Google AI Search Visibility & Brand Citation Monitor

GEO / AEO Visibility Monitor

An Apify Actor that turns search and AI-search results into recurring brand visibility and competitive-intelligence data.

The Actor orchestrates a configurable Apify source Actor and normalizes common SERP/AI-answer response shapes. Output contract version 2 separates billable results from diagnostics:

• The default Dataset contains one visibility_snapshot per successful query, market, and engine check. Completed snapshots are delivered individually during the run.
• Run report and historical changes opens RUN_REPORT in the run's key-value store: a report_summary with visibility, mention/citation rates, competitors, historical changes, source budgets, and parent event-charge accounting.
• Source errors opens SOURCE_ERRORS in the same store: a diagnostic array for partial or complete source failures.

The default source is apify/google-search-scraper. The source Actor is intentionally configurable so the monitor can later add other engines without changing its output contract.

Choose a monitoring preset

New Console inputs prefill Google web + AI Mode - two checks, extra charges, so a slow AI add-on cannot discard the separately collected web evidence. The schema uses prefill: google_web_and_ai and default: custom: the interface suggests web and AI collection while omitted fields in existing tasks and API requests preserve their current settings. Every preset works with your chosen brand and competitors; Apify and Notion in the examples are only sample brands.

monitoringPreset	Google collection settings	Source charges
google_web_and_ai	First a web-only source run, then an independent AI Mode source run, per query/market/source. Engine names end in :web_search and :ai_mode.	Two source runs; up to two billable snapshots per query, plus source search/AI fees.
google_ai_mode	Enable Google AI Mode; full AI Overview capture off.	Standard search plus the source's AI Mode add-on.
google_ai_overview	Full AI Overview capture on; AI Mode off.	Standard search plus the source's captured AI Overview add-on. Google may not show an Overview.
google_ai_full	Enable both Google AI Mode and full AI Overview capture.	Standard search plus both applicable add-ons.
web_search	Disable both Google AI switches.	Standard source search charges; this preset is not an AI-answer visibility test.
custom	Preserve sourceActorInput and engine inputOverrides.	Depends on the source and enabled options.

Google presets require apify/google-search-scraper and control the two Google switches even when advanced options conflict. The two-channel preset creates independent checks and history baselines for web and AI. If AI times out but web succeeds, the report is partial, its source diagnostic identifies SOURCE_TIMEOUT, and the web snapshot remains accessible. Missing AI measurements remain unknown. If both checks fail, the run still fails with an empty dataset. The two channels both count toward maxSourceRuns and share the source budget and parent fee cap. Allow at least two source runs to collect both. Other AI add-ons require custom; their response formats still require compatible normalization. For compatibility, saved inputs and API calls without monitoringPreset retain their existing collection settings (custom). Choose google_ai_mode for a single AI Mode check or google_web_and_ai for independent web and AI checks. Collection changes establish a new compatible-history baseline.

For a one-query AI check, use examples/ai-live-input.json. For a regular-search check, use examples/live-input.json. Both launch paid sources. See the source's current pricing; source budgets do not include this monitor's fee. Presets request AI data but cannot guarantee that an answer is returned. A successful snapshot requires usable search or AI evidence; a valid search-only snapshot is still a billable visibility check and has unknown AI-answer measurements.

For the recommended independent web + AI workflow, use examples/web-and-ai-input.json. This example makes at most two paid source calls and produces at most two snapshot charges. Change the sample brand, domains and question to your own project before scheduling recurring runs.

For Google, queries is sent as a single string according to the source's input contract. To manage the switches manually, choose custom and set sourceActorInput, for example {"aiOverview":{"scrapeFullAiOverview":true},"aiModeSearch":{"enableAiMode":true}}.

Pricing and parent charge limit

Keep the existing apify-default-dataset-item event in Console. Only successful snapshots enter that dataset, so its event now bills successful checks only. Summaries and error records do not incur per-result charges. The automatic apify-actor-start fee remains separate, and upstream Actors can charge for failed or empty source runs. Ordinary compute/storage costs also apply according to platform pricing.

At $0.01 per snapshot, 10 successful query checks produce 10 billable rows: $0.10 plus startup and upstream charges. Their summary and error diagnostics add zero result events. No-results or all-error runs have zero snapshot events, while startup/upstream costs can still apply.

