Growth marketing · Agent workflow
Diagnose your signup or checkout funnel, then brief one experiment
ZeroTwo checks the tracking first, then compares each funnel step by date and cohort, lists what else could explain the drop, and drafts a prioritised experiment brief in Slack. A person approves it before anything changes.
It finds where and for whom a funnel leaks. It cannot see why, so every finding is a hypothesis to test.
- PostHog
- Adobe Analytics
- Slack
Four questions before a drop-off number means anything
Is each step an event with one written definition?
“Signed up” should mean the account was created, not that a button was clicked.
Has tracking been checked against the real product?
A drop between two steps is sometimes just an event that stopped firing.
Are you comparing like with like?
Same date window, same traffic mix, same device split, same attribution rule.
Could something else explain the drop?
A release, a campaign, a holiday or a bot wave can all look like a UX problem.
Start from a funnel with every event defined
A diagnosis begins with a table like this one for your own funnel: what each step means, how many people reached it, the rate from the step before, and how you confirmed the tracking. Rates without counts are not evidence.
| Step | Event definition | Visitors | Rate from previous step | Tracking check |
|---|---|---|---|---|
| Landing view | Entry page viewed, once per visitor per day | [count] | [rate] | Fires once per load. Internal and bot traffic excluded. |
| Pricing view | Pricing page viewed after a landing view | [count] | [rate] | Counts the same visit path in both analytics tools. |
| Signup started | Signup form opened or first field focused | [count] | [rate] | Fires on desktop and mobile web, and on the redirect path. |
| Account created | Server-confirmed signup success | [count] | [rate] | Matches the account records, not the button click. |
| Plan selected | A plan or product chosen at checkout | [count] | [rate] | Fires once per choice, not on every re-render. |
| Payment confirmed | Payment provider confirmation received | [count] | [rate] | Reconciled against payment records for the same dates. |
Template only. The counts and rates are blank on purpose. This is not a result from a real site.
Check the tracking before you analyse the funnel
The cheapest conversion fix is discovering that an event broke. The workflow runs this check first and reports it at the top of the brief.
Tracking QA checklist
- Open the funnel in the product and trigger every step yourself, on desktop and on a phone.
- Compare event counts with a source of record, such as account or payment records.
- Look for events that start or stop on a release date.
- Confirm the same event name means the same thing in each analytics tool.
- Exclude internal users, test accounts and known bot traffic.
One definition per step
Write the definition once and use it in both analytics tools. When two tools disagree on a count, the report shows both rather than choosing one silently.
Segment by date and cohort, and say what the cut cannot prove
Each cut separates one explanation from another. Each also has a trap. For calculation and cohort analysis beyond this workflow, see the AI agent for data analysis.
Date and release
Whether the drop starts on a particular day, such as a deploy, a price change or a campaign launch.
Trap Comparing a holiday week with a normal one.
Device and browser
A step that fails on one browser or screen size, which usually points at a bug rather than copy.
Trap Treating a small device group as a trend.
Traffic source
Whether a new campaign is sending a different kind of visitor, who convert differently by nature.
Trap Blaming the page for a change in audience.
New or returning
Whether the problem is first-time understanding or friction for people who already know the product.
Trap Mixing cohorts whose entry date differs.
Confounders to name in every brief
A recommendation with no confounders listed is a guess with a chart attached.
Attribution rules differ between tools
PostHog and Adobe Analytics can count the same visit differently. Pick one rule and state it before you compare.
Small samples exaggerate
A step with few visitors can swing wildly. Report the denominator next to every rate.
Many cuts invite a false finding
Slice enough segments and one will look broken by chance. Note how many you looked at.
Seasonality and promotions
A drop that matches last year’s same week may not be a problem to fix.
A fix and a bug look alike
A sudden drop at one step is a tracking or release question before it is a design question.
The output is one experiment brief, not a dashboard
Fields in brackets are for the owner to complete. The hypothesis below is an example of the format. It is not a finding from a real site.
Experiment brief · signup to payment
Illustrative format- Hypothesis
- If the plan price and billing date are shown before the card form, fewer new visitors will abandon at payment, because the drop is concentrated at that step.
- Evidence
- Step rate from plan selected to payment confirmed, [rate], on [date range], for [segment]. Tracking QA passed on [date]. No release in the window.
- Segment
- New visitors from [source], on [device].
- Primary metric
- Payment confirmed per plan selected.
