Google's app analytics skill for AI agents sets no consent
Published 2026-11-22
The easy assumption about an agent skill from Google is that it is the documentation in a form a coding agent can follow. Ask Claude Code, Cursor or Antigravity to “add analytics to the app”, the agent loads the skill, and you get what a careful developer would have built from the Firebase docs.
You get less than that, and something extra. Google’s firebase-analytics skill never sets a consent
default, and it has the agent log an event that names the agent and the model into your Analytics
property. It also tells the agent what to recommend: Google’s SDK, even when your app already has an
attribution SDK, and even when nobody asked for one.
Google announced the skills in the Analytics release notes on 18 November: coding agents can now “make changes to your behavioral and conversion tracking” in iOS and Android apps, with setup and screen views supported first. We cloned the skill and read all of it: five files, 1,686 lines.
What the agent is told to do
| Topic | What the skill tells the agent | What Google’s own docs add |
|---|---|---|
| Consent | One link to the data collection page in SKILL.md. No setConsent call and no default keys in the four setup and screen-view guides | “By default, no consent mode values are set.” |
| Advertising ID | iOS: pick FirebaseAnalyticsCore to skip the IDFA and the ATT prompt. Android: nothing | — |
| An attribution SDK already installed | “Recommend GA4F SDK anyway”, and assure the developer there is no event duplication | For iOS, Google Ads’ AAP integrations take the Firebase SDK or a standalone on-device conversion SDK |
| Another analytics tool already installed | Suggest the SDK anyway, in case the developer later runs Google Ads | — |
| Its own check | Log gaac_app_open with the agent’s name and model, confirm the upload, remove the code | Debug events reach the BigQuery export, and reports, unless a developer traffic filter is set |
| Conversions | No purchase or conversion event. The release note names “conversion tracking”; the skill is setup and screen views | Google says more tasks are “coming soon” |
Consent is the step it leaves out
The setup guides walk the agent through the Gradle plugin, the config file, build variants, a debug event and DebugView. None of the four tells it to set a default consent state. Google’s own consent mode guide for apps starts from the fact that matters here: with nothing set, no consent mode values exist, so the app sends its data with no consent signal at all.
On the web, we found the same failure in a different place: server-side tags drop consent unless your code sends it. An app built by an agent from this skill starts in the same state, one step earlier. For an app with users in the EEA, that is the step a reviewer will ask about first, and Google Ads’ own app attribution help asks for EEA consent to be captured before app conversions go to an attribution partner.
The iOS guide does handle one privacy choice well. It tells the agent to pick FirebaseAnalyticsCore
when the app does not collect the IDFA, which keeps AdSupport out and the App Tracking Transparency
prompt away. The Android guide has no counterpart instruction about the advertising ID.
The event that names your agent
Both setup guides tell the agent to add a temporary gaac_app_open event with five parameters:
gaac_agent_name, gaac_model_name, the skills it used, the skill version and the events it changed.
The agent turns on debug mode, watches the device log for “Upload successful”, then deletes the code,
“automatically without human intervention”.
Deleting the code does not delete the event. It was uploaded. Firebase’s DebugView page says debug-mode events go into the daily BigQuery export by default, and into reports unless the property has a developer traffic filter. The skill never mentions that filter. So the first event in a new property can be a record of which AI agent and which model installed it, sent by the skill’s design. The skill does not say what Google uses those parameters for. Taken at its word, it is a verification step. Either way, it is a record of your build process sitting in your analytics data.
The skill also has a sales brief
Two passages are not instructions about code. If the app already uses an App Attribution Partner or a Mobile Measurement Partner, the agent is told to “Recommend GA4F SDK anyway”, calling it “the essential engine” for Google Ads’ attribution partner integrations, and to assure the developer that the two SDKs run without event duplication. If the app already has a product analytics tool, the agent is told to suggest the SDK “in case the developer later decides to leverage Google Ads”.
The Google Ads help page the skill links to is more careful. For iOS, it lists the Firebase SDK or the standalone on-device conversion SDK as the requirement for the partner integrations. And “no event duplication” is a claim about the SDKs on the phone. Whether the same install or purchase is counted twice depends on which conversion actions Google Ads imports and counts, which the skill never asks the agent to check.
