How to check if Gemini mentions your brand

The shortest honest answer: open Gemini, ask your category question, and search the reply for your brand name. Then check sources[] — but check whether it's populated at all before you conclude anything from it, because on this surface an empty source list is a normal delivered result rather than a failure. To do the check repeatedly, across many phrasings, with history you can diff, fetch the answer as structured data instead of reading it in a tab.

The manual way (and why it stops working)

Open Gemini. Ask exactly what a prospective customer would ask — say, "best email marketing platform for an ecommerce brand" — and read what comes back. If Acme is named, Acme is mentioned. If the reply shows links or a grounding panel, look for acme.com in there; if it shows nothing, that's information too, and it's the part people misfile. It takes a minute, it costs nothing, and for a one-off sanity check it is the right tool.

It stops working the second you need to do it twice.

• One question per session. Real category coverage is twenty to fifty phrasings. Typed by hand that shrinks to the three you happen to remember, and the ones you forget are usually where you're invisible.
• The answer isn't stable. Ask the same thing twice and the named brands can differ. A single reading tells you what Gemini said once, not what it tends to say.
• Nothing is recorded, so nothing can be compared. The useful question is whether you're named more often than you were in June. June has to exist somewhere for that question to have an answer.
• Sources are easy to misread in the app. Some answers show links, some show none, and by eye it's genuinely hard to tell whether Gemini grounded an answer or simply didn't need to — so you conclude "we're not cited" about an answer that cited nobody at all.
• It ends at the browser. A read-through can't become a row in a client report, a weekly job, or an alert. It happens right up until the week nobody has twenty minutes.

Do it with one API call

AgentGEO fetches the answer for you and hands it back as JSON: the full answerText, plus sources[] with the title, URL and position of anything the answer cited. Same reply you'd have read in the app, in a shape your code can test.

curl -s https://api.agentgeo.org/v1/fetches \
  -H "Authorization: Bearer ag_live_your_key_here" \
  -H "Content-Type: application/json" \
  -d '{"query": "best email marketing platform for an ecommerce brand", "surfaces": ["gemini"]}' \
  | jq '.answers[0] | {status, answerText, sources}'

A typical record looks like this. It is delivered, it charged a credit, it names brands — and it cites nothing, because the model answered from what it already holds:

{
  "id": "run_84c1f2ab91d3",
  "query": "best email marketing platform for an ecommerce brand",
  "surfaces": ["gemini"],
  "status": "completed",
  "recordsDelivered": 1,
  "creditsCharged": 1,
  "answers": [
    {
      "surfaceKey": "gemini",
      "status": "delivered",
      "answerText": "For an ecommerce brand the platforms suggested most often are Globex, which leads on automation, and Acme, which is the quickest to connect to a storefront ...",
      "sources": [],
      "fetchedAt": "2026-07-14T09:26:04Z",
      "latencyMs": 14310
    }
  ]
}

So write the check in that order: the mention first, then the grounding question, then the citation. Collapsing the last two is what produces phantom declines.

import requests
from urllib.parse import urlparse

BRAND = "Acme"
DOMAIN = "acme.com"
QUERY = "best email marketing platform for an ecommerce brand"


def host(url):
    h = urlparse(url).netloc.lower()
    return h[4:] if h.startswith("www.") else h


resp = requests.post(
    "https://api.agentgeo.org/v1/fetches",
    json={
        "query": QUERY,
        "surfaces": ["gemini"],
        "country": "US",  # Gemini takes country; it does not take language
    },
    headers={"Authorization": "Bearer ag_live_your_key_here"},
    timeout=200,  # live surfaces are slow - requests wait up to 180s
)
resp.raise_for_status()
answer = resp.json()["answers"][0]

if answer.get("status") == "failed":
    snapshot = (answer.get("providerFields") or {}).get("snapshot_id")
    raise SystemExit(f"no record - retry gemini alone with snapshot_id {snapshot}")

text = answer.get("answerText") or ""
sources = answer.get("sources") or []

# Signal 1 - the primary one on a chatbot surface.
named = BRAND.lower() in text.lower()

# Signal 2 - only worth asking once you know the answer was grounded.
cited = [
    s for s in sources
    if host(s["url"]) == DOMAIN or host(s["url"]).endswith("." + DOMAIN)
]

print(f"query:    {QUERY}")
print(f"named:    {named}")
print(f"grounded: {bool(sources)} ({len(sources)} source(s))")

if not sources:
    # A delivered record with no sources is a real result, not an error.
    print("cited:    n/a - Gemini answered without citing a page")
else:
    print(f"cited:    {bool(cited)}")
    for s in cited:
        print(f"  #{s['position']} {s['title']} - {s['url']}")

One delivered record costs one credit; a failed record costs nothing. Loop the same script over your query set and you have a visibility table for the surface — with a grounded column that most tools throw away.

If you already track Perplexity, it's worth seeing how differently the same fields behave here:

Let your agent do it

If the check belongs in a conversation rather than a script — the way it usually does when you're mid-edit on the page in question — install the MCP server and ask in plain language:

claude mcp add agentgeo -- npx -y agentgeo-mcp --key ag_live_...

agentgeo-mcp runs over stdio with zero npm dependencies on Node.js 18+, and exposes exactly one tool: fetch_raw_answers. query and surfaces are required; country defaults to "US". Then ask for the three signals, in words:

Fetch what gemini answers for "best email marketing platform for an ecommerce brand". Tell me whether Acme is named in the answer text. Then tell me whether the answer cited anything at all — and if it did, whether acme.com is in the list.

The advantage isn't saving fifteen lines of Python — it's that the agent already has your repo open, so it can go from "named but ungrounded" to "here's the page that should be carrying this claim" without you assembling the context by hand. The same server works in Claude Code, Cursor, Zed and any other MCP client, and the Python path hits the identical endpoint.

Common problems

• Empty is not failed. A record with status: "delivered" and sources: [] is correct, complete and chargeable. Branch on status first and on len(sources) second; conflating the two turns ordinary answers into imaginary outages.
• language isn't forwarded to Gemini. The chatbot datasets behind this surface accept only a URL, a prompt and a country, and a language field is rejected outright — so shape the market with country and read each market as its own series. Don't assume settings carry across surfaces either: Google AI Overviews does honour language.
• web_search is honoured only by chatgpt. Setting it and expecting Gemini to browse gets you nothing, because it's ignored on this surface. Grounding is something you observe here, not something you request.
• Short brand names collide. A substring test for "Acme" also matches "Acme Insurance" and "Acmegraph". Use a word-boundary regex — re.search(r"\bacme\b", text, re.I) — or a longer, more distinctive string, and spot-check the raw answerText whenever a number jumps.
• Slow scrapes come back failed. Requests wait up to 180 seconds and some scrapes exceed that budget: the record returns failed with a providerFields.snapshot_id and costs zero credits. Retry with that id and the same single surface to redeem the finished answer instead of paying to scrape it again.

Summary

For a one-off, open Gemini and read the answer — it's free and it's fast. The moment the check has to repeat, cover more than a couple of phrasings, or produce something comparable to last month, fetch the record and test the signals in the right order: named, grounded, cited. On this surface the mention carries most of the meaning, and knowing when a missing citation means nothing is what keeps the number honest.