We ran 2,636 user prompts through ChatGPT across four UK industries and logged every source it cited. Social platforms accounted for between 0.3% and 2.6% of the 13,216 citations we collected. The advice to build your AI visibility on Reddit and LinkedIn is pointed at the smallest part of the answer. The largest part, in three of the four industries we tested, was the brands' own websites.
The headlines
- Social platforms are between 0.3% and 2.6% of citations. Not zero, but nowhere near what the industry advice implies.
- Measured the way the widely quoted studies measure it, Reddit appeared in 1.3% to 5.4% of answers, against the 40% to 60% those studies report.
- Recruitment, the industry closest to LinkedIn, was the highest of the four. LinkedIn still got 70 citations out of 4,291.
- Brands' own websites were 56% to 74% of citations in accountancy, fitness and recruitment.
- Commercial law was a significant exception. Two directories, Chambers and Legal 500, accounted for 75.7% of every citation.
- Which of those two shapes your industry takes decides where your effort should go, and stops you wasting time on things that will have little impact.
How we tested where ChatGPT's citations come from
Twenty-five buying questions per industry, the kind a real customer types: best recruitment agencies for nursing jobs in the UK, what are the best running shoes for beginners, who are the best accountancy firms for small businesses. Each question was asked fifteen times, because asking once doesn't give you a stable response to analyse.
Every question ran twice. Once with web search forced on, to see the full set of sources available to the model. Once with search left to the model's own judgement, which tells you whether writing your own content is even worth it. All on GPT-5.4, across UK accountancy, UK commercial law, UK recruitment and consumer fitness.
The accountancy and law numbers come from the datasets behind our earlier studies of UK accountants and UK law firms, re-analysed for source type rather than named firms. One change worth flagging: the original accountancy study ran 25 prompts, five of which asked about small-business law firms rather than accountants. For this analysis we separated those out, so the accountancy figures below draw on 400 answers rather than the 500 in the original study. Leaving them in would have counted legal directories as accountancy citations.
The two passes matter because of the obvious objection. Force a model to search and you get search results. Left free to choose, the model still searched in 98% to 100% of cases in law, recruitment and fitness, and 74% of the time in accountancy. Social citations went down in the unforced pass, not up.
Does ChatGPT cite Reddit? Two ways to count it
There are two ways to count this, and the difference explains most of the argument about it.
Answer frequency asks whether a source appeared in an answer at all. One Reddit link in an answer with nine other sources counts as a hit. This is the measure behind the headline figures that have been quoted over the last twelve months or so.
Citation share asks what proportion of every source cited was Reddit. The same answer counts as one tenth of a hit.
We ran both, so the comparison is like for like.
| Industry | Answers | Answers citing any social | Citations | Social share of citations |
|---|---|---|---|---|
| Accountancy | 500 | 5.4% | 2,169 | 1.6% |
| Commercial law | 636 | 1.6% | 3,317 | 0.3% |
| Fitness | 750 | 3.2% | 3,439 | 0.9% |
| Recruitment | 750 | 6.0% | 4,291 | 2.6% |
Semrush analysed over 150,000 AI citations in June 2025 and found 40.1% of references pointed at Reddit. In early August 2025 they observed ChatGPT citing Reddit in close to 60% of responses. Measured the same way, in four UK commercial categories, we found between 1.3% and 5.4%.
That gap is not a methodological artefact. We can count it their way and still find roughly a tenfold difference. What separates the two is not how you count but what you count over. Horizontal studies average across every topic people ask about, and a great deal of what people ask ChatGPT is not a buying question. When someone is choosing a supplier, the sources look different.
The individual platform numbers make the same point. LinkedIn appeared in 2.4% of recruitment answers and none at all in accountancy or fitness. YouTube appeared in 0.8% of fitness answers and nowhere else. These are the two platforms most often named in AI visibility advice.
