2026-09-18, Advertising · ChatGPT

Ads in ChatGPT answers: how they work and the ads[] fields

ChatGPT answers now carry advertising. OpenAI began testing ads in ChatGPT in the US on February 9, 2026, for logged-in adult users on the Free and Go tiers, with Plus, Pro, Business, Enterprise and Education tiers excluded; it later announced pilots in Canada, Australia and New Zealand (update of March 26, 2026) and plans for the United Kingdom, Mexico, Brazil, Japan and South Korea (update of May 7, 2026) (OpenAI, “Testing ads in ChatGPT”, checked 2026-09-17). If you buy media or watch competitors, the ads attached to your category’s prompts are now worth monitoring. One limit first: result.ads[] exists in the API, but it comes back empty on the ChatGPT answers served today, so the steps below describe the field and the method, not data you can collect now. Below: requesting them with include.ads, what each field means, and turning the results into advertiser share-of-voice numbers.

What OpenAI says about how the ads work

The same OpenAI post, checked 2026-09-17, states that:

• ads are “clearly labeled as sponsored and visually separated from the organic answer”;
• ads do not influence the answers ChatGPT gives;
• during the test, ads are matched to the topic of the conversation, past chats and past interactions with ads, and when several advertisers match, the most relevant one is shown first;
• ads are not eligible to appear near sensitive or regulated topics such as health, mental health or politics.

Two consequences for measurement. First, because matching uses conversation and account context, one answer shows a single serving decision, not what every user sees; measure frequencies over many runs. Second, an empty ads[] on a health prompt is expected and says nothing about competitor spend.

Definitions

• Served ad: an ad included in ChatGPT’s ad response for an answer, whether or not it was displayed.
• Rendered ad: the served ad displayed as a visible card. At most one is displayed per response.
• Ad presence rate: the share of answers for a prompt set in which a given advertiser is served (or rendered, reported separately).
• Paid share of voice: an advertiser’s rendered ads divided by all rendered ads across the prompt set and period.

The request

Add include.ads to a ChatGPT request. Synchronously:

curl -X POST https://api.answerline.dev/v1/monitor/chatgpt \
  -H "Authorization: Bearer $ANSWERLINE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "prompt": "best project management tools for small agencies", "country": "US", "include": { "ads": true } }'

Or as an async task, which fits prompt sets:

{
  "taskType": "CHATGPT",
  "payload": {
    "prompt": "best project management tools for small agencies",
    "country": "US",
    "include": { "ads": true }
  },
  "idempotencyKey": "ads-pm-small-agencies-us-2026-09-18"
}

The response fields

When include.ads is true, result.ads is an array with one entry per served ad. On the ChatGPT answers served today it comes back empty; the shape below is the contract:

{
  "success": true,
  "result": {
    "text": "…",
    "sources": [ { "position": 1, "url": "…", "label": "…", "description": "…" } ],
    "ads": [
      {
        "brand": { "name": "…", "url": "…", "favicon": "…" },
        "cards": [ { "title": "…", "body": "…", "url": "…", "image": "…" } ],
        "visibility": "allowed",
        "adsResponseIndex": 0,
        "rendered": true
      }
    ]
  }
}

Use rendered for exposure. Do not infer exposure from adsResponseIndex 0: the index reflects served order only.

Served-but-not-rendered entries show which advertisers were in contention for the prompt even when another brand’s card was displayed, which is useful for competitive intelligence as long as you label it as “served”, not “seen”.

An ad-tracking run

1. Build the prompt set

Take the commercial prompts from your organic prompt set: category (“best X for Y”), comparison (“X vs Y”) and purchase-intent (“where to buy X”) questions. Informational and sensitive prompts rarely matter for ads. Designing a prompt set covers the rest.

2. Choose markets deliberately

Run the markets where you advertise or plan to. ChatGPT answers from 213 countries, including pilot markets such as the US, Canada, Australia, the United Kingdom, Brazil and Japan. Do not read an empty result as “nobody advertises here”: ads[] comes back empty on the ChatGPT answers served today in every market.

