Website + CRO Agent

What this agent does

Closes the loop the website program opens. The October launch ships ~103 pages against locked baselines; this agent reads what the measurement says, proposes the next test or content fix with the evidence attached, drafts variants on the template system (which makes a variant a content pass, not a build), and runs the experimentation cadence — test queue, monitoring, results write-ups. It arrives in November, exactly when launch-month data is ready to act on and WM has exited.

Why this agent

• The launch creates more optimization surface than the team can work — every treated page has a baseline and a lift expectation; a two-seat web team plus a new backfill hire can't test at the pace the surface deserves.
• Post-WM capacity is the plan's known gap — WM exits Oct 31; launch-month iteration rides Verto hours and this agent is the standing capacity answer for conversion work after.
• The whole thesis is conversion — Why now on the website brief is falling traffic against rising goals; the build gets the site to launch, but the compounding returns come from the test-and-iterate loop this agent runs.
• The template system makes it feasible — variants on tokens and templates are cheap and safe; the agent's drafts inherit the same rails human work does.

Trigger & inputs

• Trigger — weekly test-planning cycle; continuous monitors on live tests; alerts when a treated page underperforms its baseline or a test hits significance.
• Reads — GA4 (post sign-off) and the warehouse for funnel truth; the locked per-page baselines from the website program; the experimentation backlog; template and component inventory from Storybook.
• Context — the fixed template contracts (what a variant may and may not touch), brand rules, and the experimentation program's methodology standards.

What it produces

• A ranked test queue — hypothesis, evidence, expected impact, and effort per proposed test, refreshed weekly.
• Drafted variants — copy and layout variants on the template system, staged behind flags, never live-edited.
• Results write-ups — what won, why we think so, and what ships — feeding Shipped & Wins and the next cycle.

How it works

• Ingest the week's page and test performance against baselines.
• Rank opportunities: traffic × gap × confidence × effort.
• Draft the top variants on the templates; package each as a test spec (hypothesis, metric, minimum sample, stop rules).
• Route to the owner for approval; ship approved tests behind flags via the pipeline; monitor to stop rules; write up results.

Guardrails & human-in-the-loop

• Everything behind flags, everything approved — no live-page edits, ever; tests ship through the existing build workflow with its brand review.
• Never-do list — never touches pricing, legal, or product-naming copy; never runs tests that conflict with a running test on the same surface; never calls a result before stop rules are met.
• Methodology gates — minimum sample and significance standards from the experimentation program; peeking is a violation, not a shortcut.
• Escalation — measurement anomalies (tracking breaks, bot storms) pause affected tests and page the owner + Data & Tracking.

Success metrics

• North star — measured lift on treated pages vs locked baselines, compounding across the queue; test velocity as the input metric.
• Quality bar — proposal acceptance rate on the weekly cycle; post-test review of called results.
• Guardrail metric — zero unflagged changes to live pages; zero methodology violations.
• Hours returned — test setup, monitoring, and reporting hours off the experimentation contractor and web team, on the program ledger.

Build plan

• V0 · Shadow mode (Nov) — proposes tests against launch-month data alongside the human-planned queue; proposals compared, none shipped. Exit: proposal quality judged by the CRO owner.
• V1 · Assisted (Nov–Dec) — owns the queue drafting + monitoring + write-ups; humans approve every test. Rides the V3 wave's cost controls and governance from day one.
• V2 · Standing loop (Q4+) — the permanent optimization cadence; approval gate on shipping stays, monitoring runs autonomous.

Dependencies & risks

• The backfill hire is the owner — the Sr Manager, Growth Marketing (TBH) owns CRO; until they land, the daily-owner rule is unmet on this agent by definition. [owner to complete]
• GA4 sign-off + locked baselines — no trustworthy measurement, no honest lift claims; the agent inherits Data & Tracking's timeline.
• Launch slippage compresses it — if the October launch slips, November is stabilization, not optimization; the agent's V0 slips with it.
• Experimentation contractor alignment — the plan-design-build-test contractor line and this agent must be one program, not two queues.