| name | online-marketing-graph |
|---|---|
| description | A graph-structured operating system for marketing work — diagnoses which barrier is actually blocking a result, routes to the right discipline (consumer psychology, UX research, behavioral economics, brand psychology, market research, experimentation), and enforces evidence gates before anything ships. Use whenever the user works on marketing, growth, conversion, funnels, landing pages, ads, copy, messaging, positioning, branding, pricing, onboarding, activation, retention, churn, A/B tests, customer research, personas, or user behaviour — including questions like "why aren't people converting", "how do I get more signups", "review my landing page", "what should this say", "how should I price this", or "who is my customer". Also use for marketing plans, campaigns, funnel audits, and growth strategy. Apply it even when the request sounds like a quick copy or design tweak, because the commonest failure in marketing is fixing the wrong barrier fluently — the graph finds the real one first. |
A graph, not a checklist. Nodes are units of work with typed inputs, outputs, and exit conditions. Gates stop bad work from propagating. Routing is conditional — the barrier type determines the path.
Core premise: most marketing failure is not bad execution. It is well-executed work aimed at the wrong barrier. So diagnosis is gated ahead of design, and design is gated ahead of shipping.
Standing constraint (never bypassed): every mechanism must survive the customer understanding it. Persuasion addresses reasons; manipulation bypasses them. A tactic that only works in the dark — hidden costs, manufactured scarcity, obstructed cancellation, confirmshaming — is refused, and an alternative is offered instead. This is not decoration: such tactics reliably show up as a conversion win and a retention loss, measured before the damage lands.
Do not run all nodes. Enter at the node that matches what the user brought, run forward through the gates, and stop when the question is answered. Most requests touch three or four nodes.
State the diagnosis before proposing anything. If the user asked for a tweak and the barrier is elsewhere, say so in one sentence, then help with both: the thing they asked for and the thing that's actually blocking them.
N0 Intake ──► G0 open decision? ──no──► say so, answer cheaply, stop
│yes
▼
N1 Instrument check ──► G1 data trustworthy? ──no──► fix measurement first
│yes
▼
N2 Audience & job map
│
┌───────────┼───────────┐ (run in parallel)
▼ ▼ ▼
N3a Friction N3b Barrier N3c Trust
(UX) (behavioral) (brand)
└───────────┼───────────┘
▼
N4 Merge & triage
│
▼
R Route by barrier type
┌────────┬───┴────┬────────┐
▼ ▼ ▼ ▼
N6 Friction N7 Msg N8 Trust N9 Pricing
└────────┴───┬────┴────────┘
▼
G2 testable at this traffic? ──no──► N11 alternative evidence
│yes
▼
N12 Run ──► N13 Verify ──► G3 clean? ──no──► back to N1
│yes
▼
N14 Attack ──► G4 survives? ──no──► back to R
│yes
▼
N15 Ship
│
▼
N16 Write back ──► N17 Holdout & drift watch
Full procedures, exit criteria, and output formats are in references/nodes.md. Read it when running more than two nodes. The table below is enough for a single-node request.
| Node | Question it answers | Read this reference |
|---|---|---|
| N0 Intake | What decision is open, and what would change it? | market-research.md |
| N1 Instrument check | Can these numbers be trusted at all? | cro-experimentation.md |
| N2 Audience & job map | Who is this for and what are they hiring it to do? | consumer-psychology.md |
| N3a Friction pass | Where does the interface fight the user's head? | ux-research.md |
| N3b Barrier pass | Is the block ability, motivation, or prompt? | behavioral-marketing.md |
| N3c Trust & meaning pass | What does the brand mean, is it recalled, is trust real? | marketing-psychology.md |
| N4 Merge & triage | Which barrier actually costs the most? | nodes.md |
| N6 Friction removal | What do we remove or default? | ux-research.md + behavioral-marketing.md |
| N7 Message & offer | What do we say, to whom, at what stage? | behavioral-marketing.md + marketing-psychology.md |
| N8 Trust & brand asset | What costly signal do we emit? | marketing-psychology.md |
| N9 Pricing architecture | How is the choice structured? | behavioral-marketing.md |
| N11–N13 Evidence | Can we prove it, and do we believe the result? | cro-experimentation.md |
| N14 Attack | What kills this conclusion? | reasoning-standards.md |
| N16–N17 Write back | What did we learn, what expires when? | nodes.md |
Gates matter more than nodes. When time is short, run the gates alone — never a truncated node sequence, because truncation removes gates first and that is exactly backwards.
G0 — Is a decision actually open? Ask what they will do differently depending on the answer. If nothing changes either way, this is a request for reassurance. Say so kindly and answer cheaply instead of building a research programme around it.
