AI engines
Does OpenAI's web search tool give the same answer as ChatGPT?
No. The API answers from the model until you attach the web_search tool, and the model still decides: searches cost $10 per 1,000 calls plus the tokens they feed it.

By Mathias C.SEO consultant, and more recently GEO, at MentionBox
No. The OpenAI API answers from the model's own knowledge until you attach the web search tool to the request, and even then the guide for that tool states that "like any other tool, the model can choose to search the web or not based on the content of the input prompt". The ChatGPT app works the other way round: OpenAI's help article describes a product that searches on its own when a question would benefit from current information, and shows the citations with the answer. A searched API answer is billed in two parts, $10 per 1,000 search calls plus the search content tokens at model rates, which comes to about $14 per 1,000 answers on gpt-4.1-mini.
This page is for the developer who sent one web_search request, compared it with the same question typed into chatgpt.com, and found a shorter answer with fewer links. Below: what the API actually returns, the five documented differences, the settings that narrow the gap, and the bill.
What the tool returns
Web search is a hosted tool declared in the tools array of a Responses API call:
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6-astra",
"input": "best AI visibility API for an agency",
"tools": [{"type": "web_search"}],
"tool_choice": "required",
"include": ["web_search_call.action.sources"]
}'
Two kinds of output item come back. A web_search_call item records what the tool did, with an action whose type is search, open_page or find_in_page; the search action carries the queries the model ran and the sources it consulted. A message item carries the text, with url_citation annotations that give the url, the title and the character range each citation covers, per the API reference. Without the include line above you see only the inline citations, and the guide notes that "the number of sources is often greater than the number of citations".
The absence of a web_search_call item is a result too: it means the model answered from memory, which is the single most common reason a developer reports an answer with no links.
Five documented differences
| What differs | ChatGPT app | Responses API with web_search |
|---|---|---|
| Decision to search | searches when the question would benefit from fresh information; a user can also pick Search by hand | optional under tool_choice: "auto", forced by tool_choice: "required" |
| Model and instructions | OpenAI's product, its system instructions and the account's context | the model and the prompt you send, nothing else |
| Search budget | not published | search context capped at 128k tokens, even on a model with a larger window |
| What you can steer | the Search tool, on or off | filters with up to 100 allowed or blocked domains, user_location, search_context_size at low, medium or high |
| What you receive | an answer with the citations the product shows | queries, sources, url_citation annotations, and the search state itself |
Behaviour and parameter names from OpenAI's web search guide, its API reference and its pricing page, read on October 28, 2026.
A sixth difference is undocumented, and worth stating as such. OpenAI's crawler page presents OAI-SearchBot as the agent used "to surface websites in search results in ChatGPT's search features", and adds that sites opted out of it "will not be shown in ChatGPT search answers, though can still appear as navigational links". Nothing on that page describes the fetch path behind an API search call, so treat the two source pools as separate until OpenAI says otherwise.
Settings that narrow the gap
- Force the search with
tool_choice: "required", or move to Chat Completions, where only specialised search models such asgpt-5-search-apisupport web search and, in the guide's words, "always search before responding". - Set
user_locationto the country and city you want to measure, since the app answers a located user and a bare API call does not. - Raise
search_context_sizetohighwhen the question is comparative, keeping in mind the 128k cap and the extra input tokens it buys. - Send an empty or minimal system prompt when your goal is to observe the engine rather than to build a product, and log the prompt with the answer either way.
What a searched answer costs
| Setup | Billed for one searched answer |
|---|---|
web_search on gpt-4.1-mini | $0.01 for the call, plus a fixed block of 8,000 input tokens at $0.40 per 1M, so about $14 per 1,000 answers of 500 output tokens |
web_search on gpt-5.5 | $0.01, plus what the search feeds at $5 per 1M input, up to $0.64 at the 128k cap |
web_search on gpt-6-astra | $0.01, plus the same content at $10 per 1M input, up to $1.28 at the cap |
| web search preview tool, non-reasoning model | $25 per 1,000 calls, search content tokens free |
Chat Completions, gpt-5-search-api | tokens at $1.25 per 1M in and $10 per 1M out, search always run |
Rates from OpenAI's pricing page, read on October 28, 2026. The 8,000-token block and the 128k search context limit come from its web search guide; the per-1,000 line assumes 500 output tokens per answer.
