200 skills · 1230 min

The last two rows are per the editor’s own help text (“Doesn’t work with the mobile SDKs for now” / “Not available for the mobile SDKs at the moment”) rather than a per-SDK source audit.

Consequences worth memorizing:

• surveyPopupDelaySeconds is web-only. If a mobile ticket blames the delay, it’s a red herring.
• Partial responses and auto-submit are web-only too. On mobile a survey always stores one response at the end, and a rating tap never self-submits. Don’t carry a web diagnosis onto a mobile ticket.
• URL targeting is effectively web-only. Mobile SDKs decode the field but never enforce it; React Native filters those surveys out entirely. A mobile survey with a URL condition behaves as “no URL condition” (mobile/flutter) or “never shows” (RN).
• Android ships no survey UI. The app (or the Flutter plugin) must provide a PostHogSurveysDelegate. “Survey never renders on Android” is often a missing delegate, not a PostHog bug.
• Flutter is hybrid: triggering/eligibility runs in the native iOS/Android layer; rendering is Dart (SurveyService.showSurvey → showModalBottomSheet). It does not “just call native” for UI. So a Flutter rendering bug lives in Dart; a Flutter eligibility bug lives in native.
• React Native wait period is silently disabled (the check is commented out). Don’t blame the wait period on RN.

For a deeper version-by-version capability audit, see the survey-sdk-audit skill if available.

How a survey actually gets shown (the web eligibility pipeline)

The web SDK is the most complex and the most common in tickets. Mental model from packages/browser/src/extensions/surveys.tsx (checkSurveyEligibility) — checks run in order, first failure wins:

1. isSurveyRunning — has start_date, no end_date.
2. survey type is in-app (Popover / Widget / API).
3. linked_flag_key enabled (if set).
4. targeting_flag_key enabled (if set) — customer-defined property targeting.
5. _internalFlagCheckSatisfied — the auto-generated internal targeting flag.
6. hasWaitPeriodPassed — seenSurveyWaitPeriodInDays vs localStorage.lastSeenSurveyDate.
7. getSurveySeen — per-survey seen flag.

Then in getActiveMatchingSurveys: URL/device/selector match, event/action trigger fired, and flag re-check.

Two non-obvious facts that drive real tickets:

• The server returns ALL non-archived surveys (SurveyViewSet, products/surveys/backend/api/survey.py). It does not pre-filter by the internal targeting flag. All eligibility is client-side. So you cannot conclude “the backend excluded them” — the SDK did.
• The wait period has TWO independent implementations. canActivateRepeatedly (true when schedule: 'always') short-circuits _internalFlagCheckSatisfied (step 5) — so always bypasses the internal flag, including its $last_seen_survey_date rule. But hasWaitPeriodPassed (step 6) reads localStorage.lastSeenSurveyDate directly and is NOT bypassed by canActivateRepeatedly. So a schedule: 'always' survey with seenSurveyWaitPeriodInDays: 30 still enforces the 30-day wait via the localStorage path. lastSeenSurveyDate is updated whenever any survey is shown, regardless of completion. The same short-circuit also drops the internal flag’s $survey_responded/<id> is_not_set rule, so schedule: 'always' lets one person respond repeatedly — check uniq(distinct_id) against count() on survey sent before reading a response count as a respondent count.

How a response actually gets stored

Showing a survey and storing a response are separate pipelines. “Wrong data” tickets are about the second one, and its UI copy is genuinely misleading — reason from the code, not from the label.

“Response collection” maps to enable_partial_responses

The radio in frontend/src/scenes/surveys/SurveyResponsesCollection.tsx. “Any question: when at least one question is answered…” is true; “Complete survey: the response is stored when all questions are answered” is false.

• true -> sendSurveyEvent fires on every onNextButtonClick, so one survey sent per question answered, all sharing a $survey_submission_id with $survey_completed running false ... true.
• false -> one survey sent event, at the end of the path.