Recruitment is the result worth sitting with. It is the industry closest to LinkedIn, and candidates discuss agencies on Reddit constantly. If social were going to drive AI recommendations anywhere it would be here, and LinkedIn still managed 70 citations out of 4,291.
Worth flagging a discrepancy we cannot fully explain: Profound records LinkedIn cited in 14.3% of ChatGPT Search responses, and rising from eleventh to fifth most-cited domain in three months. That is far above anything we saw. It may be that ChatGPT Search behaves differently from the API, or that LinkedIn dominates in categories we did not test. We are reporting what our data shows rather than reconciling the two.
Why everyone thinks Reddit drives AI visibility
Reddit's position has been genuinely unstable, which is part of how the story took hold.
Semrush watched ChatGPT's Reddit citations fall from roughly 60% of responses to around 10% between early August and mid September 2025, then recover. Between 14 and 17 August 2026, Peec AI recorded Reddit's share of ChatGPT Search citations dropping from 3.83% to 0.52%, an 86% fall in four days.
Anyone who built a strategy at the peak has watched the ground move underneath it twice. And because the widely quoted figures are horizontal averages, a brand in a specific category could follow that advice diligently and never appear in a single answer their customers actually see.
This also updates something we published ourselves. Our guide to getting cited by ChatGPT recommends building community mentions, and we still think that holds for Perplexity, which leans on community sources far more heavily. What our own data now shows is that on ChatGPT, in these four categories, the payoff is small. We have updated that guide accordingly.
One note on definitions. Trustpilot and Glassdoor sit in a separate review category rather than folded into social. Counting them as social would take recruitment to 5.3% of citations and leave the other three industries roughly where they are, so the conclusion holds either way.
Where ChatGPT's brand citations actually come from
Forced pass shown, which gives the fuller view of available sources. The unforced pass sits within two points on every line.
| Source type | Accountancy | 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, all platforms | 1.7% | 0.3% | 1.3% | 3.8% |
Law is a directory industry. Chambers alone took 56.8% of every citation in the commercial law study. Legal 500 added 18.9%. Between two websites, three quarters of everything ChatGPT cited when asked to recommend a commercial solicitor. Individual firms' own sites managed 22.1% between all of them, and the best-performing single firm reached 0.99%. This is the same pattern that produced the 6% overlap between Google rankings and AI recommendations we found in the original law study, seen from the source side.
The other three are own-site industries. In recruitment, Hays alone took 6.7% of all citations, ahead of every directory, publisher and social platform in the study. Michael Page took 3.7%, Robert Walters 3.2%. In accountancy, Crunch took 11.3%, ahead of ICAEW at 9.4%.
Fitness sits between the two and is the one place editorial genuinely earns its keep. Runner's World alone accounted for 10.5% of every citation in that study, Tom's Guide 5.8%, RunRepeat 3.3%. Manufacturers' own sites still came out ahead in aggregate at 56.1%, but if you sell running shoes, a Runner's World roundup is worth more than any amount of social activity.
One finding we did not expect: government sources took 18.5% of accountancy citations, almost all gov.uk. Ask ChatGPT about VAT registration or Making Tax Digital and it goes to HMRC before it goes to an accountant. If your category touches regulation, part of your competition for the answer is the government.
Where to focus, by industry
Work out which shape your industry is first. The answers are different and the wrong one costs months. We built Visibly specifically to help brands with this problem.
If your industry has directories like Chambers or Legal 500
Chambers, Legal 500, ICAEW, the REC. In law these are 76% of the answer, so any approach to influence these would be a smart strategy.
- Get ranked and treat the submission as a priority rather than an admin task.
- Read your directory entry as if you were the model. The description there is what gets quoted, so it should say what you do and who for, not what band you achieved.
- Keep it current. A stale entry is what gets recommended for the next twelve months, and you can track this over time with a product like Visibly.
- Check which directories actually get cited in your category before investing. In recruitment it was AgencyCentral and the REC, not the ones with the biggest ad spend.
