On 11 November 2026, Min Choi and Felix Yu of Google DeepMind published the Gemini Decisions API, powered by a new model, gemini-decision-1. It is in the Gemini API and Google AI Studio. The post is on the Google blog.
This is a developer tool, not a ranking change. It does not touch crawling or indexing. It returns a closed-set decision instead of generating text.
Three types are supported. Yes/no is type noul. Multiple choice takes 2 to 255 choices and returns a probability for each. Ordered rubric scores start at 0 and return a weighted score plus a probability per level. Every answer has a confidence score from 0 to 1, so an ambiguous case can go to a person.
Advertisement
One request carries one multimodal input and up to 1,000 independent questions. The input can be text, images, audio, video or PDFs, up to 256K tokens. Options are scored in a single forward pass: 150-250 ms end-to-end, nearly flat from 1 question to 100.
Price is $0.10 per 1 million input tokens and $0 for output tokens. Shared input is billed once per request. Google AI Studio has a no-cost tier for prototyping. Google says that, on its benchmark suite, this is nearly 5x cheaper than Gemini 3.5 Flash-Lite via the Interactions API.
Google is already using it: model routing in Gemini Notebook (up to 40% lower latency), a rapid canvas in Stitch, and emoji picking in CC. The post says Google Search "uses the Decision API to detect adversarial behavior and content" as a safety layer for MCP tools in Search.
Integrations are rolling out for OpenRouter, LangChain, Vercel, LiteLLM, Pydantic AI, Pipecat, LiveKit, OpenCode and the google-genai SDKs.
WiseSEO's interpretation
The ranking systems did not change. What changed is the price of asking a closed question across a long list.
Label a keyword export by intent: buy, compare or learn. Keep the 0-1 confidence, and hand-review only the low-confidence rows. Grade a URL against a checklist you write, not one the model invents. "Answers the query in the first paragraph?" is a yes/no. Helpfulness can be a 0-3 rubric. An image input can check a product photo against its alt text.
The Search sentence is a product note, not a ranking note. A safety layer for MCP tools is a hint that agent-facing Search tooling is already running. The post does not say a page can influence that check.
The checklist itself is the one in our guide to on-page SEO best practices. How to group the queries before you label them is in NLP for SEO.
what to do
- Do not rewrite titles, links or indexation because this shipped. It is not a Search update.
- Lock the intent labels (buy, compare, learn, or your own) and store the confidence beside each row. Send low-confidence rows to a person.
- Grade priority URLs on your own checklist: yes/no for a first-paragraph answer, 0-3 for helpfulness. Read the misses before a score becomes policy.
- For image-led pages, pass the photo with the alt text and ask a closed question. Review the failures.
- Prototype on the no-cost AI Studio tier. Shared input is billed once, so put every question about one page or list in the same request.
refs
- [1] Gemini Decisions API: Fast, calibrated multimodal decisionshttps://blog.google/innovation-and-ai/technology/developers-tools/gemini-decisions-api/
Advertisement