The setting controls capture, not response visibility. The results include every survey sent and every survey dismissed or survey abandoned marked $survey_partially_completed = true. buildMergedSubmissionsSubquery in frontend/src/scenes/surveys/utils.ts merges the latest answer per question for each submission ID; events without an ID stay separate. Raw event counts can exceed response counts. A completed event takes precedence over earlier partial closures. Otherwise the outcome chip reads “Dismissed” when the latest response event is a dismissal, and “Abandoned” for other unfinished submissions. “Abandoned” does not prove a page-unload event or inactivity timeout occurred; a later completion updates the same row.

The defaults disagree by creation path: NEW_SURVEY sets true (frontend/src/scenes/surveys/constants.tsx) but the Django model default is False (products/surveys/backend/models.py). Read it, don’t infer it from the creation date.

“All questions answered” does not mean what it says

isSurveyCompleted is getNextSurveyStep(...) === End — the respondent reached the end of their own branching path. Nothing checks that every question holds a value, and two things then blank out cells on a complete response: questions skipped by branching have no key at all (onNextButtonClick prunes to visitedIndices before capture, so absent rather than "" or null), and optional: true questions clicked past are stored as null (submitDisabled is isNull(rating) && !question.optional).

So on a branching survey with optional tail questions, “only the first question was answered” is the expected shape of a complete response. Blank ≠ unanswered. Check branching and optional on every question before believing a bug.

Branching is in the API — read it, don’t ask for it

It’s questions[].branching, returned verbatim by both the management API and the SDK payload, in four shapes: next_question, end (the confirmation message), specific_question + index, and response_based + responseValues. Two traps: for rating questions responseValues is keyed by bucket (negative/neutral/positive, or NPS detractors/passives/promoters), not by value; and a bucket missing from responseValues silently falls through to currentQuestionIndex + 1. So {negative: 3, neutral: 3} on a 5-scale is a complete, working config that reads like an omission. Bucket boundaries per scale are in references/reading-responses.md (opens in a new tab) — don’t guess them, scale 2 is inverted.

The event vocabulary

survey shown (fired inside showSurvey() at the moment the popup becomes visible, i.e. after surveyPopupDelaySeconds, so shown→sent latency is real time-on-popup), survey sent, survey dismissed (explicit close), survey abandoned (handlePageUnload → sendSurveyAbandonedEvent). $survey_partially_completed appears on dismissed and abandoned only, never on sent. All of them carry sessionRecordingUrl, so a disputed submission can often be watched instead of theorized about. A populated URL only proves a session id was captured, not that a recording exists: replay is off by default, and the default cloud retention of 30 days is well short of these queries’ 180-day window. Open the link before offering it as evidence.

Read answers with getSurveyResponse, never a guessed key

getSurveyResponse(<index>, '<questionId>') is the HogQL helper the product itself reads answers with (posthog/hogql/functions/survey.py, used across products/surveys/backend/responses/). Prefer it over hand-writing a property key: response keys come in three formats, and the helper coalesces the current UUID-keyed one ($survey_response_<questionId>) with the legacy index-keyed one, so it returns what the customer’s own results table shows. Pass true as a third argument for multiple-choice questions. A hand-written properties.$survey_response_<uuid> misses legacy-keyed answers, and inside a countIf that silently converts a real answer into “unanswered” and inflates the incomplete ratio.

Whichever you use, the index and id must come from questions[] in the survey JSON. Guessing which UUID belongs to which question is the fastest route to a wrong conclusion: it reads one question’s answers under another’s label, and the resulting ratio looks like a serious bug. If you have no survey JSON, unroll $survey_questions, which carries the question text alongside the answer. Formats and templates: references/reading-responses.md (opens in a new tab), references/diagnostic-queries.md (opens in a new tab).

Sanity check the mapping before trusting the numbers: if a column you labeled as a single-choice question contains free text, the mapping is wrong.

Debugging workflow

1. Parse the ticket. Extract: org/project ID, instance (US vs EU — URLs differ), survey ID(s), the lib (platform), the symptom in precise terms, and what the customer already tried. If the ticket is aged or has prior support replies, the config may have been edited mid-thread — treat earlier claims as stale and re-pull current state.

