Google AI Mode

Runs prompt in Google AI Mode (the AI Mode tab of Google Search, udm=50) as a logged-out visitor and returns the answer with the sources it cited. For the AI Overview Google shows above ordinary results, use /v1/search: its response carries ai_overview when the page has one, at the price of the search.

Billing: 2 credits per answer, 1 from the cache with max_age. AI answers cost the same in POST /v1/batches: the batch lane does not halve them. A failed, refused or empty answer is refunded (503, nothing billed).

Latency: answers typically take 5–45 s; requests wait up to 90 s. Set a client timeout of at least 120 s, or queue prompts with batches or a monitor instead of holding requests open.

Variance: assistants answer the same prompt differently from run to run and by country. To measure visibility, ask each prompt several times and compare shares, not single answers. See AI visibility tracking.

Request body

Example request

Authenticate with Authorization: Bearer $SERPKITE_API_KEY (GET requests may pass ?api_key= instead). The official SDKs read SERPKITE_API_KEY for you. See Authentication.

curl https://api.serpkite.com/v1/ai_mode \
  -H "Authorization: Bearer $SERPKITE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"prompt":"how to remove a red wine stain from a wool carpet"}'

Response fields

Every billed response also carries credit and latency headers (X-Credits-Used, X-Credits-Remaining, X-Request-Id…).

Example response

{
  "request": {
    "endpoint": "ai_mode",
    "prompt": "how to remove a red wine stain from a wool carpet",
    "country": "us",
    "language": "en"
  },
  "text": "Blot (don't rub) as much wine as you can, then dab the stain with cold water and a little wool-safe detergent…",
  "markdown": "**Blot, don't rub**, as much wine as you can, then dab the stain with cold water and a little wool-safe detergent [1](https://www.example.com/wool-carpet-stains)…",
  "sources": [
    {
      "position": 1,
      "title": "Cleaning wool carpets: stains",
      "link": "https://www.example.com/wool-carpet-stains",
      "domain": "example.com",
      "cited": true
    },
    {
      "position": 2,
      "title": "How to get red wine out of carpet",
      "link": "https://www.example.org/red-wine-carpet",
      "domain": "example.org",
      "cited": false
    }
  ],
  "search_queries": [],
  "products": [],
  "meta": {
    "request_id": "req_01J8ZK4M6Q2V7",
    "credits_used": 3,
    "cached": false,
    "engine": "ai_mode",
    "latency_ms": 9800,
    "parse_quality": "ok"
  }
}

Errors

Errors use one shape: {"error":{"code","message","request_id"}}. Errors are never billed. Full list in Errors.

Notes

• AI Mode isn't offered in every country; a prompt for a country where Google doesn't show it fails with 503 and is free.
• We ask like a logged-out visitor in a standard browser and never sign in. When the assistant refuses the visit or asks the visitor to sign in, the call fails with 503 upstream_blocked and is free. See AI assistants.
• engine is not accepted (the path names the assistant), and neither are search options such as q, num or location: they return 400 invalid_request.
• Monitors take endpoint: "ai_mode" with prompt and track, and report every run's answer with new_sources and mentions. See Monitors.
• The remote MCP server has a ai_mode tool. See MCP server.
• The prompt is handled like a search query: never logged, and cached for at most an hour under a hashed key. See Privacy and data retention.

Related

• AI assistants overview
• AI visibility tracking guide
• ChatGPT
• Gemini
• Copilot Search
• Brave AI
• Google AI Mode API overview and pricing
• OpenAPI spec