Loading the catalogue…
Loading the catalogue…
Compliance posture
Sovereignty ladder · how the level was reached
Stav's assessment · serving-side
Google serves inference through Gemini API, AI Studio, and Vertex AI under a US-incorporated parent (Google LLC), which remains reachable under the CLOUD Act even when EU customers contract with Google Ireland Limited or select EU-region Vertex AI endpoints — this is the composite's dominant drag, landing legal_exposure at 38. Security posture is the clear strength, backed by a broad and currently-sourced certification stack (ISO 27001/27017/27018/27701, SOC 2 Type II, SOC 3, BSI C5, FedRAMP) and an active Cloud VRP. Serving residency and retention are configurable rather than default: EU-pinned processing only applies once a customer deliberately moves to paid Vertex AI EU endpoints, whereas the no-training guarantee applies to all paid Gemini API and Vertex AI usage — free-tier Gemini API/AI Studio traffic trains on prompts by default with human review and processes globally. Contractual posture is solid — a published DPA, assumed EU SCCs, and subprocessor transparency — but doesn't offset the underlying jurisdictional exposure. Given the mix of a strong security/contractual layer against unresolved US legal exposure and default-off EU residency, Stav rates this elevated_risk: sovereign serving is achievable only through deliberate, correctly-configured paid Vertex AI EU deployments, not as a default posture — routed-only unless that derogation path is explicitly engineered.
EU-pinned processing is available only via explicit configuration of Vertex AI EU regional endpoints, while default Gemini API/AI Studio traffic is processed globally with no residency guarantee.
Certifications & legal documents
Endpoints served · 42
Google LLC, the US-incorporated parent and infrastructure operator, remains subject to CLOUD Act/FISA reach regardless of whether EU customers contract with Google Ireland Limited or store data in EU regions.
Paid Gemini API/Vertex AI traffic carries a contractual no-training guarantee, but free-tier Gemini API and AI Studio use trains on prompts by default with human review, and even paid grounding features retain data for 30 days.
Google holds multiple current, sourced certifications (ISO 27001/27017/27018/27701, SOC 2 Type II, SOC 3, BSI C5, FedRAMP) plus an active, well-funded bug bounty program.
A published DPA, Google's assumption of EU SCC obligations for EEA/UK/Swiss customers, and a detailed subprocessor list support contractual clarity, though the split between Google Ireland contracting and Google LLC's US exposure leaves counterparty certainty imperfect.
Risk assessment
Google Cloud Platform Agreements are entered into with Google Ireland Limited (or regional affiliates) for EMEA customers, but Google's ultimate parent and infrastructure operator, Google LLC, is US-incorporated, leaving group-wide exposure to US CLOUD Act / FISA reach independent of the contracting entity or region where data is stored. source ↗
LEGAL_EXPOSUREOn Unpaid Services (free Gemini API quota and direct Google AI Studio use), Google may use submitted prompts and responses to improve its products and machine-learning technologies, and human reviewers may read, annotate, and process API input/output — a default posture unsuitable for sensitive or regulated data. source ↗
SERVING_RETENTIONOn June 12, 2025, a flawed Service Control update without feature-flag protection triggered a multi-hour global Google Cloud outage affecting over 70 services across dozens of regions, cascading to major downstream customers (Cloudflare, Spotify, OpenAI, Shopify, Discord). source ↗
RESILIENCEThe consumer/default Gemini Developer API and AI Studio process data globally with no residency guarantee; EU-pinned processing is only achieved by explicitly using Vertex AI with a specified EU regional endpoint, meaning residency is opt-in rather than default. source ↗
DATA_RESIDENCYEven under Paid Services, Google states prompts, contextual information, and output submitted for 'Grounding with Google Search' are stored for thirty (30) days for creating results and for debugging/testing purposes. source ↗
SERVING_RETENTIONThird-party open-weight models served via Vertex AI 'Model-as-a-Service' (e.g. Llama, Qwen, DeepSeek) do not automatically inherit the same data-residency guarantees Google applies to its own Gemini models, requiring per-model verification for regulated deployments. source ↗
DATA_RESIDENCYSafeguards
Under Paid Services (billed Gemini API projects and Vertex AI), Google contractually commits not to use customer prompts, system instructions, cached content, or responses to improve its products, i.e. no training on customer data. source ↗
Vertex AI offers data-residency guarantees for generative AI (including Gemini and Imagen) with customer-selectable storage/processing regions across North America, Europe (Netherlands, France, UK, Germany, Belgium and others), and Asia. source ↗
For customers located in the EEA, UK, and Switzerland, Google has assumed all responsibilities imposed by the new EU Standard Contractual Clauses, simplifying cross-border transfer compliance without requiring a separate contracting entity. source ↗
Google Cloud publishes a detailed, regularly updated list of third-party and Google Group subprocessors used to deliver its services. source ↗
Google operates a dedicated Cloud Vulnerability Reward Program (launched Oct 2024) covering 460+ products, paying out over $3.5 million to 143 researchers in 2025 alone as part of a broader $17.1M annual VRP payout. source ↗
Following the June 2025 outage, Google published a detailed public incident report and committed to specific remediations (modularizing Service Control, enforcing feature-flag protections, improved external communications). source ↗
Privacy-policy issues
Training-on-prompts by default for free tier source ↗
Unpaid Gemini API quota and direct Google AI Studio use permit Google to use submitted prompts/outputs for product improvement and model training, with human reviewer access, unless the customer is in the EEA/Switzerland/UK where paid terms apply universally.
Retention beyond request lifecycle for grounding feature source ↗
Even on Paid Services, prompts and context used for 'Grounding with Google Search' are retained for 30 days for debugging/testing, creating a serving-boundary retention window some regulated customers may not expect.
Region defaults do not guarantee EU-only processing source ↗
Absent explicit configuration of Vertex AI regional endpoints, Gemini API/AI Studio traffic can be processed and transiently stored in any country where Google or its agents maintain facilities.