Product roadmap · Roughly five months of focused work, in five phases
From a working demo to a weekly habit
Meridian is a weekly growth strategist for solo e-commerce owners. It reads a store, recommends a short list of credible actions, helps the owner do them, and reviews what changed without pretending to know why.
A working weekly loop exists today: findings can become tasks, tasks land on a calendar, results are logged and reviewed. The gap is not ambition — it is trust, capacity and data quality.
What every phase is judged against
01
Action over instrumentation
Every screen has to end in something the owner can do this week. New data is only added when it changes a decision.
02
Show the evidence, not just the answer
Each recommendation says where it came from and how confident Meridian is. Owners can correct what the app got wrong.
03
Honest about cause
Results are movement alongside work that was shipped — never proof. When data is too thin, the app says so instead of guessing.
04
Manual stays first-class
Typing a number and importing a CSV never become second-class paths. Connections are optional accelerators, not requirements.
Sequence and dependencies
- 01A first week that actually works
- 02Data worth trusting
- 03Read-only connections
- 04Close the loop honestly
- 05Make it a habit, and make it honest
- Phase 01 comes first and nothing runs in parallel with it: every later phase is judged against a first week that works.
- Phase 02 must land before 03, so connections inherit validation rather than duplicating it.
- Phase 04 can begin once 02 is stable, even while 03 is in progress — annotation only needs trustworthy numbers.
- Phase 05 is scheduled deliberately, not deferred: trust and habit work fails if it is treated as leftover time.
The phases
Each phase states what it changes, what “done” means, and the gate it must pass before the next one starts.
A first week that actually works
The product is only as good as the first seven days. Today a new owner has to complete a profile, wait for an analysis and interpret a lot at once. This phase makes the first useful week reachable and makes every recommendation explain itself.
- 01.1
A resumable setup checklist
Four explicit steps — store profile, baseline numbers, first analysis, first plan review — with progress kept, optional fields marked optional, and a route back to the checklist after each step.
Done when A new account always lands on the next unfinished step; skipping optional fields never blocks progress; no step dead-ends.
- 01.2
Evidence and confidence on every recommendation
Each finding and task states its source — your store profile, a number you logged, or a public page Meridian read — plus a confidence label separating observation from hypothesis.
Done when Every AI-created task renders a source and a confidence label; findings that rely on assumptions say so in the same block.
- 01.3
Capacity-aware weekly planning
Plan generation respects the owner's stated weekly time and budget. Work that does not fit is offered as a later option rather than filling the three priorities.
Done when A generated plan never exceeds stated capacity; the owner can override; overflow items are visible and clearly deferred.
- 01.4
No duplicate tasks from the same finding
Converting findings into tasks checks what is already in the plan or recently completed, and shows that state on the finding itself.
Done when Running the conversion twice adds nothing new; a planned finding is labelled as already in the plan.
- 01.5
One obvious next action
The Growth Hub leads with a single recommended next step and shows how it connects to its source finding.
Done when Exactly one primary action is rendered at a time, keyboard reachable, with a defined state when the week is complete.
- 01.6
Baseline capture during setup
Ask for this week's sessions, orders and revenue as part of onboarding, with a plain explanation of why — so the first review has something to compare against.
Done when Skipping is allowed; when present, the values appear on the first plan as the comparison baseline.
Phase gate
A new account with no history reaches a reviewed weekly plan in under ten minutes, and every recommendation can answer “why am I seeing this?”
Touches
Surfaces: Onboarding · Growth Hub · Growth Plan · Meridian AI
Data: Business · Task · WeeklyPlan · SeoIssue · GrowthBlocker
- Checklist steps can become a wall if they are mandatory — keep every field but the store name skippable.
- Confidence labels are only useful if they are honest; a label on everything reads as decoration.
Data worth trusting
Recommendations and reviews are only as good as the numbers behind them. Import currently accepts what it can parse and drops the rest silently, and entries cannot be corrected afterwards.
- 02.1
Import preview before anything is saved
Parse the file first, then show every row as valid, duplicate or rejected with the reason, and let the owner confirm the import.
Done when Nothing is written until confirmation; each rejected row names its line number and why.
- 02.2
Duplicate and conflict handling
Detect an existing value for the same metric and week and let the owner choose to replace or skip, per row or in bulk.
Done when Importing the same file twice creates no duplicate weeks; the choice made is reported in the result.
- 02.3
Correct and remove entries
Edit or delete a logged value, and show when each metric was last updated.
Done when An edited value is reflected on both Results and Performance without a reload; deletions ask for confirmation.
- 02.4
Coverage prompts instead of empty charts
When a metric has not been logged for two weeks, say so where it matters and link straight to the entry form.
Done when A missing fortnight produces a prompt with a link, never a silent flat line.
Phase gate
A fifty-row import containing seven bad rows saves forty-three, reports every rejection with a reason, and leaves no silent partial failure.
Touches
Surfaces: Results · Performance
Data: MetricEntry
- Validation that is too strict rejects plausible real-world formats — accept common date and number shapes.