With google_web_and_ai, 10 queries in one market/source can produce 20 successful snapshots ($0.20 at $0.01 each). This intentionally performs two checks per query. If the AI run fails, that failed check adds no snapshot charge, though upstream charges can still apply.

Before starting each new source run, the monitor reads that run's pricing and charged event counts from Apify, reserves the startup fee according to memory, and checks that one additional snapshot fits within the smaller of ACTOR_MAX_TOTAL_CHARGE_USD and the run's API maximum. Local delivery receipts cover delayed event counts. If a snapshot does not fit, the run stops with parent_charge_limit and saves its report without launching more sources. Malformed or unsupported pricing fails before paid source work. No charge request is sent manually; Apify bills dataset writes automatically. This is an event-charge guard for the parent run, separate from the source-run budget, and is not an account-wide or entire-invoice spending cap.

report_summary.billing includes the run-specific unit price, maximum parent charge, accounted event charges, remaining snapshot capacity, confirmed delivery count and pending delivery count. Pending writes reserve their full event price until reconciled; these figures are conservative accounting, not a final invoice. Platform compute/usage charges and downstream spending are not estimated by the parent event guard. Actor memory is limited to 1 GB.

When updating your published Actor, keep the event name and existing unit price. You can rename its Title to Successful visibility check and its Description to One successful query, market and source-Actor visibility snapshot. Run summaries and source errors are not charged as results. No new custom event is needed. Start a new run after rebuilding; version 1 run checkpoints cannot be resurrected into this output contract.

Run locally with a fixture

The fixture path keeps local development deterministic and avoids paid Apify calls:

$env:GEO_INPUT_FILE = "$PWD/examples/input.json"
$env:GEO_FIXTURE_FILE = "$PWD/examples/source-fixture.json"
$env:GEO_OUTPUT_DIR = "$PWD/test-output"
npm start

The run writes test-output/dataset.jsonl, keyed checkpoints under test-output/records/, and persistent history under test-output/named-stores/. Keep that directory for a second local run to verify history; delete it after verification. Fixtures report zero source charges and make no paid calls.

Run on Apify

Build and run the Actor with an Apify token. The runtime reads the platform INPUT record, invokes the configured source Actor with the same token, writes dataset items to the default dataset, and stores checkpoints in the run's default key-value store. Cross-run history uses a separate named store in the running customer's account. Normal limited Actor permissions support creating storage and using storage created by earlier runs; full permissions are not required. Storage authorization failures are surfaced before launching paid sources.

The runtime supports the current ACTOR_DEFAULT_KEY_VALUE_STORE_ID, ACTOR_INPUT_KEY, and ACTOR_DEFAULT_DATASET_ID variables, as well as legacy APIFY_* names. Input defaults to the INPUT record when the platform does not set a record key.

For a small live smoke test, paste examples/live-input.json into the Actor's Input → JSON editor, save it, and start a new run of the latest successful build. This checks one Google query for Notion without paid AI add-ons. It still incurs source Actor and platform usage charges; the $0.50 source cap does not include this monitor's fee. A successful source call produces one visibility_snapshot in the Dataset and a separate RUN_REPORT; verify snapshotCount: 1, errorCount: 0, and sourceRecordCount > 0. A regular Google result alone does not verify AI answer visibility.

Remove GEO_USE_EXAMPLE when testing Console input. Setting it to 1 explicitly allows the bundled example as a fallback when no input is found, which can hide an input-delivery problem. If a run fails before producing output, inspect that run's Log and Input tabs. The missing-input error distinguishes missing runtime credentials, missing store IDs, and a missing input record.

Persistent history

Use a stable monitorId such as notion-us for recurring runs. historyEnabled defaults to true; historyRetentionDays defaults to 90. The Actor creates/reuses a named key-value store, geo-history-<actor hash>, scoped to the running Apify account. The report exposes its name and ID in history. It does not use a global developer-owned store.

Each completed run writes its own record. There is no shared mutable "latest" pointer, so concurrent runs cannot overwrite each other's history. Each query uses the latest compatible observation completed before the current run started. Compatibility includes the brand and aliases/domains, competitors, query, country, language, location, engine/source Actor, actual source input (including AI settings), and analysis version. Adding queries preserves existing query baselines. Changing measurement settings starts a new baseline.

History retains successful observations from partial runs. Failed, empty, unresolved, and budget-skipped tasks never replace valid baselines. History omits rawSource, including when raw source output is enabled. The first compatible run has scoreDelta: null and baselineRunId: null. Later changes entries include baseline run/time, score delta, brand/AI mention gained/lost, AI owned-citation gained/lost, competitor changes, and citation-domain changes. A missing AI answer remains unknown rather than becoming a false observation.

