# How does ChatGPT rewrite your question into search queries?

> ChatGPT Search does not necessarily send your wording directly to a search engine.

Source: https://bestqueryfanouttools.com/chatgpt-fan-out-queries · Publisher: Fan-out · Updated: 2026-11-13

## TL;DR
- ChatGPT Search typically rewrites a request into one or more targeted queries and may run follow-up searches that become more specific.[1]
- Published counts vary by dataset: Ahrefs observed approximately 1.6 queries among responses containing fan-out, while StatWharf found a median of two searches in a 100-prompt software panel.[7][3]
- ChatGPT sometimes searches in English for non-English prompts, but reported rates range from 17.1% in Howseen’s panel to 43% in Peec’s non-English fan-out analysis.[12]
- Observed rewrites add years, vendors, comparison terms, “official” wording and `site:` restrictions, although the frequencies change by dataset and period.[8]
- The exact production rewrite algorithm and a universal ChatGPT query count are not published, so every fan-out claim should include its source, sample and observation window.[6]

ChatGPT Search does not necessarily send your wording directly to a search engine. OpenAI says it typically turns a request into one or more targeted queries and may run additional searches after reviewing initial results.[1] The available evidence suggests that these ChatGPT fan-out queries often add dates, named entities, comparison terms, locations, language changes and domain restrictions—but there is no published universal query count or rewriting formula.[2]

This article separates three kinds of evidence:

- **Official documentation:** what OpenAI says ChatGPT Search can do.[3][4]
- **Observed datasets:** what researchers and vendors have captured in specific panels and periods.[5]
- **Academic research:** frameworks that explain query rewriting without proving how production ChatGPT is implemented.[6]

### What ChatGPT query rewriting looks like

OpenAI’s example begins with the question `what's the latest on the development of drugs that target CCR8 for cancer?`. ChatGPT may first search for `CCR8 immunotherapy drug development 2025`, then refine the research with `CHS-114 conference 2025`.[1]

That sequence illustrates two separate actions:

1. The original question is converted into a retrieval-oriented query.
2. A later search adds a more specific drug or conference-related term.

The example does not establish how frequently ChatGPT performs a second search, nor does OpenAI publish a global average, median or maximum number of queries per prompt.[2]

Context can also change the rewrite. OpenAI gives the example of `What are some good restaurants near me?` becoming `top restaurants San Francisco`; with relevant memory about a vegan preference, the query may become `good vegan restaurants San Francisco`.[1]

The practical implication is that a ChatGPT search query can incorporate information not present in the latest sentence, including:

- Location
- Memory or prior conversation context
- Current-year language
- Search-intent terms
- Named products or companies
- Follow-up entities discovered during research

OpenAI says general location information may be shared with search providers, while the user’s IP address and ChatGPT account information are not shared for the search; device-location sharing is optional and off by default.[1]

## How many queries does ChatGPT run per question?

There is no single reliable ChatGPT-wide number. Published figures measure different populations, model configurations, topics, dates and observation methods, so a number should always be reported with its denominator.[6][2]

### Findings from observed ChatGPT fan-out studies

| Study and observation period | Population | Reported finding | How to interpret it |
|---|---:|---:|---|
| Ahrefs, August–September 2026 | 66.6 million free, logged-out responses; 15 million recorded fan-outs | 14.4% of responses contained fan-out; approximately 1.6 queries among responses with fan-out | The 1.6 figure does not describe every ChatGPT response.[3] |
| StatWharf, September 12, 2026 | 100 prompts across 50 software categories | 211 queries; median of 2 searches per prompt; maximum of 8 | This is a software-recommendation panel, not a product-wide maximum.[7] |
| Howseen, October 7–November 5, 2026 | 17,439 searches from 10,559 answers | Average of 1.65 searches per answer | The sample was principally French and based on customer accounts.[8] |
| Peec, October 2025–January 2026 | More than 20 million ChatGPT fan-outs; country comparisons | Approximately 2.3–2.8 fan-outs per prompt across reported country and time comparisons | These are country and period observations, not a universal range.[9] |

The most defensible summary is therefore conditional: **observed ChatGPT fan-out counts vary by panel, topic, period, interface and capture method**.[7][4]

Ahrefs provides scale but studies a specific setup: free, logged-out responses identified as GPT-5.6 Luna during August and September 2026. Its approximately 1.6 queries applies to responses in which fan-out was recorded, while its 14.4% figure describes the share of responses containing fan-out.[3]

StatWharf provides more detailed query visibility but uses only 100 software-selection prompts. Its median of two searches and observed maximum of eight describe that controlled panel.[10]