Read-only connections
Manual entry is the honest starting point, but it is also the reason the loop decays in week three. Connections remove that work — deliberately after data quality is fixed, so a bad import and a bad sync are not fixed twice.
- 03.1
Connect with the scope stated in plain language
Each provider lists exactly what will be read, why, and what will never be written. Consent is explicit and revocable.
Done when No connection is created without a visible scope summary; disconnecting removes stored credentials.
- 03.2
Source attribution on every number
Metrics carry their origin — manual, CSV, or a named provider — shown on charts and cards.
Done when Every plotted series states where it came from; mixed-source weeks are distinguishable.
- 03.3
Sync freshness and honest failure
Show the last successful sync, a clear failure state, and a retry. Never fill a gap with an estimated value.
Done when A failed sync shows its last good timestamp and does not alter stored values.
- 03.4
Parity with the manual path
Everything available by hand stays available: entry, import, edit and delete continue to work alongside a connection.
Done when With every connection disconnected the app behaves exactly as it does today.
Phase gate
Disconnecting a provider stops syncing completely, and the product remains fully usable on manual data alone.
Touches
Surfaces: Results · Performance · Settings
Data: MetricEntry · new: MetricConnection
- Provider APIs differ enough that each connection is its own small project — ship one provider end to end before starting a second.
- Trust is fragile here: a single silent sync gap costs more than the time it saves.
Close the loop honestly
The current review compares completed work with metric movement. That is a useful signal, but owners learn most from a stated hypothesis, a review date and a written conclusion — including the actions that did not help.
- 04.1
Annotations on the trend line
Mark task completions, launches, discounts and events directly on the chart, with the date and a short note.
Done when An annotation is visible on the chart and readable as text without hovering; it appears on the matching week only.
- 04.2
Simple experiment tracking
A hypothesis, the action taken, the metric that will show an effect, and a review date — one small object, not an analytics suite.
Done when Each experiment records a result and can be referenced by a later plan; results are never presented as proof.
- 04.3
An editable weekly review
Draft the review, let the owner amend it, and carry the confirmed version into the next plan.
Done when The draft is editable before it is saved; each conclusion links back to its task and its metric.
- 04.4
Guards against false precision
Warn when a comparison rests on too few weeks or on non-comparable periods, and say what would make it meaningful.
Done when A one-week history never renders a percentage change without a caveat.
Phase gate
No screen states or implies causation, and every review line is traceable to one task and one metric.
Touches
Surfaces: Results · Performance · Growth Plan
Data: Task · MetricEntry · new: Experiment · new: Annotation
- Experiments can quietly become a second product — keep to a single measure and a single review date.
Make it a habit, and make it honest
The loop works; this phase makes it survive a busy week and be trustworthy to read. It also sets out plainly what Meridian analyses, what it stores and what it never claims.
- 05.1
Notification preferences that stick
Choose channel and frequency, with reminders linking straight to the task they refer to and a visible automation history.
Done when Turning a reminder off stops it; every reminder opens the exact item it mentions.
- 05.2
Search and filtering as the record grows
Filter tasks, findings, opportunities and past plans by status, category, source and date, with paging that does not silently truncate.
Done when Filters combine and persist while navigating within a section; counts match the filtered result.
- 05.3
Real empty and error states
Every surface explains what is missing and offers the smallest next step, and a failed save preserves what was typed.
Done when No screen shows a bare blank or a generic spinner; a failed submission can be retried without retyping.
- 05.4
Accessibility and mobile audit
Keyboard order, focus states, contrast, chart text summaries, touch targets and the tablet breakpoint between phone and desktop.
Done when The core weekly loop can be completed with a keyboard alone, and every chart has a text summary.
- 05.5
Data-use and AI transparency
One plain page covering what is read from public pages, what is sent for analysis, what is stored, and how to remove it.
Done when The page is reachable from Settings and from the footer; it names every category of data the product touches.
- 05.6
Demo is unmistakably demo
Sample figures, seeded findings and example tasks are labelled wherever they appear, and live accounts never inherit them.
Done when No demo number can be mistaken for the owner's own result, including on shared screenshots of charts.
Phase gate
Accessibility reaches AA on every app surface, the keyboard-only path completes the weekly loop, and a demo screen can never be read as live data.
Touches
Surfaces: Automations · Notifications · Settings · all surfaces
Data: Notification · AutomationSetting · User
- Polish work competes with new features for attention — schedule it as a phase, not as leftover time.
Deliberately not on this roadmap
- Billing, tiers and paywalls — the product stays free while the loop is proven.
- Team and client workspaces — the owner is one person; shared accounts are a different product.
- Native mobile apps — the responsive web app is the delivery surface.
- Attribution modelling and incrementality testing — Meridian reports signals, not proof.
- A recommendation template marketplace — advice has to stay specific to one store.
Scope is intentionally narrow. Meridian wins by being the tool an owner trusts on a Monday morning, not by covering every marketing channel.
See the loop this roadmap improves
The sample store shows today's behaviour, unedited — including the parts these phases set out to fix.