3. Submit and collect

import collections
import datetime as dt
import os
import time
import requests

API = "https://api.answerline.dev"
HEADERS = {"Authorization": f"Bearer {os.environ['ANSWERLINE_API_KEY']}"}
PROMPTS = {
    "pm-small-agencies": "best project management tools for small agencies",
    "pm-vs": "Asana vs Monday for a marketing team",
    "time-tracking": "which time tracking app should a consultancy buy",
}
MARKETS = ["US", "CA", "AU"]
day = dt.date.today().isoformat()

tasks = [
    {
        "taskType": "CHATGPT",
        "payload": {"prompt": p, "country": c, "include": {"ads": True}},
        "idempotencyKey": f"ads-{pid}-{c.lower()}-{day}",
    }
    for pid, p in PROMPTS.items()
    for c in MARKETS
]
batch = requests.post(f"{API}/v1/async/task/batch", json=tasks, headers=HEADERS, timeout=60)
batch.raise_for_status()
ids = [r["task"]["id"] for r in batch.json()["results"] if r["success"]]

results = {}
while len(results) < len(ids):
    for tid in ids:
        if tid in results:
            continue
        data = requests.get(f"{API}/v1/async/task/{tid}", headers=HEADERS, timeout=30).json()
        if data["task"]["status"] in ("COMPLETED", "FAILED"):
            results[tid] = data
    time.sleep(10)

For recurring runs, replace polling with a webhook on each task (webhooks).

4. Aggregate served and rendered ads

served = collections.Counter()
rendered = collections.Counter()
creatives = collections.defaultdict(set)
answers = 0

for data in results.values():
    if data["task"]["status"] != "COMPLETED":
        continue
    answers += 1
    for ad in data["response"]["result"].get("ads", []):
        name = ad["brand"]["name"]
        served[name] += 1
        if ad["rendered"]:
            rendered[name] += 1
        for card in ad["cards"]:
            creatives[name].add(card["title"])

total_rendered = sum(rendered.values())
for name, count in served.most_common():
    print(
        f"{name:30} served {count / answers:6.1%} "
        f"rendered {rendered[name] / answers:6.1%} "
        f"paid SoV {rendered[name] / total_rendered if total_rendered else 0:6.1%} "
        f"creatives {len(creatives[name])}"
    )

Store the raw task JSON with the date, prompt id and market. Card titles and bodies over time are a record of competitors’ messaging tests.

Worked example

These numbers are invented to show the calculation. 30 prompts in one market give 30 completed answers. Brand A is served in 12 and rendered in 7; brand B is served in 9 and rendered in 5; no other brand is rendered. Brand A’s served presence rate is 12 / 30 = 40%, its rendered presence rate 7 / 30 = 23.3%, and its paid share of voice 7 / (7 + 5) = 58.3%. Brand B is served in 9 answers but rendered in only 5, so in 4 answers it was in contention and another ad (or none) was displayed.

What to do with the data

• Competitor ad monitoring: which advertisers appear on your category prompts, in which markets, and how often they are rendered.
• Paid vs earned in one response: compare ads[] with organic sources[] and brand mentions in text from the same answer. A competitor rendered in ads but absent from citations is buying coverage it has not earned organically; the reverse suggests you may not need to buy.
• Creative tracking: cards[].title and cards[].body per brand over time show which messages advertisers test.
• Market entry: run your category prompts in a pilot market before launching there and see who already advertises.
• Filtered ads: a share of visibility: "hidden" entries for a brand shows ads OpenAI filtered; report them apart from allowed ones.

For the organic side of the same answers, see how AI engines choose citations and measuring brand mentions.

Ads on Google surfaces

Google’s surfaces carry ads too, and the Google endpoints return ads with their own shapes.

AI Overview ads need include.aioverview on the Google request (+$0.00071 on the $0.00103 Google Search task). A combined view (ChatGPT ads, Google text ads and AI Overview ads for the same keyword) shows where a competitor concentrates spend. See Google AI Overviews monitoring, SERP feature tracking and the Google Search engine page.

Cost

A ChatGPT task is $0.0006 at the Starter tier, and include.ads and include.shopping are free. So a task with ads, or with ads and shopping, is $0.0006.

New accounts get 100 free requests on any engine, ChatGPT included, enough for 100 tasks with ads; after that you top up your balance. See pricing for tier prices and cost planning for budgeting.

Pitfalls

• Treating served as seen. Only rendered: true was displayed. Report served and rendered separately.
• Reading adsResponseIndex as rank on screen. It is served order.
• One run per prompt. Ad selection depends on context, so single runs swing. Aggregate across runs and days.
• Mixing field shapes. ChatGPT ads[] has brand and cards; Google ads[] has title, url and domain. Normalize to advertiser domain before combining.
• Sensitive prompts. OpenAI says ads are not eligible near health, mental health or political topics; exclude them from paid share-of-voice denominators.
• Assuming a fixed market list. OpenAI has changed pilot markets over time; re-check its announcements before planning coverage.

Field definitions are in the API reference; engine details on the ChatGPT engine page.

Keep reading