Reports now include changeDetectionVersion: 2. Each changes entry also compares AI response counts, total citation counts, AI citation counts, and attached-source counts using previous..., current..., and ...Delta fields (for example previousAiResponseCount, currentAiResponseCount, aiResponseCountDelta). Exact URL additions/removals appear in citationUrlsGained/Lost, aiCitationUrlsGained/Lost, and aiAttachedSourceUrlsGained/Lost. URL replacements within the same domain count as changes even when counts and scores stay constant. List order and duplicate URLs do not count as URL changes.

aiOwnedAttachedSource on a snapshot indicates whether any AI channel includes a brand-owned attached link. It remains distinct from aiOwnedCitation: attached links do not become formal citations or contribute to the visibility score. History reports previousAiOwnedAttachedSource, currentAiOwnedAttachedSource, and aiOwnedAttachedSourceChange (gained, lost, unchanged, or null when unavailable).

aiChannelChanges compares each AI channel separately: channel and answer availability, counts, formal/attached URLs, brand mention, competitor mentions, and owned-source transitions. Changes cannot cancel out merely because the combined URL sets are unchanged. changedFields explains which tracked measurements changed; changed becomes true when any aggregate or channel comparison changes. Missing answers can change response availability while their brand-mention transition remains unknown. First runs, incompatible versions, and fields absent from older history do not invent losses or gains. Existing analysis-version-2 baselines remain usable, and the owned-attached flag is derived from their saved evidence when needed. Raw source output is not required.

changed: false means these tracked measurements are unchanged. Full answer prose, ranking within AI recommendations, and URL order are not compared; this flag does not assert identical answer text.

historyKey remains a supported namespace, combined with monitorId. Existing run-local history arrays are not imported because their brand and collection configuration cannot be verified. The first run after this upgrade establishes a fresh baseline.

Expired records are deleted only for the current monitor namespace when it next runs. Named stores otherwise persist indefinitely; an inactive project's history is not cleaned on a background timer. Deleting that named store removes its history. See Apify storage retention and limited Actor permissions.

Source budget and recovery

Input	Default	Meaning
maxSourceBudgetUsd	2	Combined source-run budget for this monitor run.
maxTotalChargeUsd	Remaining aggregate budget	Optional maximum for each individual source run.
maxSourceRuns	20	Maximum launch attempts across all queries, markets and engines.
sourceTimeoutSeconds	300	Timeout sent to the source platform run, shortened near the parent's deadline.

Sources run sequentially. Before each launch, the Actor checkpoints a reservation equal to min(per-run cap, remaining source budget). It sends this amount as the source run's API query parameter maxTotalChargeUsd, together with the platform timeout. It never raises a cap to satisfy a source's minimum charge. Zero budget, zero per-run cap or zero source-run limit prevents new launches.

Terminal source runs are accounted using the authenticated run's usageTotalUsd, including failed and aborted runs. Missing charges consume the full reserved allowance instead of being treated as free. No provisional running cost or separate event charge is added to usageTotalUsd. USD arithmetic uses integer micro-dollars; budget allowances round down and charges round up. For example, with a $0.30 aggregate budget, a $0.20 per-run cap and $0.10 actual charges, the successive caps are $0.20, $0.20 and $0.10.

The aggregate budget covers the source runs launched by this monitor. This monitor's own compute, storage and traffic charges are additional. Platform billing granularity and late settlement also prevent promising an absolute cap on the entire Apify invoice. Separate concurrent monitor runs each have their own source budget; this is not an account-wide monthly spending limit. See Apify source run cost and timeout parameters.

report_summary.budget exposes known charges, reserved unknown charges, accounted charges, remaining budget, and a per-source ledger with run IDs. The report also includes planned/completed/failed/unresolved/skipped counts and stopReason (budget_exhausted, max_source_runs, parent_deadline, parent_charge_limit, ambiguous_launch, or source_reconciliation_required). A budget stop preserves completed results and returns status: "partial". All-source failures produce status: "failed" and a nonzero process exit after saving diagnostic output. RUN_REPORT is also updated during collection with status: "running", so an interrupted run retains its latest progress.

