Connect Google Gemini to WhatsApp

A normal chatbot can only answer what you scripted. Connect it to Gemini and it can handle questions you never anticipated, in natural language, drawing on the model's general knowledge or on context you feed it. You do this with one External API Request step that sends the customer's message to Gemini and brings the answer back into the chat.

This guide gives you the exact values to use. If you have not met the External API Request step before, read the field-by-field foundation guide first; here we assume you know what the fields are and focus on what Gemini specifically needs.

Step 1: Get your Gemini API key

The key is what proves the request is yours. You get it from Google AI Studio, which is the lightweight developer path (as opposed to the heavier enterprise route through Vertex AI).

Step 2: Fill in the External API Request step

Now the request itself. Here is exactly how the step should look, followed by each value.

Method

POST. You are sending the model a prompt, so this is a send, not a read.

Request URL

The generateContent endpoint for your chosen model. The shape is fixed; only the model name in the middle changes as Google releases new versions.

https://generativelanguage.googleapis.com/v1beta/models/gemini-3.5-flash:generateContent

The endpoint uses v1beta (this is correct, not a typo; most current features land there first). Replace the model name with the current stable model from Google's models documentation if it has changed.

Auth and headers

Gemini takes the key in a header. Add two header parameters:

The x-goog-api-key header is Google's current recommended method. Avoid the older key= URL parameter, which exposes the key in logs.

Body

Set Body Type to JSON and send the contents structure Gemini expects. Put the customer's message where the text is, pulling it from your flow rather than hard-coding it.

{
  "contents": [
    {
      "parts": [
        { "text": "{{customer_message}}" }
      ]
    }
  ]
}

Here {{customer_message}} is a placeholder for the value you map in from the conversation. You can also prepend instructions, for example a sentence telling Gemini to answer as your business's support assistant, before the customer's text.

Step 3: Map the reply back into the chat

Gemini answers with a JSON structure. The text you want sits inside the candidates list, in the content's parts. Set Response Type to Default Response and map that field into the next step, a WhatsApp message back to the customer.

{
  "candidates": [
    {
      "content": {
        "parts": [
          { "text": "...the model's answer is here..." }
        ]
      }
    }
  ]
}

So the path to the answer is the first candidate, then content, then the first part, then text. Map that into your reply message and the customer receives Gemini's response inside WhatsApp, in the same conversation, with no app-switching.

A worked example: an AI FAQ assistant

A customer messages your store asking something your scripted bot does not cover, say, whether a fabric is machine washable. Your flow catches the unmatched question and routes it to the Gemini step. You prepend a short instruction so the model stays on-brand, then send the customer's question.

{
  "contents": [
    {
      "parts": [
        { "text": "You are the support assistant for a clothing store. Answer briefly and politely. Customer question: {{customer_message}}" }
      ]
    }
  ]
}

Gemini replies with a short, polite answer, your flow maps it into a WhatsApp message, and the customer gets a useful response in seconds. For anything the model should not handle, an order change, a refund, you route to a human instead. The AI covers the long tail of open questions; people handle the decisions that matter.

Troubleshooting

The response the step returns almost always contains Google's own error message, which names the exact problem. Read it first.

Common questions