | sources, review, limits, rollback | output volume is the only KPI | | context | product knowledge and customer access | knowledge remains tribal | | measurement | reproducible observations and outcomes | proprietary score hides the denominator | | economics | full capacity, management, tooling, exit costs | fee is compared alone | | risk | ownership, privacy, conflicts, continuity | assets or data are trapped |
The right model may be a combination: internal ownership with specialist support, or an agency with a named internal decision maker. Choose the arrangement that makes responsibility, evidence, and exit explicit. Ask for the raw definitions, sample records, and review notes behind the report before treating the chart as evidence of delivery. Keep the request in the evaluation file so every model is scored on the same proof.
How did this article land?
Choose one reaction. You can change it anytime.
