GEO-121·Issue 11·Strategy & Market·121 / 132
A 90-Day GEO Implementation Roadmap
A 13-week plan in three gates: make the store readable to AI, ship 12–20 answer units, then prove movement with owners, sign-offs, and one fixed metric.
Issue focusScale It: SOPs, Organization, and the 2027 Roadmap
On the first Monday of a budget cycle, a DTC espresso-machine brand’s growth team gets ninety days and one line from the founder: when a US buyer asks an assistant which machine to buy under $800, we should be in the answer. By Wednesday the catalog’s 132 articles fill a twelve-week schedule. In week nine someone opens a product page with JavaScript off: a spinner, no price. At week twelve, forty-one drafts have produced nothing a model quoted.
Ninety days does not buy a finished content library. It buys one working chain — a crawler that reaches the page, a page that says something quotable, a count that shows whether the answer moved — and the order is the plan.
Why the roadmap slips in week five
GEO has a dependency order: crawlable, then parseable, then quotable, then corroborated off-site, then transactable, then measurable. Most teams start at step three, because writing looks like progress. A draft produces an artifact by Friday; a crawl audit produces a spreadsheet nobody presents.
The first failure is the wrong opener, and the three popular ones all sit mid-chain: a monitoring tool with no pages worth watching, 132 drafts of one unquotable paragraph, structured data that gives contradictions a formal syntax.
Ownership is the second: inputs sit on four desks — engineering owns rendering, content owns pages, marketing owns mentions, support holds the real questions. When nobody owns “is the brand in the answer,” each function waits for the next.
The third is cadence: answers vary by hour, account, and region, so week five looks flat. Milestone it as “AI mentions us” and the team pivots after one sample; milestone it as “the deliverable is reusable” and gate two starts on time.
Three gates in thirteen weeks
Ninety days is thirteen weeks, not twelve — the only arithmetic that fits three sign-offs. The phases are gates, not parallel streams: if gate one fails, gate two repairs and ships nothing new.
| Gate | Weeks | Deliverables | Owner | Passes when |
|---|---|---|---|---|
| 1 Visibility | 1–4 | Crawl audit, crawler policy, entity sheet | Engineering | Scripts off, title, price, specs still read |
| 2 Answer assets | 5–9 | Question chains, answer units on 12–20 pages, structured data | Content | Page hidden, four sentences still say what it is, why trust it, who it is not for |
| 3 Measurement | 10–13 | 30-prompt baseline, citation log, retro | Growth | Same prompts at fixed hours give a trend, not a screenshot |
Gate one tests two things: does the core product page still yield its fields without scripts, and are major crawlers blocked by accident. Use the crawl audit instead of inventing a bar.
Gate two is depth, not page count: twelve to twenty pages is what ninety days can rewrite — five product, five comparison, five category pages, each with one answer unit and one spec table in consistent units. Teams spread across a hundred pages usually have none passing the extraction test.
Gate three builds a baseline before anyone argues attribution, using the AI visibility baseline method. Pick the first market in week one: without it, pages and evidence have nowhere to land.
Name each gate’s signer before week one, and make it someone who can move headcount: M1 signs the crawl audit at week four, M2 the answer units at week nine, M3 the baseline at week thirteen. A failed gate sends the next phase into repair mode, and M3 judges whether the method repeats, not whether GEO works.
Example: ninety days at a DTC espresso brand
Example — brand, team, and figures are invented to show sequencing, not results.
Six people: one engineer, two content, one growth lead, one support lead, one owner, across the US and Germany. Weeks 1–2 pull forty real questions from support tickets and site search into six question chains, and fetch thirty core URLs with scripts disabled. Weeks 3–4 fix three things only — variant pages returning an empty shell, price rendered client-side, one retrieval crawler stopped by the WAF; crawler policy follows vendor documentation, which still moves.
Weeks 5–9 ship fifteen pages, each with one answer unit and one spec table in consistent units, plus price, stock, and warranty in structured data — Germany’s two-year statutory warranty kept as its own sentence, not written over the US thirty-day return window. Weeks 10–13 run thirty prompts at the same hour weekly, logging mention, cited page, and position.
Four ways the schedule dies early
| Cause | Looks like | Defuse it |
|---|---|---|
| Waiting on front-end | Copy final; template untouched | Sprint the crawlable text block in week one |
| Waiting on master data | Two prices; localized versions contradict | One value per field, not real-time sync |
| Waiting on approval | Off-site pieces and legal wording queue | Ship owned pages first; off-site after gate three |
| Drifting target | No mention in week five, so direction changes | Sign deliverables; review effect quarterly |
The most expensive is not the first. The fourth costs no hours and erases the previous nine weeks.
Where this breaks
First, regulated categories: medical and financial claims carry compliance cycles longer than ninety days, so schedule to the license first. Second, a business that sells mostly inside a marketplace, where listings and reviews feed the platform’s AI, so the website can wait. Third, no engineering capacity — change the fields your hosted platform exposes and stop promising schema and CI checks. Fourth, treating ninety days as a performance guarantee: you can deliver a system and a baseline, not a promise that an assistant recommends you by February. Fifth, no single market — two markets mean three gates at half depth and every sign-off becomes theater.
This piece opens the November line, which covers content QA, build-time checks, the multilingual pipeline, a crisis playbook, and a scorecard for management. Why the entry layer comes before writing is in how AI search split e-commerce traffic into four entry points; to step up a level, start at the strategy and market hub.
This month, do one thing: cut the ninety days into thirteen weeks and let week four deliver exactly two artifacts — a list of pages whose price survives with scripts off, and a reconciliation sheet where the brand’s name, category, and description match across site, marketplace, and knowledge base. If those fail, do not commission the 133rd article.