G1 — Is the instrument trustworthy? Before analysing any funnel number: what does the metric count, does it reconcile against a source of truth (payments, database rows), is bot and internal traffic filtered, does identity survive the flow? Any figure that looks dramatic is a data-quality alert until proven otherwise. Fixing tracking outranks every optimisation downstream of it.
G2 — Is this testable at this traffic? Estimate the minimum detectable effect before proposing an A/B test. If the plausible effect is smaller than what the traffic can detect, the test cannot answer its question. Route to N11 instead: painted-door test, qualitative sessions, sequential change with a control cohort, or an explicit judgement call labelled as one.
G3 — Are the health checks clean? Sample ratio mismatch, segment gradient, time-slice stability, guardrail metrics, funnel consistency. Check these before looking at whether the result is good, so the check isn't influenced by the outcome. A failed health check routes back to N1, not to a re-run.
G4 — Does it survive attack? Novelty, segment cannibalization, retention shadow, regression to mean, winner's curse — and the daylight test. State the mechanism in plain words as you would to the customer. If it describes bypassing their judgement rather than informing it, it fails, and the barrier was probably misdiagnosed: route back to R.
Never present findings at uniform confidence. Tag each claim C1–C4, using the normalisation table in references/evidence-tiers.md:
- C1 Established — controlled, powered, replicated
- C2 Supported — real evidence, single path, unreplicated
- C3 Indicative — directional, correlational, stated preference, borrowed from literature
- C4 Assumed — judgement, internal consensus, projection
Two propagation rules:
- Confidence floor — a recommendation is only as strong as the weakest thing it assumes. Say so when the floor is low.
- Expiry — date findings and name what invalidates them. A finding with no expiry trigger is treated as C4 within a year.
Three phrases that must never appear unlabelled: a specific number, a claimed mechanism, and a statement about what users want.
Lead with the answer. Then reasoning. Then risk. Never the reverse.
[One-sentence verdict or recommendation — actionable without reading further]
Diagnosis: [which barrier, from which node, with evidence tier]
Mechanism: [why this works — the psychological chain, labelled as inference if it is one]
Do this: [specific action, or required property rather than a prescribed pixel]
Risk:
- Evidence tier and what it rests on
- What would make this wrong
- Kill metric / expiry trigger: [the specific number or event that means stop]
For multi-node work, add a severity-ranked list rather than an unranked findings dump. Rank by cost × frequency × recoverability × silence — not by how often something was mentioned. The loudest problem is usually the cheapest one.
Say so plainly rather than running it anyway:
- Below testability traffic (roughly under a few hundred conversions a month) — G2 fails everywhere; use judgement plus qualitative work and ship.
- Pre product-market fit — only N2 and N3c are useful; the rest optimises an exchange that does not exist yet.
- Cheap, reversible decisions — if shipping and watching costs less than diagnosing, ship and watch.
- Solo operator — the parallelism has no value; run the gates and skip the ceremony.
These look like competence:
- Selling motivation to the already-motivated. The reflex diagnosis is "they need more persuasion"; the barrier is usually ability. Check for returning users, repeat cart visits, and stalls before proposing a discount or a bolder headline.
- Tactic-first design. Reaching for scarcity, social proof, or loss framing because they are famous. Barrier → mechanism → tactic, never the reverse.
- Uniform confidence. Presenting a survey result, a hunch, and a controlled test in the same declarative voice.
- Prevalence from small samples. "Most users struggle" from five sessions. Give denominators; qualitative work establishes that a problem exists and why, never how common.
- Correlation as cause. "Users who do X retain better, so push everyone to X" — almost always selection.
- Exhaustive lists instead of ranking. Twelve findings with no severity order exports the job back to the user.
- Optimising the visible step rather than the step losing the most people in absolute numbers.
- Removing protective friction. Ask what a hesitation was doing before smoothing it away. Some friction is the design working.
Read on demand — do not load everything.
| File | Read when |
|---|---|
references/nodes.md |
Running more than two nodes, or need exit criteria and output templates |
references/evidence-tiers.md |
Labelling findings, or reconciling evidence across disciplines |
references/consumer-psychology.md |
Segments, jobs-to-be-done, what the customer is actually buying |
references/ux-research.md |
Usability, cognitive friction, session design, why a step is hard |
references/behavioral-marketing.md |
Interventions, choice architecture, pricing structure, testing strategy |
references/marketing-psychology.md |
Brand meaning, distinctive assets, trust, persuasion, advertising |
references/market-research.md |
Sampling, surveys, message testing, market sizing, competitive work |
references/cro-experimentation.md |
Test design, power, validity checks, analytics integrity |
references/reasoning-standards.md |
Attacking a conclusion, verifying a claim, structuring the answer |