We asked ChatGPT 16 shopping, software, local and how-to questions and sent the same questions to the OpenAI API with web search. ChatGPT cited 49 pages and the API cited 11 of them. What the raw response adds for GEO: the use case label, every page read with its passage, entity cards, and the feeds behind places and products.
How we collected the answers
Each question was one CHATGPT task on querying.ai with include.rawResponse. The task opens a ChatGPT session in the chosen country with web search on and returns the answer together with the stream of events the ChatGPT web app received. 14 questions were asked from the United States and 2 from South Korea, between 17:59 and 18:01 UTC on November 10, 2026. Every answer came from gpt-6 in ChatGPT's logged-out experience (plan_type: guest).
The same 16 questions went to the OpenAI Responses API within the hour: model chat-latest, the web_search tool set to the same country and required on every call, and web_search_call.action.sources included so the response lists the pages the search consulted. The request:
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "chat-latest",
"input": "best CRM for a small team",
"tools": [{ "type": "web_search", "user_location": { "type": "approximate", "country": "US" } }],
"tool_choice": "required",
"include": ["web_search_call.action.sources"],
"store": false
}'Each question ran once on each side, and ChatGPT writes a new answer on every run, so read the counts as one day's sample of 16. How many runs a stable number takes is its own post.
| Question | Use case | Cited / returned | Cards |
|---|---|---|---|
| best running shoes for flat feet | shopping | 0 / 11 | 5 products |
| best robot vacuum for pet hair under $500 | shopping | 3 / 13 | 4 products |
| 건성 피부에 좋은 수분크림 추천해줘 (a moisturizer for dry skin, asked from Korea) | shopping | 3 / 13 | 5 products |
| best CRM for a small team | instant search | 3 / 12 | 5 software tools |
| best project management software for remote teams | instant search | 3 / 60 | 7 software tools |
| HubSpot vs Salesforce for a 20-person startup | instant search | 3 / 23 | 2 software tools |
| is Notion good for project management | instant search | 2 / 12 | 5 software tools |
| best coffee shops near me for working on a laptop | local | 3 / 13 | 6 places |
| 강남역 근처 피부과 추천해줘 (a dermatologist near Gangnam Station, asked from Korea) | local | 3 / 12 | 6 places |
| how does compound interest work | instant search | 0 / 14 | None |
| what is generative engine optimization | instant search | 3 / 14 | None |
| is creatine safe to take every day | instant search | 5 / 20 | None |
| how to remove a red wine stain from carpet | instant search | 3 / 14 | None |
| what are the latest AI model releases this month | search | 9 / 28 | 8 companies |
| 3 day itinerary for Tokyo in spring | local | 2 / 12 | 6 places |
| best business credit card for startups | search | 4 / 21 | 5 products |
What each response carries
The API response for the CRM question, trimmed, holds the searches, a list of URLs and the answer text:
{
"output": [
{
"type": "web_search_call",
"action": {
"type": "search",
"queries": [
"best CRM small business 2026 HubSpot free CRM pricing seats",
"Pipedrive pricing 2026 CRM official",
"Zoho CRM pricing free 3 users official",
"monday sales CRM pricing official 2026"
],
"sources": [
{
"type": "url",
"url": "https://www.spotsaas.com/blog/best-crm-software-2026-reviews-pricing-feature-comparison"
},
{
"type": "url",
"url": "https://www.zoho.com/en-us/crm/zohocrm-pricing.html?price=FREE"
}
]
}
},
{
"type": "message",
"content": [
{
"type": "output_text",
"text": "… I'd recommend **Pipedrive, HubSpot, or Zoho CRM** as the top three options in 2026. …",
"annotations": [
{
"type": "url_citation",
"start_index": 802,
"end_index": 925,
"title": "Best CRM for Small Business in 2026: 10 CRMs Compared on Price and Free Plans",
