A client asked me to "pull the AI answer" for forty questions and put it in a spreadsheet. The hard part was not the request. It was that the answer, the Markdown version, and the pages Google cites are three different things, and I wanted all three in one record.
AI Mode is not AI Overview. An AI Overview is a module on a normal results page. AI Mode answers the question in a conversation-style block and attaches the pages it used. SerpBase exposes it as POST /google/ai_mode, and the response has a clean ai_mode object: text, markdown, and sources.
Here is the whole call:
import json
import os
import requests
API = "https://api.serpbase.dev/google/ai_mode"
def ai_mode(question: str, gl: str = "us", hl: str = "en") -> dict:
resp = requests.post(
API,
headers={
"X-API-Key": os.environ["SERPBASE_API_KEY"],
"Content-Type": "application/json",
},
json={"q": question, "gl": gl, "hl": hl},
timeout=60,
)
resp.raise_for_status()
data = resp.json()
# HTTP 200 is not the success signal. status == 0 is.
if data.get("status") != 0:
raise RuntimeError(
f"status={data.get('status')} request_id={data.get('request_id')}"
)
block = data.get("ai_mode") or {}
return {
"question": question,
"answer": block.get("text", ""),
"markdown": block.get("markdown", ""),
"sources": [
{
"title": s.get("page_title") or s.get("title"),
"url": s.get("link"),
"snippet": s.get("snippet"),
}
for s in block.get("sources", [])
],
"request_id": data.get("request_id"),
"credits_charged": data.get("credits_charged"),
}
if __name__ == "__main__":
result = ai_mode("how does a heat pump work")
print(json.dumps(result, ensure_ascii=False, indent=2))
The response fields I rely on:
ai_mode.text— the answer as plain text.ai_mode.markdown— the same answer with headings, lists and tables kept as Markdown.ai_mode.sources[]— each cited page withtitle,page_title,link,snippet.linkis the real destination, not agoogle.com/gotowrapper.status,request_id,credits_charged— from the envelope, next to the payload.
A trimmed response looks like this (fields only, text shortened):
{
"status": 0,
"request_id": "r-...",
"credits_charged": 1,
"product_id": "google_ai_mode",
"query": "how does a heat pump work",
"ai_mode": {
"text": "A heat pump moves heat instead of generating it...",
"markdown": "## How it works\n\nA heat pump moves heat...",
"sources": [
{
"title": "Heat pump",
"page_title": "Heat pump - Wikipedia",
"link": "https://en.wikipedia.org/wiki/Heat_pump",
"snippet": "A heat pump is a device that transfers heat energy..."
}
]
}
}
Checking whether Google cited you
The reason I wanted sources as structured data was to answer a simple question: did the AI answer use our page? Once the call returns a dict, that is a one-liner:
from urllib.parse import urlparse
def cited(result: dict, domain: str) -> list[dict]:
return [
s
for s in result["sources"]
if urlparse(s["url"]).netloc.endswith(domain)
]
hits = cited(result, "serpbase.dev")
print(len(hits))
Run it for a fixed list of questions once a day, store the JSON, and you have a citation monitor. The same loop also gives you the answer text, so you can diff what changed.
Two details that matter
format changes the response type. The default json gives the envelope above. With "format": "markdown" the endpoint returns Markdown directly (text/markdown, no envelope), which is the right shape for feeding an LLM. Keep the JSON mode when you need the citations and the request ID.
The endpoint is priced per request. At the time of writing (2026-11-16) it is 1 credit under a launch promo; the docs say it becomes 3 credits on 2026-11-26 16:00 UTC. Failed requests return a non-zero status and are not charged.
The request contract and field names are in the Google AI Mode API docs. I work on SerpBase, so that is our endpoint; the code above is the same call I use when I need the answer and its sources in one record.
Pick twenty questions your buyers actually ask, run them through the function, and check the sources list before you write anything else. The answer is the easy part. The citations tell you whether you are in the conversation at all.
Comments
Join the discussionNo comments yet. Be the first to comment.