Key numbers
of all citations went to social platforms, depending on the industry. The lowest was commercial law and the highest recruitment.
of answers cited Reddit at least once. Counted as a share of all citations, Reddit was 0.2% to 1.6%.
recruitment citations went to LinkedIn, the most for any social platform in the study. They came from just 12 of the 750 answers, while Reddit appeared in 33.
of citations in accountancy, fitness and recruitment went to companies' own websites, the largest single source type in each.
of commercial law citations went to two directories: Chambers and Partners (56.8%) and Legal 500 (18.9%).
of accountancy citations went to government sources, almost all of them gov.uk pages.
of answers in law, fitness and recruitment used web search even when the model could choose not to. In accountancy it was 73%.
In the forced pass the model had to search the web before answering, and in the unforced pass it decided for itself. "Both passes" combines the two, and the definitions are under sample and method.
Charts
Social platforms
Social platforms in ChatGPT's citations, by industry (both passes)
| Industry | Answers | Citations | Answers citing social | Social share of citations |
|---|---|---|---|---|
| Accountancy | 500 | 2,169 | 5.4% | 1.6% |
| Commercial law | 636 | 3,317 | 1.6% | 0.3% |
| Fitness | 750 | 3,439 | 3.2% | 0.9% |
| Recruitment | 750 | 4,291 | 6.0% | 2.6% |
An answer counts once in the "answers citing social" column, however many social links it contains.
Social citations by platform (both passes, number of citations)
| Industry | YouTube | All social | ||
|---|---|---|---|---|
| Accountancy | 34 | 0 | 0 | 34 |
| Commercial law | 8 | 2 | 0 | 10 |
| Fitness | 28 | 0 | 3 | 31 |
| Recruitment | 43 | 70 | 0 | 113 |
Reddit, LinkedIn and YouTube were the only social platforms cited. There were no citations to Facebook, Instagram, X, TikTok, Pinterest, Threads, Quora, Medium or Mumsnet. Review sites such as Trustpilot and Glassdoor are counted separately as review platforms.
Reddit, counted two ways (both passes)
| Industry | Answers citing Reddit | Reddit share of citations |
|---|---|---|
| Accountancy | 5.4% | 1.6% |
| Recruitment | 4.4% | 1.0% |
| Fitness | 2.9% | 0.8% |
| Commercial law | 1.3% | 0.2% |
Widely quoted Reddit figures usually count the first way, whether Reddit appears in an answer at all. One Reddit link among ten sources is a full hit on that measure and a tenth of a hit on the second. Accountancy and law answers are four-fifths forced-pass, against half for fitness and recruitment, so compare industries with care.
Source types
Share of citations by source type (forced pass)
| Source type | Accountancy | Commercial law | Fitness | Recruitment |
|---|---|---|---|---|
| Brand's own site | 60.2% | 22.1% | 56.1% | 74.1% |
| Directories and rankings | 12.1% | 76.3% | 0.3% | 6.8% |
| Publishers and editorial | 7.1% | 0.7% | 30.3% | 4.3% |
| Government and official | 18.5% | 0.6% | 6.6% | 7.2% |
| Marketplaces and retailers | 0% | 0% | 5.3% | 2.2% |
| Review platforms | 0.4% | 0% | 0% | 1.4% |
| Social and UGC | 1.7% | 0.3% | 1.3% | 3.8% |
Based on the 8,461 forced-pass citations. Each cited website was assigned one source type. Any named company selling a product or service counts as a brand's own site, so a software company such as Xero sits in the same group as a five-person practice. Columns may not sum to 100% because of rounding.
Individual sources
Sources named in the study write-up (both passes)
| Industry | Source | Website | Share of citations |
|---|---|---|---|
| Commercial law | Chambers and Partners | chambers.com | 56.8% |
| Commercial law | Legal 500 | legal500.com | 18.9% |
| Accountancy | Crunch | crunch.co.uk | 11.3% |
| Accountancy | ICAEW | icaew.com | 9.4% |
| Fitness | Runner's World | runnersworld.com | 10.5% |
| Fitness | Tom's Guide | tomsguide.com | 5.8% |
| Fitness | RunRepeat | runrepeat.com | 3.3% |
| Recruitment | Hays | hays.co.uk | 6.7% |
| Recruitment | Michael Page | michaelpage.co.uk | 3.7% |
| Recruitment | Robert Walters | robertwalters.co.uk | 3.2% |
These are the sources discussed in the write-up. They are a selection rather than a full ranking: in accountancy, for example, gov.uk was cited more often than Crunch, at 14.8% of citations. Shares are of all citations in that industry across both passes, matched on the website address.
