Build, Agency, or Tools: Assembling GEO Capability

Three quotes sit on the table at a DTC portable-power brand’s Q4 planning meeting: an agency retainer billed per article, an AI visibility platform billed per seat, and a part-time content lead. All three are signable. The harder question is narrower: in March, which of these can you still explain to finance?

In GEO, buy throughput and keep judgment: market and budget calls, the buyer question chain, defensible facts, and the definition of done stay inside, while crawl audits, structured data, and sampling can be bought, provided every deliverable is scored against the same criteria.

Five capabilities, not one purchase

The most common negotiating error is pricing GEO as a single package. It is five capabilities that outsource with very different failure rates.

Capability	What usually happens when outsourced	Call	Acceptance artifact
Market and budget trade-offs	The vendor executes the brief; nobody makes the trade-off	Keep	One page: market priority and spend cap
Question chain and answer units	Drafts arrive; facts and differentiation do not	Hybrid	One page per question, with a four-slot answer unit
Crawling, structured data, feeds	Work is standardized and can be checked objectively	Buy	Audit sheet and validator-passing markup
Off-site media and communities	Relationships can be bought; framing cannot	Hybrid	Disclosure list and citation source table
Monitoring and interpretation	Tools produce numbers; reading them takes a person	Hybrid	Versioned prompt set and a monthly read-out

Rows two and five are where this goes wrong. A vendor will hand you a stack of answer units that all say “more premium, more durable,” because the vendor cannot see your margin, inventory depth, or certification language. Models do not quote that sentence and buyers do not believe it.

The second trap is ownership. In a hybrid setup with no named final fact-checker, you get copy written outside, edited inside, and claimed by no one — half the capacity gone before the first review.

Why the all-in-one retainer fails

Start with the invoice. Retainers bill per article, platforms bill per seat, projects bill per report. All three bill output. None bills “the AI answer names you.” Unless the contract carries that sentence, delivery stops at output.

Then the inputs. Which market first, which SKU to rewrite, which claim is safe — EU product-safety labeling, US substantiation, the return window — live in your margin sheet, return rates, and legal threads. A vendor without those writes generic benefits, which is what models are not short of.

Then the clock: between a rewritten paragraph and a changed AI answer sit repeated sampling and regional variance. Without a fixed prompt set and the four-layer AI visibility metrics, you cannot tell whether copy worked or the model shipped an update.

Finally, the quote. A bundled number hides which part failed; unbundled, you can see whether the audit, the writing, or the read-out earned its line.

The counterexample is familiar: six months of English posts, all on brief, no fixed question set, no before-and-after. The quarterly review reports volume and nothing else. That is not an execution failure — the definition of done was handed over on signing day.

Decide, define, buy, contract — in that order

Reverse the sequence and teams buy articles before they know which questions matter.

Step	Action	Owner	Passes when
1 Draw the line	Mark each capability keep, hybrid, or buy, with a named decision-maker	Founder and growth lead	Exactly one owner per row
2 Define the read	Pick 20–30 real buyer questions, fix region and language, state whether you score mention, citation, or recommendation	Growth and content	Question set is versioned and repeatable
3 Buy the work	Procure the crawl audit, structured data, and sampling against artifacts	Engineering and vendor	Scored with the same crawl audit checklist
4 Write the contract	Put artifacts, review cadence, and data ownership into the terms	Procurement and legal	Monthly read-out filed; raw data exportable

Step four’s three clauses matter more than price. The artifact decides what arrives; the cadence decides how fast you correct course; data ownership decides whether the question set and samples leave with you when you switch. Compare tools by reproducibility, not feature count: run the same URLs twice. A tool whose answers drift badly is not cheap at any price.

Example: one budget, two configurations

Example — staffing, output, and spend are invented; they are not market rates.

A portable-power brand sells in Germany and the US with a team of about a dozen. Option A buys the package: an agency retainer plus an annual platform license, eight long articles and one monthly report. Six months later the team can count articles but cannot explain a citation change; nobody owns the question set.

Option B stays hybrid: a half-time content lead in-house defines the question chain, writes answer units, and clears claims with legal; crawling and structured data go to a technical contractor, accepted against an audit sheet; monitoring is licensed, but the prompt set and monthly read-out stay in-house. Same budget, half the articles, each mapped to a question — and a quarter-end answer to “which questions moved from never mentioned to cited.”

The difference is not who is more professional. It is whether judgment sits with the person who carries the consequence. The package is not a failure and hybrid is not automatically better: Option A’s real defect is that no one owned the question set.

Where this breaks

Single-market test. One country, one category, a tight budget: a three-month package is reasonable, provided you keep the right to define the prompt set.

Marketplace-led sales. If most revenue closes inside a marketplace, owned-site content is not the priority and agency value sits in listings and reviews; do not build a content team for it.

Regulated categories. Medical devices, children’s products, supplements. Claims need legal sign-off, so outsourced copy must pass internal compliance. A package deal signs the risk away.

Two misuses. Buying a tool to answer a strategy question — tools show what the answer contains, not which market to enter. And reading “keep” as “write everything ourselves”: what you keep is judgment and acceptance; building capacity in-house is usually the expensive path.

The hidden cost: reclaiming judgment three months in without a question set or a baseline.

The cost side continues in the GEO budget model; the rest of the line is on the strategy and market hub.

This month, put the five capabilities on a single page, mark each keep, hybrid, or buy, and write a name beside it. Any row without a name is a contract not to sign yet.