"url": "https://www.spotsaas.com/blog/best-crm-software-2026-reviews-pricing-feature-comparison?utm_source=openai"
}
]
}
]
}
]
}The ChatGPT stream for the same question holds the same three things and the data the app uses to draw the answer. querying.ai parses the stream into named fields and returns the stream itself as rawResponse:
| Data | OpenAI API | ChatGPT raw response |
|---|---|---|
| Answer | Markdown text | Markdown with component tags for cards, badges, pictures and forms (text, markdown, layout) |
| Searches | action.queries | searchQueries |
| Pages the search returned | action.sources, the URL of each page | Title, site, passage, publish date and result type of each page (sources[]) |
| Citations | url_citation: URL, title and character range | The cited pages and the sentence each one backs (cited, citedText, citationPills) |
| Use case of the question | None | turn_use_case in the closing metadata event |
| Named entities | None | Each name with its type (entities[]) |
| Badges on picks | None | Badge nodes in layout |
| Follow-up form | None | Every question and option (forms[]) |
| Places | None | Feed, rating, review count, coordinates and hours (map[]) |
| Products | None | Price, merchant, rating and offers (inlineProducts[]) |
| Pictures | None | The image search ChatGPT wrote and the page each picture came from (images[], layout) |
| Model and session | model | model, plan, mode and thinking time |
One raw stream held 37 to 343 events and 80 KB to 1.6 MB of JSON, with a median of 268 KB. The parsed fields came to a median of 34 KB. One API response was 4.5 to 17.4 KB.
The app and the API give two answers
| Question | Searches | Pages cited | Cited by both |
|---|---|---|---|
| best running shoes for flat feet | 1 / 2 | 0 / 2 | 0 |
| best robot vacuum for pet hair under $500 | 1 / 4 | 3 / 2 | 1 |
| 건성 피부에 좋은 수분크림 추천해줘 (a moisturizer for dry skin, asked from Korea) | 1 / 4 | 3 / 4 | 0 |
| best CRM for a small team | 1 / 4 | 3 / 6 | 1 |
| best project management software for remote teams | 5 / 4 | 3 / 4 | 0 |
| HubSpot vs Salesforce for a 20-person startup | 2 / 2 | 3 / 3 | 1 |
| is Notion good for project management | 1 / 2 | 2 / 5 | 1 |
| best coffee shops near me for working on a laptop | 1 / 1 | 3 / 0 | 0 |
| 강남역 근처 피부과 추천해줘 (a dermatologist near Gangnam Station, asked from Korea) | 1 / 2 | 3 / 2 | 0 |
| how does compound interest work | 1 / 1 | 0 / 0 | 0 |
| what is generative engine optimization | 1 / 2 | 3 / 3 | 0 |
| is creatine safe to take every day | 2 / 2 | 5 / 4 | 2 |
| how to remove a red wine stain from carpet | 1 / 1 | 3 / 2 | 1 |
| what are the latest AI model releases this month | 2 / 4 | 9 / 5 | 1 |
| 3 day itinerary for Tokyo in spring | 1 / 2 | 2 / 4 | 2 |
| best business credit card for startups | 2 / 4 | 4 / 5 | 1 |
| Total | 24 / 41 | 49 / 51 | 11 |
ChatGPT ran 24 searches for the 16 questions and the API ran 41. Of the 292 pages ChatGPT's searches returned, 53 (18%) were among the 431 pages the API consulted. ChatGPT cited 49 pages, the API cited 51, and 11 were the same page. Three answers show where the two part ways.
- Language. Asked in Korean from Korea for a moisturizer for dry skin, ChatGPT searched in English, “best moisturizer cream for dry skin ceramide glycerin hyaluronic acid dermatology recommendations”, cited the American Academy of Dermatology and Cleveland Clinic, and recommended five creams by name. The API ran four Korean searches, one per product, each with the word for “official”. The two read zero pages in common.
- Location. Asked for coffee shops “near me”, the ChatGPT session placed the reader in Chicago, searched “best laptop-friendly coffee shops Chicago Wi-Fi outlets seating work remotely” and mapped six cafes. The API call, set to the United States, asked which city to search.