Sample and method
Each industry had a set of buying questions written the way a customer would ask them, such as "Who are the best accountancy firms for small businesses in the UK?". Every question was sent to GPT-5.4 through the OpenAI Responses API several times in two passes. In the forced pass the model had to use its web search tool, and in the unforced pass it decided for itself whether to search. Every source URL listed in an answer was logged as a citation.
The accountancy and commercial law answers come from the datasets behind our earlier studies of UK accountants and UK law firms, which asked each question 20 times with search forced and 5 times without. The original accountancy study included five questions about small-business law firms. They are excluded here, which leaves 20 accountancy questions. Fitness and recruitment were run for this study, 15 times in each pass.
Questions, answers and citations by industry
| Industry | Questions | Runs per question, forced / unforced | Answers, forced / unforced | Citations | Fieldwork, 2026 |
|---|---|---|---|---|---|
| Accountancy | 20 | 20 / 5 | 400 / 100 | 2,169 | 10 to 11 Aug |
| Commercial law | 25 | 20 / 5 | 511 / 125 | 3,317 | 19 Aug |
| Fitness | 25 | 15 / 15 | 375 / 375 | 3,439 | 27 Aug |
| Recruitment | 25 | 15 / 15 | 375 / 375 | 4,291 | 27 to 28 Aug |
| All four | 95 | 1,661 / 975 | 13,216 | 10 to 28 Aug |
One commercial law question has 31 forced-pass answers rather than 20, because 11 of its runs were made twice. Each gave a different answer, so all 31 are included. Fieldwork dates are when the answers were logged.
- Answer
- One response from the model to one question.
- Citation
- One source URL listed in an answer. No answer listed the exact same URL twice. Four listed the same page twice, once with a tracking tag added to the address, and both copies are counted.
- Share of citations
- Citations to a source or source type, divided by all citations in that industry and pass.
- Answers citing
- The share of answers with at least one citation to that source.
- Social platforms
- Reddit, LinkedIn and YouTube, the only social platforms that appeared.
- Individual sources
- Matched on the website address with any leading "www." removed, so a subdomain such as find.icaew.com is counted separately from icaew.com.
Limitations
This covers one model, GPT-5.4, queried through the API in August 2026. The ChatGPT app, other engines such as Gemini, Claude and Perplexity, and later months may cite differently.
The questions are buying questions in four categories, so the figures describe what the model cites when someone is choosing a supplier or a product, rather than ChatGPT use in general.
Assigning websites to source types involves judgement. The social, Reddit and individual source figures don't depend on it, because they are matched on the website address.
The data shows which sources were cited. It can't show why the model chose them.
Downloads
- chatgpt-citation-sources-by-industry.csvAnswers, citations and social platform figures for each industry, both passes.
- chatgpt-citation-sources-source-types.csvShare of citations by source type for each industry, forced pass.
- chatgpt-citation-sources-named-sources.csvThe individual sources named in the write-up, with citation counts.
- chatgpt-citation-sources-sample.csvQuestions, runs, answers, citations, search rate and fieldwork dates for each industry and pass.
- chatgpt-citation-sources-reddit-chart.pngChart: Reddit counted two ways.
- chatgpt-citation-sources-source-types-chart.pngChart: share of citations by source type.
How to cite
The charts and tables are free to reuse with a link to this page. Please credit Visibly.
Press and questions
For a breakdown that isn't shown here, or a question about the method, email Jack, founder of Visibly, at [email protected]. Visibly is a UK company building tools that measure how AI answer engines such as ChatGPT recommend brands.
Changes
7 October 2026: data page published, and corrections to the write-up
Rebuilding every table from the raw answer logs turned up some errors in the original write-up, which we have corrected there. None of the headline findings changed.
- Accountancy and law questions were asked 20 times with search forced and 5 times without, not 15 times in each pass. Fitness and recruitment were 15 times in each pass, as stated.
- With search optional, the model searched in 73% of accountancy answers, not 74%. The earlier figure included the five excluded law questions.
- LinkedIn appeared in 1.6% of recruitment answers and YouTube in 0.4% of fitness answers across both passes. The 2.4% and 0.8% given before were forced-pass figures.
- The source-type chart is based on the 8,461 forced-pass citations, not all 13,216.
- Chambers and Partners and Legal 500 together took 75.7% of law citations. A chart label gave 76.3%, which is the share for all directories in the forced pass.
- Hays was cited 1.8 times as often as Michael Page, the next agency, not three times.
- We removed a sentence saying the unforced pass was within two points of the forced pass for every source type. In law, Chambers and Partners and Legal 500 took 73.5% of forced-pass citations and 85.1% of unforced ones.
- We removed a sentence saying a test of the source-type rules moved the headline figures by under two points. The test's results aren't in the study files, so we can't show it.