- Guardrails
- Refund and cancellation requests, support contacts about billing, and revenue per visitor must not worsen.
- Duration and sample
- [decide before launch, not after seeing results]
- Decision rule
- Ship, iterate or stop, agreed in advance with the owner.
- Owner and review date
- [name], [date]
Prioritise by the evidence you trust, the size of the audience affected and the effort to ship, using a scale your team already agrees on. The agent can rank candidates, and the owner chooses.
How ZeroTwo runs the diagnosis
It runs before a scheduled growth review, or when a funnel step moves by an amount you decide matters. Slack carries the brief.
Pull the funnel and event data
Read event and funnel data from PostHog for the date window you set, and note which events and definitions were used.
Bring in the second source
Use an Adobe Analytics export you provide, so the two tools can be reconciled instead of trusted blindly.
Check tracking before analysing
Flag events that start or stop around a release, mismatched definitions and counts that disagree with your source of record.
Segment and rank the drop-offs
Compute step rates by date and cohort with denominators, then rank where the loss is largest and best supported.
Post the brief for approval
Send the experiment brief to Slack with evidence, confounders, and approve, revise or defer actions for the owner.
A starting prompt
Edit the bracketed values and the channel. It asks for the tracking check before the analysis, the denominators and the confounders, and it stops at the brief. PostHog is available as a connector. Check the connectors page for Adobe Analytics, and attach an export if it is not connected.
Before each weekly growth review, diagnose our signup-to-payment funnel for the last [28] days.
1. Pull the step events from PostHog and use the Adobe Analytics export I attach. List the event definition used for each step.
2. Run a tracking check first. Flag any event that starts or stops near a release, and any count that disagrees between the two tools.
3. Report each step rate with its count and denominator, split by week, device and traffic source.
4. List confounders: campaigns, releases, seasonality and small samples.
5. Propose one prioritised experiment with a hypothesis, primary metric, guardrails and a decision rule. Say what evidence is missing.
Post the brief to #growth for approval. Do not change anything on the site.Where this stops and the next workflow starts
This page owns funnel diagnosis and experiment hypotheses. Browse every workflow in agent use cases, or see how the Work agent and scheduled tasks run them.
- AI agent for data analysisFor the calculation and cohort-analysis work behind a funnel diagnosis, with deeper analytical capability than this workflow alone.
- Traffic Drop DiagnosisWhen the problem is fewer visitors, not fewer conversions per visitor.
- Content Performance AnalysisConnects blog and landing page engagement to pipeline and revenue.
- Cross-Channel PerformanceA daily Slack brief across analytics, ad platforms, CRM and email.
- AI for marketersThe role-level view for marketers. This page stays on funnel evidence and experiment hypotheses.
- Technical SEO for agenciesFor crawl and indexing issues, which are a different problem from conversion behaviour.
Questions about conversion diagnosis
What is an AI website conversion analysis agent?
It prepares a funnel diagnosis for a person to review. It reads event data, checks tracking and definitions, segments drop-offs by date and cohort, lists confounders, and proposes one prioritised experiment with a primary metric and guardrails. A person approves anything consequential.
Which systems does the workflow use?
The workflow is built around PostHog event and funnel data, a second analytics source such as an Adobe Analytics export you provide, and Slack for approval. Check the connectors page for current availability of each connector before you rely on it, and attach an export where one is not connected.
Why check tracking before looking at the funnel?
A large drop between two steps is often an event that stopped firing or changed meaning. Confirm the events against a source of record such as account or payment records before you design an experiment around the number.
Can the agent tell me why people drop off?
It can show where and for whom the drop is concentrated, and propose hypotheses. It cannot see intent. Treat the output as a hypothesis to test with an experiment, ideally alongside session recordings or user conversations you already have.
Can ZeroTwo change my site or launch the experiment?
The workflow stops at the brief. Changes to the site, tracking or experiment tool should stay behind a human approval gate with a named owner.
How do we measure whether the agent is useful?
Track how often the tracking check finds a real problem, how many of its hypotheses reviewers keep, how much time the weekly diagnosis takes, and how often a decision is reopened. Faster, better-supported decisions matter more than the number of briefs produced.
What happens if a connector or model fails?
The run should keep its source list, the current stage and the draft so a person can resume manually. Test this in the first run, because a useful workflow fails visibly rather than losing context.
Turn a leaking funnel into one reviewed experiment
100 free credits to start on the Free plan. Work agent access starts on Plus at $14.99/mo. See pricing.