- Format. Asked how compound interest works, ChatGPT drew a line chart of $10,000 growing at 8% a year to $217,245 in 40 years and linked the Investor.gov calculator. The API ran a calculator query and answered with two tables.
- Referral tag. Links in the ChatGPT answer carry
utm_source=chatgpt.comand links in the API answer carryutm_source=openai, so web analytics separates the two audiences.
| Question | ChatGPT | OpenAI API |
|---|---|---|
| best running shoes for flat feet | Brooks Adrenaline GTS 25 | ASICS Gel-Kayano 33 |
| best robot vacuum for pet hair under $500 | Dreame D20 Pro Plus | Roborock Qrevo S Pro |
| 건성 피부에 좋은 수분크림 추천해줘 | 에스트라 아토베리어365 크림 | 에스트라 아토베리어365 크림 |
| best CRM for a small team | HubSpot CRM | Pipedrive |
| best project management software for remote teams | ClickUp | ClickUp |
| best business credit card for startups | Chase Ink Business Unlimited | Chase Ink Business Unlimited |
| 강남역 근처 피부과 추천해줘 | 아름다운나라피부과의원 | 클린업피부과 강남점 |
| best coffee shops near me for working on a laptop | Capital One Café | Asked for a city |
The first pick matched in 3 of the 7 questions where both named one. Across the 9 recommendation answers, 27 of the 49 names ChatGPT showed as cards appear anywhere in the API's answer to the same question. A brand that checks its visibility through the API reads a shortlist that buyers never see. The API answer also costs more, and we measured its bill.
The use case ChatGPT assigns to a question
The stream closes with a metadata event. Its turn_use_case field is ChatGPT's own label for the question, next to the model, the plan and the mode of the session. From the robot vacuum answer, trimmed:
{
"type": "server_ste_metadata",
"metadata": {
"model_slug": "gpt-6",
"plan_type": "guest",
"requested_model_experience": "instant",
"turn_use_case": "shopping",
"tool_invoked": true,
"tool_name": "SonicBrowserTool",
"did_auto_switch_to_reasoning": false
}
}shoppingon the three product questions. The answers carried product cards, and the robot vacuum answer carried four cards with a price and a merchant.localon the two place questions and the Tokyo itinerary. All three drew a map with a record for each place.searchon the news question and the business credit card question. ChatGPT ran two searches for each.instant searchon the other eight: four software questions, a definition, an explainer, a health question and a how-to question.
The label tells you what a question competes on. A shopping question is decided on product cards and store catalogs, a local question on place records, and an instant search question on the pages ChatGPT reads and cites. All four software questions were instant search, so a SaaS brand competes on pages.
Every page ChatGPT read, with the passage
The API lists the pages its search consulted as URLs. The raw response lists each search result with its title, site, publish date and the passage handed to the model, and querying.ai returns them as sources[] with the search that found each page. A cited page from the CRM answer, trimmed:
{
"position": 2,
"url": "https://www.spotsaas.com/blog/best-crm-software-2026-reviews-pricing-feature-comparison",
"label": "Best CRM for Small Business in 2026: 10 CRMs Compared on Price and Free Plans",
"attribution": "spotsaas.com",
"datePublished": "June 4, 2026",
"description": "BEST CRM FOR SMALL BUSINESS IN 2026: 10 CRMS COMPARED ON PRICE AND FREE PLANS … Published June 4, 2026 Updated October 1, 2026 …",
"searchQueries": [
"best CRM for small sales team 2026 HubSpot Pipedrive Zoho Freshsales pricing features"
],
"cited": true,
"citedText": [
"Pricing is indicative, generally based on annual billing, and can vary by plan and promotion. Sources:"
]
}- 292 pages came back from 24 searches, a median of 13.5 per answer: 273 web results, 18 news results and 1 academic result.
- 269 of the 292 (92%) carry the passage, a median of 202 characters from the opening of the page: the title, the byline, the date and the first sentences.