The search content, not the search call, is the variable that moves a bill. One comparative prompt on a large model can read more than a hundred times the tokens of a lookup on gpt-4.1-mini, for the same $0.01 of tool call.
Reading the consumer answer instead
Vendors that collect the app say so in writing, and that sentence is the thing to check before buying. cloro's documentation states that "cloro does not call the ChatGPT API: each request runs a real browser session against ChatGPT's web interface", with web search forced on and a disableWebSearch parameter to turn it off. DataForSEO sells both routes side by side: its LLM Scraper reads the interfaces, while its LLM Responses endpoint calls the official model APIs and documents web_search with a "default value: false", checked on October 23, 2026. API for GEO runs each prompt natively on each engine, in the country and language asked for, and returns brand mentions, position, share of voice, sentiment and the cited sources at $13 per 1,000 answers. MentionBox, which publishes this site, has a commercial link with API for GEO.
Frequently asked questions
Does the OpenAI API search the web by default?
No. The model answers from its own knowledge until you add `{"type": "web_search"}` to the request's tools array, and OpenAI's guide for the tool states that "like any other tool, the model can choose to search the web or not based on the content of the input prompt". With the default `tool_choice: "auto"` search is optional, and `tool_choice: "required"` is the documented way to make it run. On Chat Completions the behaviour is reversed: only specialised search models such as gpt-5-search-api support web search, and they always search before answering.
Why does the API web search answer differ from what ChatGPT shows me?
Three documented differences. You pick the model and the instructions, while the app carries OpenAI's own; the app searches whenever a question would benefit from fresh information, while the API model decides under `tool_choice: "auto"`; and the search context of an API call is capped at 128k tokens even on a model with a larger window. OpenAI documents OAI-SearchBot as the crawler that surfaces sites "in search results in ChatGPT's search features" and says nothing about which path serves an API request, so the source pools cannot be assumed identical.
How do I get the full list of pages the model read?
Add `web_search_call.action.sources` to the request's `include` parameter. OpenAI's guide states that "unlike inline citations, which show only the most relevant references, sources returns the complete list of URLs the model consulted when forming its response", and that the number of sources is often greater than the number of citations. The inline citations arrive separately as `url_citation` annotations carrying the url, the title and the character range they cover in the answer.
How much does one searched answer cost through the API?
The search call is $10 per 1,000 calls and the pages it feeds the model are billed as input tokens at the model's rate. On gpt-4.1-mini the tool bills a fixed block of 8,000 input tokens per call, $0.0032 at $0.40 per 1M, so a 500-token answer lands at about $0.014, or $14 per 1,000. The 128k search context cap sets the ceiling: $0.64 of input on gpt-5.5, $1.28 on gpt-6-astra. Rates read on October 28, 2026.
Can I restrict the API search to certain sites?
Yes, with the `filters` parameter, which takes up to 100 `allowed_domains` or `blocked_domains`, written without the http or https prefix. You can also set `user_location` with a country code, city, region or timezone to get results for that place, and `search_context_size` to low, medium or high, which the guide warns "does not set an exact token count or guarantee a specific number of sources or citations". None of those three controls exist for a ChatGPT user.
Can I measure brand visibility in ChatGPT with the official API?
You can measure the model you configured, which is a different figure from what a buyer reads in the app. Vendors that sell the consumer surface say so in writing: cloro's documentation states that "cloro does not call the ChatGPT API: each request runs a real browser session against ChatGPT's web interface", and DataForSEO splits the two into an LLM Scraper endpoint and an LLM Responses endpoint where `web_search` carries a "default value: false". Checked on October 23, 2026.