2. Disambiguate “none” vs “fewer.” Customers say “no responses” when they mean “fewer.” Pull the survey shown vs survey sent counts before/after the suspected change (see references/diagnostic-queries.md (opens in a new tab)). If the response rate (sent/shown) is stable, the problem is upstream eligibility (fewer people shown), not rendering or submission. This single check redirects most investigations correctly.

2b. For a data-quality ticket, disambiguate “incomplete” vs “complete but sparse.” Before anything else, check whether the blank cells are questions the respondent was never asked. Branching plus optional questions makes a complete response look abandoned, and this is the single most common reason a survey gets reported as broken when it isn’t. See How a response actually gets stored.

3. Platform parity check. Confirm the lib and consult the parity table. Eliminate features the platform doesn’t support before investigating them.

4. Pull the survey definition. GET /api/projects/<id>/surveys/<sid>/. Inspect conditions (events, url, seenSurveyWaitPeriodInDays, repeatedActivation), appearance.surveyPopupDelaySeconds, appearance.position, schedule, linked_flag, targeting_flag, internal_targeting_flag.filters, responses_limit, iteration_*, enable_partial_responses — and then go inside each question: id, type, scale, optional, skipSubmitButton, branching. The per-question fields answer most “wrong data” tickets on their own and are easy to miss, since a shallow look at top-level keys doesn’t surface them.

5. Pull the targeting-flag activity log for any “stopped showing” ticket. Cohort swaps and rollout changes are invisible in the current config but show up here: GET /api/projects/<id>/activity_log/?scope=FeatureFlag&item_id=<flag_id>&limit=20. Also ?scope=Survey&item_id=<sid> to see whether the survey itself was edited.

6. Confirm with events. Use $feature_flag_called to see what the gating flag actually returned for affected users, and whether $groups is set (see group-aggregation cause below). Use survey shown to see real reach vs the stats UI.

7. Diagnose against the known-cause catalog, confirm with one targeted query, then write the reply.

Before you call it an SDK bug

These tickets attract a specific failure mode: the symptom looks impossible, so the SDK gets blamed and the customer is told to pause the survey, file a bug, and rebuild their questions. That advice is expensive and hard to walk back — restructuring destroys the comparability of every response already collected, and reordering questions risks the UUID problem below. Clear all four gates before writing “this is a bug”:

1. Read the survey JSON, per-question branching, optional, skipSubmitButton and scale included. Most “impossible” behavior is configured behavior.
2. Reproduce it in Preview.
3. Read the SDK source for the handler you’re accusing. Grep the symbol; don’t reason from what a setting is named. enable_partial_responses and “all questions answered” both mean something narrower than they sound.
4. Re-check your own query before trusting a shocking ratio. “89% of responses are incomplete” is more often a bad response-key mapping than a real defect.

If it survives all four it’s a bug: name the file and handler, and file it upstream.

Known-cause catalog

Ordered roughly by how often they’re the answer.

“Survey shows to fewer users than expected”

• surveyPopupDelaySeconds + URL re-check (web only). After the event fires and eligibility passes, the SDK waits N seconds, then re-checks doesSurveyUrlMatch against the current URL before rendering (handlePopoverSurvey). If the user navigated during the delay, the survey is silently dropped — no survey shown. Common on navigation-heavy apps with a non-trivial delay. Fix: lower the delay to 0–2s.
• seenSurveyWaitPeriodInDays + the customer’s other surveys. Any survey shown to a user updates lastSeenSurveyDate; this survey is then blocked for the wait window. Completion status is irrelevant. Verify by checking whether the unshown cohort saw another survey recently — and confirm against a control group (do the shown users differ?). Fix: lower the wait period, or pause competing surveys.
• Cohort composition changed. If the survey targets a cohort and someone edited the source dynamic cohort (e.g. added a behavioral filter), every static snapshot taken afterward inherits the narrower definition. Reach drops without any survey-side change. Find it in the flag activity log (cohort swap) and confirm cohort sizes via static_cohort_people.