- 155 (53%) carry a publish date. 127 of those were published within the last year and 46 within the last 30 days.
- 48 of the 292 (16%) were cited, a median of 3 per answer. The first result was cited in 10 of the 14 answers that cited pages.
The passage is what ChatGPT receives with each result, and it starts at the top of the page. In the first result for the CRM question, the passage read, after the page title, “HubSpot CRM is the best CRM for most small businesses in 2026, thanks to its free plan for up to 2 users”, and the answer gave HubSpot CRM its “Best overall” badge.
The searches also show the shortlist ChatGPT starts from. The CRM search named HubSpot, Pipedrive, Zoho and Freshsales before ChatGPT read a page, and all four became cards. Across four recommendation questions, 17 of the 18 names written into the searches became cards. Query fan-out covers how these searches are written.
Entities, badges and the follow-up form
In the raw stream the answer is markup. ChatGPT wraps each name it treats as an entity in a tag with a category, puts a badge on its picks and writes an image search beside each card. The first card of the CRM answer, trimmed:
<row align=start gap={3}>
<AsyncImage query="HubSpot CRM software logo" aspectRatio="1:1" maxWidth="76px"/>
<box flex="1" gap={1}>
**1. <Entity value="HubSpot CRM" category="software"/>**
<badge color="success">Best overall</badge>
- Easy to learn and implement.
- Free option for up to 2 users.
Best for: Small businesses wanting an all-around CRM.
<text color="secondary" size="xs"><Link url="https://www.hubspot.com/products/crm" title="Website and plans"/></text>
</box>
</row>- 64 entity cards in 12 answers: 19 products, 19 software tools, 17 local businesses, 8 companies and 1 point of interest. A reader who clicks a card opens its detail panel. The four answers without cards were the definition, the explainer, the health question and the how-to question.
- 16 badges in 10 answers, such as “Best overall”, “Best for sales teams” and “Best for avoiding a personal guarantee”. A badge is ChatGPT's one-line positioning of a brand in its own words.
- A follow-up form in 11 answers, every answer with cards except the news answer. The CRM form asked for team size, business type, the main job of the CRM, monthly budget per user and current tools. The form lists the criteria ChatGPT uses to narrow a category.
querying.ai returns these as entities[], badge nodes in layout and forms[]. In the API answer the same names are bold text and the picks are headings.
The feeds behind places and products
Place and product cards are filled from data feeds. A place record from the Chicago answer, trimmed:
{
"name": "Overflow Coffee",
"provider": "yelp-feed",
"rating": 4.4,
"review_count": 111,
"is_claimed": true,
"price_str": "$$",
"categories": [
"Coffee shop"
],
"website_url": "https://www.overflowchicago.com/",
"provider_url": "https://www.yelp.com/biz/overflow-coffee-chicago-2?adjust_creative=openai&utm_campaign=yelp_feed&utm_medium=feed_v2&utm_source=openai"
}- Chicago coffee shops: six places from Yelp's feed, rated 4.0 to 4.5 with 111 to 645 reviews. The raw record adds
is_claimed, true for all six,price_str, opening hours and tags. - Dermatologists near Gangnam Station: three place records from Google Maps, rated 4.5 to 5.0 with 32 to 178 reviews.
- The Tokyo itinerary: nine places on the map, seven from Tripadvisor and one from Google Maps.
- Robot vacuums: four product cards priced $359.99 to $419.99, each with a rating, a merchant and its offers. Two link to the maker's own store, Dreame and Eureka, with
utm_campaign=openai_catalog, and two to Walmart and Lowe's.
A Yelp, Google Maps or Tripadvisor listing and the product catalog behind a store are GEO surfaces alongside web pages. querying.ai returns the records as map[] and inlineProducts[].
Where the pictures come from
ChatGPT writes an image search for each card, such as “HubSpot CRM software logo”, and the response records the page each picture came from. Of 47 pictures drawn beside a named brand, product or place in 9 answers, 11 came from that brand's own domain.