Howseen’s average of 1.65 searches per answer comes from a different population: non-English buyer questions collected from October 7 to November 5, 2026.[8] Peec’s figures compare countries and periods, so they should not be combined with the other studies into one global average.[4][9]

### Why the denominator matters

The statement “ChatGPT runs 1.6 searches per question” is misleading because it removes the condition attached to the Ahrefs figure. A more accurate formulation is: **Ahrefs observed approximately 1.6 queries among responses containing recorded fan-out in its free, logged-out August–September 2026 dataset**.[3]

Likewise, “ChatGPT runs two searches” would overstate StatWharf’s findings. The study found a median of two searches in a 100-prompt software-recommendation panel, with one prompt producing as many as eight observed queries.[2][10]

The collection method also affects the result. StatWharf reports that live response streams exposed search arguments while saved conversation records returned blank search-call arguments.[7] Missing query arguments should therefore not automatically be interpreted as proof that ChatGPT did not search.

## Does ChatGPT search in English when the prompt is not English?

Yes, observed audits show that ChatGPT sometimes uses English fan-out queries for non-English prompts. However, the available studies report materially different percentages because they use different languages, topics, panels and classification methods.[11]

| Study | Sample | English-search finding | Main qualification |
|---|---:|---:|---|
| Peec, February 2026 | More than 10 million prompts and 20 million fan-outs | 43% of fan-outs for non-English prompts were in English | Exact model and interface conditions were not published in the cited analysis.[12] |
| Howseen, October–November 2026 | 17,439 searches from non-English questions | 17.1% of searches were English | 349 of 356 non-English brand setups were French; ten brands generated 58.7% of English searches.[8] |
| aclipp, August–October 2026 | German questions about PR software | ChatGPT averaged 84% English queries, with a monthly range of 74–94% | The dataset concerned one German PR-software project.[13] |

These figures are not evidence of one universal “English bias.” They are separate observations from different samples.[8]

aclipp provides an exact example. The German prompt `Welche PR-Tools sind DSGVO-konform und aus der EU?` produced searches including:

- `EU PR media database tool GDPR compliant Germany press distribution`
- `deutsches PR Tool Medienkontakte DSGVO EU Hosting`
- `PR software Europe GDPR media monitoring press distribution EU company`

Two of those three examples are primarily English, while one is German.[2][13]

Howseen reports a lower pooled English share of 17.1% in its predominantly French customer panel. Removing its ten largest contributors reduced the share to 9.7%, showing how strongly account and industry mix can affect the aggregate result.[3]

Peec also distinguishes between conversations and individual searches. Its reported 81.8% French figure describes conversations containing English fan-outs, while its 43% figure describes individual fan-outs.[12] Those are different denominators.

## What does ChatGPT add when it rewrites a query?

Observed ChatGPT fan-out queries often preserve the original topic while adding information that helps retrieval or comparison. Reported additions include years, vendor names, review language, comparison terms, locations, “official” wording and `site:` restrictions.[8]

| Observed pattern | Reported evidence | What it suggests |
|---|---:|---|
| Adding a year | 161 of 211 StatWharf queries, or 76%, contained a year | ChatGPT may introduce recency even when the original prompt has no date.[11] |
| Preserving the category | 143 of 211 queries, or 68%, retained the exact category phrase | Rewriting can keep the user’s subject while expanding the query.[1] |
| Avoiding exact repetition | None of StatWharf’s 211 queries exactly reproduced the prompt | Fan-out is more than simple duplication.[8] |
| Adding evaluation terms | Peec observed `best` in 15.33%, `review(s)` in 6.84% and `comparison` in 4.48% of reported results | Recommendation prompts may acquire evaluation language.[2][11] |
| Adding “official” | Ahrefs observed `official` in 13.1% of English fan-outs in September, compared with 10.1% in August | Query vocabulary can vary over time.[3][11] |
| Restricting domains | Ahrefs found `site:` in 6.8% of its 15 million fan-outs | Some searches explicitly target a domain or site type.[14] |
| Increasing query length | Peec reported average length rising from about 6 words in October 2025 to 12 words in January 2026 | Longer queries do not necessarily mean more queries.[11] |

### Example: a software question becomes a vendor comparison

StatWharf records the prompt `best board portal software` becoming:

`best board portal software Diligent Boards OnBoard BoardEffect Nasdaq Boardvantage iBabs features 2026`

The generated query retains the category, adds several vendors, adds “features” and adds a year. StatWharf reports a median of five words for the original prompts and 11 words for generated queries in its panel.[7]

Peec records a similar pattern in the prompt `Which Samsung Galaxy should I get?`, which became `best Samsung Galaxy phones comparison 2026`.[15]

These examples show why ChatGPT query rewriting matters for AI visibility. A page may need to be discoverable not only for the wording used by a person, but also for the expanded comparisons and attributes that appear in the underlying search queries.[14] For a structured introduction to making pages visible in these answers, Finseo offers a [free GEO/AEO course](https://www.finseo.ai/geo-course) with 30 short video lessons on visibility in ChatGPT, Perplexity, Gemini and AI Overviews.