“Event-based survey never fires”

• Timing race at session start. Event captured before /api/surveys returns and the capture hook registers. Signature: event fires very early in session. Unavoidable client-side; mitigate by triggering on a slightly later event.
• Group-aggregated linked_flag with no group context. If linked_flag (or targeting flag) has aggregation_group_type_index set, it evaluates against a group, not the person. Without posthog.group(<index>, <key>) set before the event fires, the flag returns false and the survey never shows. Signature: $feature_flag_called returns false with empty $groups, and the $feature/<key> property is missing from the trigger events. Fix: set group context in the SDK, or switch the survey to a person-level flag.
• Customer wired the survey to the wrong flag. They create a flag with email/property targeting but the survey’s linked_flag/targeting_flag points at a different flag. Always confirm the actual linked_flag.key / targeting_flag.key from the API — don’t trust the customer’s description.
• Behavioral cohort in a realtime flag. A cohort with performed_event/behavioral filters can't be evaluated in realtime flag bytecode ("Unsupported behavioral filter for realtime bytecode", posthog/api/cohort.py). The cohort shows a bytecode_error. Surveys/flags can’t use it directly — the customer must make a static copy of the cohort and target that.

“Responses are incomplete — only the first question has a value”

Usually not a bug — see How a response actually gets stored. Work these four first; only then is the completion condition genuinely not being honored.

• Branching skipped the blank questions. Map it from questions[].branching yourself. Customers describe their intent, which may not be what’s saved, and a rating bucket absent from responseValues falls through to the next index.
• The blank questions are optional: true and the respondent clicked past them. A rating plus two skipped optional open-text questions is a complete submission with one value in it.
• Partial responses is on and you’re reading intermediate events in raw SQL that the UI collapses by $survey_submission_id.
• The response keys were guessed. Re-run keyed off questions[].id.

“The respondent says they didn’t mean to submit”

• skipSubmitButton, shown in the editor as “Automatically submit on selection”. When true no submit button renders at all (BottomSection.tsx) and one click both records the answer and advances — setRating(response) then handleSubmit(response) → onNextButtonClick in the same handler invocation, no debounce, no confirm. Rating and single-choice-without-open-choice only (canQuestionSkipSubmitButton); web-only. It is on by default in every survey template (constants.tsx) and in quick-create, so customers rarely know they enabled it.
• It compounds with popup placement. appearance.position: middle_center, a large maxWidth, and a non-zero surveyPopupDelaySeconds mean the popup materializes seconds after the trigger under a cursor that’s still moving, and the next click lands on an answer. Consecutive auto-submitting questions let someone click through most of a survey without reading it.
• Diagnose with shown→sent latency, then confirm from the replay (sessionRecordingUrl on survey sent). A cluster of sub-10-second completions on a multi-question survey is the signature; genuine respondents take longer and leave text behind. Fix is to uncheck the setting — not to move the question, since the behavior follows the question wherever it sits.

“Responses show as zero in the UI but raw events exist”

• Max AI corrupted the survey definition. Max’s edit_survey tool (products/surveys/backend/max_tools.py) has two failure modes: (a) on reorder/edit it rebuilds each question from QUESTION_TYPE_MAP (nps→scale 10, csat→scale 5, etc.), so picking the wrong semantic type silently changes a question’s scale; (b) the id field expects 1-indexed labels ("1","2") — passing a real UUID falls through and a fresh UUID is generated, so responses keyed by $survey_response_<old_uuid> no longer join to the question. Raw events are intact; only the definition is wrong. Fix: PATCH the questions array back to the original UUIDs (recoverable from the response events) and restore the question type. Tell the customer to edit question text via the UI, and avoid asking Max to reorder questions on a survey with historical responses until the tool guards UUIDs.

“Cohort count shows 0 but the cohort is populated”

• Cosmetic UI bug, does not affect targeting. Confirm the real count via static_cohort_people. NOTE: this is not a simple one-line bug — the normal insert_cohort_from_query path does recompute count via count_cohort_members; the count=0 display only appears on certain failure paths. Do not promise a quick fix without reproducing the specific path.