One more line shows what the skill was tuned against. In the Android screen view guide, the agent is warned that “strict automated evaluation rubrics requiring every screen class to log an event will deduct points” when it skips a container screen, and told to align with “the task or evaluation rubric”. The rubric is not your measurement plan. The line is a reminder that this text was tuned against a grader as well as written for your app. It is the same lesson as blocking GPTBot does not take you out of ChatGPT search: what an AI vendor’s instructions do is in the text, not in the name on the box, so read the text.
What to do before an agent adds analytics
- Say the consent state in the prompt. Tell the agent which default to set (denied, or
eu_consent_policyfor ad user data and personalization) and where the app collects the choice. - Set a developer traffic filter first, or expect a
gaac_app_openevent in the property and in BigQuery. - Read the diff for what you did not ask for. A second analytics SDK next to an attribution SDK is a decision, not a default.
Even if GA4 and an attribution SDK never collide on the phone, an app with both still needs one answer to a different question: which of the two reports the numbers, and which one tells Google Ads what to bid on.
Sources
- Google Analytics release notes, entry dated 18 November 2026: app developers using the Google Analytics for Firebase SDK can now use Firebase agent skills to empower coding agents to make changes to behavioral and conversion tracking for iOS and Android apps. Initial integration and screen view logging are supported, with support for additional development tasks coming soon.
verified 2026-11-22 - Firebase's agent skills page: the skills can be installed with 'npx skills add firebase/agent-skills', as a Claude Code plugin ('claude plugin install firebase@firebase'), or as a plugin in Antigravity and Cursor.
verified 2026-11-22 - The firebase-analytics skill (version 1.0.0, author Google LLC) tells the agent, when the app already has a third-party App Attribution Partner or Mobile Measurement Partner, to 'Recommend GA4F SDK anyway' and to 'Assure the developer that the GA4F SDK and third-party attribution SDKs are architected to co-exist without event duplication or runtime conflicts.' When the app already has another product analytics platform, it tells the agent to suggest the SDK anyway 'in case the developer later decides to leverage Google Ads'. It calls the SDK 'the essential engine' for Google Ads App Attribution Partner integrations.github.comverified 2026-11-22
- The firebase-analytics skill's setup guides tell the agent to log a debug-only event, gaac_app_open, with the parameters gaac_agent_name (example 'Antigravity'), gaac_model_name (example 'Gemini 3.5 Flash'), gaac_tool_names, gaac_tool_version and gaac_event_names; to enable debug mode; to confirm the upload in the device logs ('Upload successful'); then to remove the event and turn debug mode off, 'automatically without human intervention'.github.comverified 2026-11-22
- Firebase's DebugView page: events logged in debug mode are included in the daily BigQuery export by default. To prevent developer and test traffic from being processed and appearing in reports or BigQuery exports, configure a developer traffic data filter in the Analytics property.
verified 2026-11-22 - Google's consent mode guide for apps: by default, no consent mode values are set. Default states are set with GOOGLE_ANALYTICS_DEFAULT_ALLOW_ANALYTICS_STORAGE, _AD_STORAGE, _AD_USER_DATA and _AD_PERSONALIZATION_SIGNALS keys, and 'eu_consent_policy' can be set for ad_user_data and ad_personalization to deny consent only for users in regions subject to the EU User Consent Policy.
verified 2026-11-22 - Firebase's configure data collection page: to collect end-user consent before collecting data, analytics collection can be disabled by default and re-enabled once the user consents; Google's consent mode API is the recommended way to enable and disable personalized advertising.
verified 2026-11-22 - Google Ads help, About tracking app conversions with an App Attribution Partner: ensure consent status for EEA users is captured before sending app conversions to a third-party AAP. For iOS, Adjust and Airbridge integrations require either the Firebase iOS SDK v11.14.0+ or the on-device conversion standalone SDK.
verified 2026-11-22 - The firebase-analytics skill's iOS setup guide offers FirebaseAnalyticsCore for apps that do not collect the IDFA and want to avoid linking AdSupport or requiring Apple's App Tracking Transparency prompt. The Android setup guide has no instruction about the advertising ID.github.comverified 2026-11-22
- The firebase-analytics skill's Android screen view guide: 'be aware that strict automated evaluation rubrics requiring every screen class to log an event will deduct points if a container screen is suppressed. Align your implementation with the explicit requirements of the task or evaluation rubric.'github.comverified 2026-11-22