The patterns are not permanent rules. Promptwatch reported `site:` usage rising from 0.37% to 16.8% of fan-outs on August 8, 2026, while Ahrefs reported 6.8% across its August–September aggregate.[3][6] Those figures come from different datasets and should not be treated as one trend line.

## Does ChatGPT run one search or a sequence of searches?

ChatGPT can run a sequence. OpenAI’s CCR8 example shows an initial targeted query followed by a more specific query about a named drug and conference.[1]

The sequence can be understood as:

1. **Interpret the question:** identify the subject, intent and relevant context.
2. **Construct an initial query:** turn the question into retrieval-oriented language.
3. **Inspect early results:** identify missing entities, dates or evidence.
4. **Refine the search:** add a specific product, organization, location or event.
5. **Prepare the answer:** use retrieved pages and selected citations.

This is a conceptual description of the documented behavior, not a claim that every ChatGPT request follows all five steps.[6]

OpenAI’s API documentation also separates search actions from page operations. A `web_search_call` can contain searched queries, while `open_page` and `find_in_page` represent other actions.[16] Counting tool actions therefore does not automatically count search queries.

The API may expose searched queries, but OpenAI says the query field is present usually, not always.[16] The same documentation distinguishes inline citations from the complete list of sources consulted; applications can request the latter through the `web_search_call.action.sources` field.[16]

That creates three different observability questions:

- **What queries did ChatGPT run?**
- **Which URLs did it retrieve?**
- **Which sources did it display or cite?**

A citation list cannot by itself prove the number or wording of searches. Conversely, a visible query list may not reveal every internal step if the interface or API does not expose it.[4][2]

## What does academic research say about query rewriting?

Academic research provides a useful framework but does not reveal ChatGPT’s production implementation. Ma, Gong, He, Zhao and Duan’s 2023 paper, *Query Rewriting for Retrieval-Augmented Large Language Models*, describes a Rewrite-Retrieve-Read approach.[11]

The framework separates the process into three stages:

1. Rewrite the user’s input into a search query.
2. Retrieve relevant context from an internet search engine.
3. Read the retrieved context and answer the original question.

The paper addresses a common mismatch: a conversational question may not use the wording that retrieves the best evidence. Its experiments use datasets including HotpotQA, AmbigNQ, PopQA and MMLU, with a trainable T5-large rewriter and ChatGPT or Vicuna-13B readers.[11]

The authors report improved retrieval-augmented performance in their experiments.[2][11] That supports query rewriting as a useful research technique, but it does not prove that OpenAI uses the same architecture, model or prompt in ChatGPT Search.

The distinction matters because three claims are often conflated:

- **Query rewriting is a recognised retrieval technique.**[11][14]
- **OpenAI documents that ChatGPT turns requests into targeted searches.**[2][1]
- **The exact production rewriting algorithm used by ChatGPT is publicly known.** It is not published in the cited documentation.[6]

## How should ChatGPT fan-out evidence be interpreted?

The strongest interpretation combines official product documentation with dated observations, while keeping each source’s scope visible.[4]

Use the following rules when analysing ChatGPT search queries:

1. **Attach every count to its population.** Say “in the Ahrefs fan-out subset” or “in StatWharf’s software panel,” not simply “per question.”[7][3]
2. **Separate query count from query length.** Peec observed longer queries over time while reported fan-out counts remained broadly stable.[9]
3. **Separate prompt language from query language.** An English fan-out can occur inside a non-English conversation, but the reported percentages use different denominators.[8][6]
4. **Show exact prompt/query pairs.** Concrete rewrites reveal added years, vendors and comparison language more clearly than aggregate percentages.[8]
5. **Date every pattern.** Search behavior can change between observation windows, as shown by differences in `site:` usage and query length.[2] Cited sources shift as well: in [Finseo's own German-language (DACH) citation-stability data](https://www.finseo.ai/blog/ai-citation-stability), only 38.4% of URLs ChatGPT cited were still cited the next day for the same prompt.
6. **Distinguish missing data from no search.** API and interface capture methods do not always expose the same arguments.[2][4]

The evidence supports a concise conclusion: **ChatGPT is documented to rewrite questions into targeted searches and may refine those searches after seeing initial results. Observed fan-outs commonly add dates, entities, evaluation terms, language changes and domain restrictions, but no fixed product-wide query count or universal rewrite formula has been established.**[2]


What should an AI visibility team record when studying ChatGPT fan-out?

Record the original prompt, timestamp, locale, account or session conditions, visible search queries, retrieved URLs, displayed citations, response text and capture method. Keep query counts separate from URL counts and citations because these represent different stages of the search process.[11][8]

Also preserve the exact observation window and model or interface details when available. Without that context, a percentage such as 43% English fan-outs or 1.6 queries per fan-out response can easily be misread as a universal ChatGPT property.[12][3]



## FAQ
### Does ChatGPT rewrite every question into search queries?
No published evidence establishes that every question becomes a search query. OpenAI says ChatGPT can search automatically when a question would benefit from current information, and users can also select Search manually.[1] The available studies measure observed fan-outs in particular panels rather than every ChatGPT interaction.[11][10]

### What is a ChatGPT fan-out query?
A ChatGPT fan-out query is an underlying search query generated while ChatGPT researches a user’s request. OpenAI describes ChatGPT as converting a request into one or more targeted queries, and external studies observe those queries in specific interfaces or datasets.[17]

### How many searches does ChatGPT run for one prompt?
There is no authoritative product-wide number. Published observations include approximately 1.6 queries among Ahrefs responses containing fan-out, a median of two in StatWharf’s software panel and an average of 1.65 searches per answer in Howseen’s predominantly French sample.[8]

### Why does ChatGPT search in English when I ask in another language?
Observed audits show that ChatGPT sometimes generates English searches for non-English prompts, but the rate varies substantially by sample and topic. Reported findings range from 17.1% of searches in Howseen’s panel to 43% of non-English fan-outs in Peec’s analysis, while aclipp observed an 84% average in one German PR-software project.[12]

### Does ChatGPT add the current year to search queries?
It often does in some observed datasets, but year insertion is not a documented universal rule. StatWharf found a year in 76% of its 211 generated queries and reported that 62% of queries derived from no-year baseline prompts added one.[7]

### Can I see every query ChatGPT Search runs?
Not necessarily. OpenAI’s API documentation says searched queries are usually, but not always, exposed, and distinguishes query data from consulted sources and displayed citations.[16] Interface and capture method also matter; StatWharf found differences between live streams and saved conversation records.[7]

## Sources
[1] Searching the web with ChatGPT | OpenAI Help Center — https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt
[2] Characterizing Web Search by Conversational LLM Agents:From Search Decisions and Strategies to Results and Responses — https://arxiv.org/html/2609.19244
[3] How ChatGPT Does Query Fan-Out (15 Million Queries Studied) — https://ahrefs.com/blog/query-fan-out-chatgpt/
[4] What We Can Learn from Evolving ChatGPT Fan-Out Queries — https://lilyraynyc.substack.com/p/what-we-can-learn-from-evolving-chatgpt
[5] ChatGPT Hidden Searches 2026: 1.48 Web Searches Per Question — https://www.suffdigital.com/resources/data-studies/chatgpt-hidden-searches-per-question
[6] From Prompt to Recommendation:A Fitted Stage Model of Brand Visibility in AI Search — https://arxiv.org/html/2609.23162
[7] How ChatGPT searches for software recommendations: a query fan-out study — https://statwharf.com/research/chatgpt-query-fan-out
[8] ChatGPT Fan-Out Queries: 1 in 6 Run in English | Howseen — https://howseen.ai/data/chatgpt-fan-out-queries-english
[9] ChatGPT fan-outs have doubled in length in 4 months - Peec AI — https://peec.ai/blog/country-analysis-20-million-search-qfos
[10] Query Fan-Out in ChatGPT and Gemini Across 6,352 Searches | Bee LLM — https://beellm.ai/blog/query-fan-out-chatgpt-gemini-study
[11] Query Rewriting for Retrieval-Augmented Large Language Models — https://aclanthology.org/2023.emnlp-main.322.pdf
[12] ChatGPT searches in English, even when you don't - Peec AI — https://peec.ai/blog/chatgpt-searches-in-english-even-when-you-don-t
[13] Asked in German, searched in English: three AI visibility findings from our own measurement — https://aclipp.com/en/blog/ai-assistants-search-in-english
[14] How Retrieval, Fan-out & SERPs Shape ChatGPT Citations — https://www.airops.com/report/influence-of-retrieval-fanout-and-google-serps-in-chatgpt
[15] Patterns we see in ChatGPT query fanouts - Peec AI — https://peec.ai/blog/patterns-we-see-in-chatgpt-query-fanouts
[16] Web search — https://developers.openai.com/api/docs/guides/tools-web-search
[17] ChatGPT Now Uses the site:operator at Scale | Promptwatch — https://promptwatch.com/data/chatgpt-site-operator-fanouts