Checkpoints are saved before launch, after obtaining the child run ID, after settlement, and after analysis. Resurrection of the same Apify run reuses its checkpoint and known child IDs instead of relaunching them. A brand-new run has a fresh budget and performs fresh measurements. Do not edit input while resurrecting an existing run. Completed output is not emitted again when its delivery checkpoint is confirmed.

If a launch POST loses its response, a child may already exist. The Actor reserves its full allowance and stops with ambiguous_launch; it does not blindly repeat that POST. Inspect the source Actor's run list before starting another monitor run. If a known running child cannot be reconciled or aborted, the Actor stops other launches; resurrection can inspect that same child again. Local timeouts and polling errors request an abort. The child also has a platform timeout, since forced parent termination cannot guarantee cleanup code executes.

Dataset delivery and checkpoint writes are separate API operations. Each snapshot write is checkpointed as pending before delivery and confirmed afterward. On resurrection, an interrupted write is checked against (runId, comparisonKey) in the existing dataset. A matching row is confirmed without replaying its paid write. If delivery cannot be confirmed, the run stops with a dataset_delivery_unconfirmed report and does not repeat the write or launch more sources; inspect the dataset and diagnostic before starting another run. Reports are overwritten in the run store, never appended as billable rows. Concurrent instances of the same run ID are not supported; separate scheduled runs have separate IDs and independent history records.

Output contract

The default dataset defines two Console views in .actor/dataset_schema.json: Visibility overview and AI evidence and competitors. Both contain successful snapshots only. Select Visibility overview in the Dataset schema section when publishing an example task. Open Run report and historical changes or Source errors in the run's Output tab for the separate diagnostics. API clients that previously read report_summary or source_error from dataset items must use the output URLs or the RUN_REPORT and SOURCE_ERRORS records in the run's default key-value store. Views control presentation without changing stored records or the number of dataset items.

Every successful snapshot contains the query, market, engine, overall and AI-response brand mention flags, owned-citation flag, best detected position, visibility score, evidence confidence, competitor mentions, citations, and a suggested next action. Raw source records are omitted by default; set includeRawSource to true when audit-level reproducibility is needed.

aiResponses separates AI Overview (aiOverview) and AI Mode (aiModeResult) with their own brand mentions, competitors, and source lists. aiCitations contains AI source citations; aiAttachedSources contains links that the source labels section: "attached". Attached links do not count toward aiOwnedCitation or citation totals. The combined citations list includes search result URLs and AI citations, while excluding query/navigation URLs and snapshot metadata. bestPosition measures a matching search result's position, not the order of brand recommendations in AI prose.

Brand mentions are detected in result text and AI answer text, excluding the user's query, related queries, source titles, and runtime metadata. A missing AI answer yields brandMentionedInAiResponse: null; a captured answer without the brand yields false. aiOwnedCitationRate reports the rate of snapshots with brand-owned AI citations separately from brand mentions. These are observed references and text matches, not an assessment of endorsement or citation quality. The visibility score is a heuristic, not a Google metric.

Analysis version 2 corrects AI Mode detection and citation classification. Snapshots include analysisVersion: 2; historical score deltas require matching analysis versions. The test fixture in tests/fixtures/google-ai-mode.json is a reduced, sanitized representation of a successful source response and contains no signed snapshot links.

Development checks

The input schema includes explicit prefill values for the required brand and query fields, as required by Apify's daily Store checks. Restoring the example selects Apify as a sample brand and one query, with independent web and AI Mode checks, at most two source runs, a 90-second timeout for each source, and a $0.50 shared source budget. The published workflow uses genuine upstream results for every run. Source and parent fees still apply to this live example; it does not guarantee upstream availability. These UI/test prefills do not reduce limits on existing saved inputs or API calls.

The official check OKU91PmoaAfgeLqaH on October 21, 2026 failed because its Google AI Mode child run timed out after 277 seconds with Nothing extracted yet. The parent had no successful snapshots and exited with code 1. Its billing and input paths completed normally. The revised example limits each source's wait and keeps web/AI evidence independent. SOURCE_ERRORS includes a diagnostic code, upstream run ID/status, and channel; the parent publishes a terminal status message with these details. Source outages remain visible rather than being converted into fabricated AI measurements. After a new build passes the platform's daily check, Apify can clear its maintenance flag automatically; local tests do not establish cloud health.

npm test
npm run lint

The implementation intentionally uses Node's built-in fetch and test runner, so a local install does not need a dependency cache outside this workspace.