Triaging Web Analytics Support Agent Skill

Triaging web analytics support tickets

The job: turn a pile of open support tickets into (a) reply drafts grounded in code or data, and (b) draft PRs for real bugs.

Most reported "bugs" are explainable semantics; most real bugs show up in error tracking or raw data before they show up in the code.

Diagnose before writing code, and always determine which layer a symptom lives in before proposing a fix.

1. Enumerate the queue

Tickets live in the conversations product and are queryable via the PostHog MCP execute-sql tool against system.support_tickets (project 2, US).

Zendesk mirrors carry full comment history in the data warehouse.

See references/ticket-queries.md for ready-to-run SQL: open-ticket scans, keyword filters, full Zendesk comment extraction (the child_events JSON pattern), and resolving a requester email to an org/team across US and EU regions.

Slack channel #support-web-analytics mirrors new Zendesk tickets; the in-app ticket link in each message carries the conversations UUID.

2. Classify the shape, then run its playbook

Detailed walk-throughs with worked examples are in references/diagnostic-playbooks.md. The shapes:

| Shape | Trigger phrases | First move |

| -------------------------------------------- | --------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |

| Frontend crash | "everything crashes", exception ID, stack trace | Error tracking lookup; sourcemapped frames name the file. Check both US and EU projects |

| Two numbers don't match | "two different bounce rates", "insight X disagrees with tile Y" | Semantics first, not code: event-level vs session-entry scoping, "landing vs containing", any-event vs entry-event filters explain most of these |

| Count drop over time | "pageviews declined", "tracking loss" | Layer split: raw stored counts vs query-side exclusion. $pageview vs $pageleave ratio, UA segmentation, SDK version pin. Bot-shaped traffic disappearing is common and is not a PostHog bug |

| Tracker not loading / undercounts competitor | "numbers lower than <other tool>", GTM, consent, ad blockers | Runtime loading audit with Playwright against their live site: load method, first-request timing, blocklist simulation. See references/loading-audit.md |

| Ad-platform integration error | "can't re-add source", OAuth errors, "no conversions" | Source re-creation paths, OAuth failure modes (for example Microsoft AADSTS650052), attribution join keys (exact campaign name + normalized source, both UTMs required for the fallback) |

| Channel type misclassification | "shows as Direct", "wrong channel" | posthog/models/channel_type/channel_definitions.json + the decision tree in posthog/hogql/database/schema/channel_type.py; unknown source + stripped referrer falls through to Direct |

Two cross-cutting rules:

• Determine the layer before the fix. Capture → ingestion → stored events → query-time classification → UI. A drop in raw count() can't be caused by query-time bot exclusion; a classification change can't alter stored counts. State which layer the evidence points at.
• Check for prior art before building. Search open issues/PRs and the channel history; several recurring asks (self-referral exclusion, AI channel type, OAuth error surfacing) have open issues with context that changes the right response.

3. Produce artifacts

• Reply drafts: ground every claim in a file:line, a query result, or a doc link. Offer the customer the aligned filter/property instead of only explaining why they're "wrong" (for example: session $entry_utm_campaign instead of event utm_campaign).
• Fix PRs: one worktree + branch per fix, conventional commit, draft PR using the repo template. Public-repo safety: describe bugs generically; never include customer names, Zendesk numbers, or customer traffic volumes. Slack/ticket links behind auth are acceptable as origin context.
• Session note: keep a running triage note (.notes/) with one section per ticket and an explicit "action left" marker per ticket, so a human can pick up the queue.

4. Verification tools

• Runtime loading audits and traffic simulation: references/loading-audit.md.
• Production query-side checks (per-team event series, UA splits, ingestion warnings): the query-clickhouse-via-metabase skill covers prod-us and prod-eu access.
• Error tracking: MCP query-error-tracking-issues-list / query-error-tracking-issue-events with verbosity: stack gives sourcemapped frames.