- All seven project management tools showed a product screenshot from the vendor's own domain.
- HubSpot's logo came from a software review site, and four of the five CRM logos came from third-party sites.
- The Brooks Adrenaline GTS 25 photo came from a UK running store, and one of the five shoe photos came from the maker's site.
- The photo of Dollop Coffee came from Yelp and the photo of Philz Coffee from Eater Chicago.
What to do with it
- Measure the app. Buyers read the ChatGPT app's answer, and the API returns another shortlist. Collect the app's answer in each country you sell in with Answers.
- Sort your questions by use case. Pair
shoppingquestions with catalog work,localquestions with listings, andinstant searchandsearchquestions with content. - Open the page with the claim. The passage attached to a search result is the opening of the page, about 200 characters. State what the product is, who it is for and the number that proves it in the first two sentences.
- Answer the form. The follow-up form lists what ChatGPT asks a buyer next, such as team size, budget and main use. Give each of those cases a clear answer on product and comparison pages.
- Complete the feeds. Places arrived with ratings, review counts, hours and a claimed flag, and products with a price, a merchant and offers. Fill every field of the listing and the catalog.
- Host the picture. ChatGPT searches for a logo, a product photo or a dashboard screenshot by name. Publish a clean one on your own domain, on a page whose title names the product.
- Track the questions over time with a monitor. The share of answers that carry each field is charted daily on the model tracker.
Collect the raw response
Send a CHATGPT task with include.rawResponse. The named fields arrive on every answer, and the flag adds rawResponse, the array of stream events.
curl -X POST https://api.querying.ai/v1/async/task \
-H "Authorization: Bearer $QUERYING_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"taskType": "CHATGPT",
"payload": {
"prompt": "best CRM for a small team",
"country": "US",
"include": { "rawResponse": true }
}
}'The metadata event, the search results and the place records sit in the stream under ChatGPT's own names, so a field ChatGPT adds tomorrow is in your data the same day. An answer costs 1 credit with or without the flag. The fields are on the ChatGPT engine page, and response options lists the flags.
Frequently asked questions
What is ChatGPT's raw response?
The stream of events the ChatGPT web app receives while it writes an answer: the answer markup, the searches, the search results, the cards and a closing metadata event. querying.ai returns it as rawResponse when a CHATGPT task sets include.rawResponse.
Is the OpenAI API answer the same as the ChatGPT answer?
They are two answers. Across 16 questions ChatGPT cited 49 pages and the API cited 11 of them, and the first recommendation matched in 3 of 7 questions. The API follows your call and its settings, and the app runs its own searches for where the reader is.
Which data is only in the raw response?
The use case label (turn_use_case), the plan and mode of the session, the thinking time, is_claimed and price_str on place records, and the safety check ChatGPT ran on each link. The rest of the data in this post also arrives as parsed fields: sources, searchQueries, entities, layout, forms, map and inlineProducts.
How large is a raw response?
80 KB to 1.6 MB in this sample, with a median of 268 KB. The Tokyo itinerary, with nine place records and their reviews, was the largest. Request it for the answers you study and leave it off for routine tracking.
How many answers make a reliable number?
ChatGPT writes a new answer on every run, so one answer is one draw. This post reads 16 answers from one day. For a visibility rate, run each question several times in each country; the sample size guide gives the numbers.
Put it to work
- ChatGPT Scraper API for Answers & Citations
Scrape consumer ChatGPT answers, citations and search queries as JSON. Build brand visibility tracking with async requests, polling and webhooks.
- AI Answer Scraping API for GEO Data
Collect consumer AI answers from ChatGPT, Gemini, Perplexity, Google, Meta AI, Qwen, Doubao, Baidu and Naver through one async API, with source citations and market settings.
- AI Visibility API for GEO Monitoring
Build AI brand visibility tracking with scheduled prompts, mentions, citations and competitor share of voice. Connect GEO monitoring to your product via API.