General caution

• Aged tickets are dirty. Config may have been edited by the customer or a prior agent during troubleshooting. Pull activity logs; frame secondary findings as “while you’re in there, double-check X” rather than “we found X is broken.”
• The stats UI can undercount vs raw survey shown events. If the numbers don’t reconcile, trust the raw events and flag the discrepancy as a separate follow-up.
• Check both $email and email on the person. person.properties.$email is what the SDKs set; plenty of customers also set a bare email. Querying one and concluding “this user was never identified” is a common miss — use coalesce(person.properties.$email, person.properties.email).

Diagnostic queries

Read-only HogQL templates for the disambiguators and confirmations above live in references/diagnostic-queries.md (opens in a new tab). Run them via the PostHog MCP execute-sql against the customer’s project.

Access for debugging

Only investigate a project tied to a genuine support request from that customer — the IDs should come from a real ticket, not from someone asking you to look up an org/survey they can’t point to a request for. Staff access is broad; don’t freelance across projects.

Prefer read-only paths in this order:

1. PostHog MCP tools (survey, feature-flag, cohorts, execute-sql, activity-log, persons) against the customer’s project. This is read-only by default and the safest way to inspect config and run queries — no impersonation, no write risk. Use this first.
2. Survey/flag API endpoints read via the browser while impersonating (staff). Good for the full JSON the MCP may not surface verbatim (e.g. raw internal_targeting_flag.filters).
3. Django admin only when 1 and 2 can’t answer it. It’s powerful and write-capable, so treat it as read-only by discipline: look, don’t change. Never edit a customer’s survey/flag/cohort from admin without explicit customer consent.

When you need a value the MCP can’t infer (project ID, instance, which survey), ask the operator to paste the survey API JSON — it skips several round-trips.

Writing the customer reply

Voice derived from the PostHog handbook support values (reassuringly human, humble, ship fixes, clear with no jargon). Rules:

• Lead with the cause, then the fix. One line on what’s happening, then what to do.
• Bold the issue and bold each action. Make the problem and the next step scannable.
• Use the labels the customer sees in the UI, never internal field names. Grep frontend/src/scenes/surveys/ for the real string. E.g. surveyPopupDelaySeconds → “Delay survey popup by at least N seconds once the display conditions are met”; the wait period is “Survey wait period” / “Don’t show this survey if another one was shown to the user in the last N days”; skipSubmitButton → “Automatically submit on selection”; enable_partial_responses → the “Response collection” radio (“Any question…” / “Complete survey…”); optional → the question’s “Optional” toggle. The customer’s own event names stay verbatim.
• When correcting an earlier wrong answer, say so plainly and early, then give the real cause. Don’t bury it or let the customer keep acting on advice we’ve retracted. If they’ve already been told to pause or rebuild something, tell them explicitly that they can stop.
• Link every PostHog entity by ID. Cohorts: https://<us|eu>.posthog.com/project/<id>/cohorts/<cohort_id>. Flags: .../feature_flags/<flag_id>. Surveys: .../surveys/<survey_id>. Match the customer’s instance (US vs EU).
• Predict the expected outcome so the customer can verify the fix worked (“you should see ~X going forward”).
• Gauge the customer’s technical level. If they can run SQL and hit the API, offer the patch path. If not, offer to apply the fix ourselves (and confirm any destructive detail first).
• Do NOT offer to “hop on a call” or book a meeting. PostHog support is async-first. Close with “We’re always here if you need a follow-up.”
• Never leak internals — no MCP tool names, code paths, line numbers, Django admin, staff impersonation, or other customers. Keep it to product concepts a customer recognizes.
• Run the final draft through a humanizer skill before sending (if you have one, e.g. humanizer). Strip em dashes, setup phrases (“Here’s the thing:”, “Three things to know”), rule-of-three padding, and tidy parallel list structure. The reply should read like a person typed it.

Reply skeleton: