# How does ChatGPT search for local news?

What happens between asking ChatGPT about local news in your town and getting an answer with sources.

By Connor Lahey (Content Lead) and JP Garbaccio (Head of AEO/SEO). Updated 21 September 2026.

Searchable asked five AI assistants 30,000 questions about local news in 100 UK towns and cities. For ChatGPT, we also recorded the web searches it ran before answering. This page follows those searches from the question to the answer.

## In three steps

1. ChatGPT decides to search the web, then rewrites your question into a search of its own.
2. It turns that one question into anything from 1 to 35 searches, and most answers run between 1 and 7. The first searches are broad, and later ones search inside specific websites, such as the council's.
3. Most answers get back pages from 5 to 21 websites, and ChatGPT cites 2 to 6 of them in its reply, drawn from 1 to 5 kinds of website, such as council sites, newspapers and Wikipedia.

## What a search means here

When you ask ChatGPT a question, it can type its own queries into a web search tool before it replies, much as you might try a few phrasings in Google. On this page, a search is one of those queries. You ask once, and ChatGPT decides how many searches to run.

We asked ChatGPT for today's top news stories in Grimsby three times. It ran two searches the first time, three the second and four the third:

- Ask 1 (2 searches): `Grimsby Lincolnshire news today September 9 2026`, `Grimsby news September 9 2026 Lincolnshire`
- Ask 2 (3 searches): `Grimsby Lincolnshire news today September 9 2026`, `Grimsby news September 9 2026 Lincolnshire Live Grimsby Telegraph`, `Grimsby Lincolnshire latest news today`
- Ask 3 (4 searches): `Grimsby Lincolnshire news today September 9 2026`, `Grimsby news September 9 2026 Lincolnshire Live Grimsby Telegraph`, `site:grimsbytelegraph.co.uk "9 SEP 2026" Grimsby`, `site:grimsbytelegraph.co.uk "09 SEP 2026" Grimsby news`

Every count of searches on this page means these queries, the ones ChatGPT ran for a single answer.

---

## Step 1: ChatGPT rewrites your question before it searches

The question we asked was "What are the top news stories in Grimsby, Lincolnshire, UK today?" ChatGPT never searched for that sentence. Its first search, the same in all three asks, read `Grimsby Lincolnshire news today September 9 2026`.

Set side by side, the changes are easy to see. "What are the top" and "stories" went, and so did "UK". Grimsby, Lincolnshire, "news" and "today" stayed. The date was added, because "today" on its own would not tell a search engine which day. In its later searches ChatGPT added the names of two local titles it already knew, Lincolnshire Live and the Grimsby Telegraph, which nobody had mentioned.

Before any of this, ChatGPT decides whether to search at all, and in our study it searched the web for 5,863 of 6,000 answers. Across the 9,700 searches we recorded, none repeated the question word for word. The searches were shorter, 7.8 words against 11 in the question, and 296 of the 300 questions asking about "today" or "right now" were searched with the exact date.

Kept from the question (share of searches whose question contained it):

| Part of the question | Share of searches | Searches |
|---|---|---|
| Town or city name | 98.4% | 9,544 of 9,700 |
| The word "news" | 83.1% | 3,423 of 4,119 |
| County | 55.1% | 3,683 of 6,687 |
| The word "local" | 38.1% | 1,209 of 3,176 |
| "Today", "this week" and similar | 28.5% | 1,507 of 5,283 |
| "UK" | 1.6% | 153 of 9,700 |

Added by ChatGPT (share of all 9,700 searches; no question contained these):

| Added to the search | Share of searches | Searches |
|---|---|---|
| The year (2026) | 94.5% | 9,166 of 9,700 |
| The month (September) | 75.8% | 7,356 of 9,700 |
| A site: search | 32.5% | 3,155 of 9,700 |
| A phrase in quotation marks | 27.1% | 2,630 of 9,700 |

---

## Step 2: One local news question becomes several searches that narrow to specific websites

We asked ChatGPT "What decisions has the local council made in Doncaster, South Yorkshire, UK recently?" The three asks produced two, four and eight searches. ChatGPT sends its searches in rounds: a batch at once, then another batch after the first results come back. The four-search answer ran two rounds of two:

1. Search 1, round 1: `Doncaster Council recent decisions September 2026 cabinet council South Yorkshire`
2. Search 2, round 1: `site:doncaster.gov.uk council decisions Doncaster cabinet September 2026`
3. Search 3, round 2: `site:doncaster.gov.uk "2026" "Cabinet" "resolved" Doncaster June July August 2026 council`
4. Search 4, round 2: `site:doncaster.gov.uk/news 2026 Doncaster Council decision cabinet August 2026`

The first search is broad. The rest use the site: operator, which limits a search to one website, here the council's own. In the second round ChatGPT added exact phrases in quotation marks, "Cabinet" and "resolved", the words a council uses when it minutes a decision.

Google calls this query fan-out. It [introduced the term with AI Mode in March 2025](https://blog.google/products/search/ai-mode-search/) to describe one question being split into several related searches whose results are combined into a single answer.

Across the study, ChatGPT ran anything from 1 to 35 searches for an answer, 3.2 on average. It almost always opened broad: 2,993 of 3,002 answers began without naming a website. Then it narrowed. In the first round 11.8% of searches used site:, and in the second round 70.2%.

The number of searches depends on what you ask. Planning questions averaged 9.5 searches per answer and council decisions 5.6. "Where can I read local news" averaged 1.3, and none of those answers searched inside a website. The most searches for one answer came from a planning question about Sheffield: 35 searches over eight rounds, down to a single road, Orgreave Lane, and individual application numbers such as `26/02534/FUL`.

Average number of searches ChatGPT ran to write one answer, by question (about 150 answers per question):

| Question | Searches per answer | Answers searching inside a named website |
|---|---|---|
| Planning applications | 9.5 | 93% |
| Council decisions | 5.6 | 88% |
| Local news to know about | 3.9 | 40% |
| Catch me up on the news | 3.7 | 29% |
| This week's headlines | 3.7 | 28% |
| Top news stories today | 3.6 | 33% |
| Events this weekend | 3.5 | 39% |
| What's in the local paper | 3.2 | 65% |
| Summarise the local news | 3.2 | 26% |
| Housing and new homes | 3.2 | 35% |
| What people are talking about | 3.0 | 17% |
| Police incidents | 2.7 | 52% |
| What's happening this week | 2.6 | 21% |
| Local sport news | 2.6 | 19% |
| Road closures and roadworks | 2.4 | 53% |
| Biggest news story right now | 2.3 | 7% |
| Breaking news | 1.8 | 1% |
| Who reports local news | 1.5 | 0% |
| Crime reported recently | 1.4 | 7% |
| Where to read local news | 1.3 | 0% |

---

## Step 3: ChatGPT's searches bring back a pool of pages, and it cites a few of them in the answer

Back in Doncaster, the four searches returned 25 pages from 8 websites. ChatGPT's answer cited 8 of them. Six citations went to doncaster.gov.uk, the website it had searched inside, one to the council's meeting records and one to opencouncil.network, which republishes council data.

The Guardian, The Times and six Wikipedia pages came back in the same searches, and none of them was cited. Returned pages are the ones ChatGPT had in front of it. Cited pages are the ones it linked in the text of its reply.

| Website | Pages returned | Pages cited |
|---|---|---|
| doncaster.gov.uk | 12 | 6 |
| en.wikipedia.org | 6 | 0 |
| youtube.com | 2 | 0 |
| doncaster.moderngov.co.uk | 1 | 1 |
| opencouncil.network | 1 | 1 |
| theguardian.com | 1 | 0 |
| thetimes.com | 1 | 0 |
| facebook.com | 1 | 0 |

The common range for a ChatGPT answer about local news, from searches run to websites cited (middle 80% of 3,002 answers; 13 cited nothing):

| Per answer | Lowest common | Highest common | Average |
|---|---|---|---|
| Searches run | 1 | 7 | 3.2 |
| Pages returned | 10 | 44 | 24.6 |
| Websites returned | 5 | 21 | 12.3 |
| Pages cited | 3 | 9 | 5.8 |
| Websites cited | 2 | 6 | 4.0 |
| Kinds of website cited | 1 | 5 | 2.8 |

One page in four that the searches return ends up cited. Searching inside a website makes a difference: in the 981 answers where ChatGPT used site:, it cited at least one of the websites it had searched inside 76% of the time.

A site: search can only return what ChatGPT's search is allowed to reach. [OpenAI says](https://developers.openai.com/api/docs/bots) sites that block its OAI-SearchBot crawler "will not be shown in ChatGPT search answers", however precisely ChatGPT searches for them.

---

## Methodology

- Searchable asked 20 questions about local news and local services for each of 100 UK towns and cities, three times each, to five AI assistants: ChatGPT, Perplexity, Google Gemini, Google AI Mode and Google AI Overviews. That made 30,000 prompts, run on 9 and 10 September 2026 with a UK location set and no signed-in account. The answers carried 244,725 citations from just over 10,000 websites.
- For ChatGPT, the collection tool also returned the searches ChatGPT ran and the pages they brought back for 3,002 of its 6,000 answers. Every search figure on this page is based on those 3,002 answers and the 9,700 searches behind them.
- "Most answers" means the middle 80% of answers. Kinds of website come from Searchable's classification of every cited page into types such as council, local news, national news, police and reference.

---

[Start for free](https://app.searchable.com/signup) | [Book a demo](/contact-sales)