- Your own site still matters, but it is the second job, not the first.
If it doesn't, your website is the battleground
This is accountancy, recruitment, fitness and most other sectors. Between 56% and 74% of what the model cites is companies' own sites, which means you need to work harder on your website than ever before.
- Confirm the crawlers can see you. Pages rendered entirely in JavaScript are invisible to GPTBot and ClaudeBot, which do not execute it. Our free AI crawler checker tells you in under a minute. Rule this out before anything else.
- Write pages that answer the specific question a customer asks. “We recruit data engineers for early-stage companies in Manchester” gets cited for that question. “We are a people-first talent partner” does not.
- Make each page self-contained. A paragraph that only makes sense after three paragraphs of preamble will not survive being extracted as a citation.
- Put the facts a model needs in crawlable text: what you do, who for, where, what it costs, what makes you the right answer. Not in an image, not behind a script.
- Think about which content formats end up in AI answers. The prompts in this study were full sentences, and the pages that got cited answered them directly. That is the core of what GEO actually asks of you, and there is more detail in our guide to how to get cited by AI search.
- Check whether an official source owns part of your category. If gov.uk answers the question before an accountant does, your page needs to add something HMRC doesn't.
Whatever shape you are in
- Monitor it rather than doing it once. Reddit's collapse happened over four days.
- Track which sources get cited in your category specifically, not the horizontal averages. Those averages are the reason the Reddit advice spread. Measuring your own share of AI answers compared to competitors beats following a benchmark drawn from every topic at once.
- Watch competitors' citation share, not just your own. Hays being cited three times more than the next agency is a position someone has to take off them.
- If you are doing this across several engines and competitors, a spreadsheet stops working quickly. That is the point at which an AI visibility tool starts paying for itself.
None of this shows up in your analytics. The customer who asked ChatGPT and got three other names never visited your site, so there is no lost click to find. That is the same blind spot behind zero-click search, and it is why the only way to see it is to ask the engines directly.
What this study can't tell you
This is ChatGPT only, sampled across August 2026. Gemini, Claude and Perplexity cite differently and we have not measured them at the same depth. Source classification involved judgement, chiefly treating any named company selling a product or service as brand-owned, which puts Xero and a five-person practice in the same bucket. We tested the effect and it moved the headline figures by under two points. And we can show what is cited without proving what causes citation.
Find out where your category's citations come from
Join the waitlist and I'll send what we're finding out about which sources AI engines read when they recommend a brand, and what you can do about it. You'll be first in when Visibly goes live.
Join the waitlistFrequently asked questions
Where do brand citations in ChatGPT come from?
Across 2,636 UK buying questions we logged 13,216 citations. In accountancy, recruitment and fitness, companies' own websites were the largest source at 56% to 74%. In commercial law, two directories accounted for 75.7%. Social platforms were between 0.3% and 2.6%.
Where does ChatGPT get its information from?
For commercial questions it searches the live web rather than answering from training data. In our tests it searched in 98% to 100% of cases in law, recruitment and fitness, and 74% in accountancy, even when it was not instructed to.
Does ChatGPT cite Reddit?
Less than the industry assumes for commercial buying queries. Reddit appeared in 1.3% to 5.4% of answers across the four UK industries we tested, and made up between 0.2% and 1.6% of all citations. Its share has also moved sharply, falling 86% on ChatGPT Search in four days in August 2026. Perplexity leans on community sources much more heavily, so the answer depends on which engine your customers use.
Which sources does ChatGPT cite most?
It depends on the industry. In UK commercial law, two directories accounted for 75.7% of citations. In accountancy, recruitment and fitness, companies' own websites were the largest source at 56% to 74%.
Does LinkedIn matter for AI search?
It is rising. Profound recorded LinkedIn moving from eleventh to fifth most-cited domain on ChatGPT in three months. In our recruitment study, the industry where it should matter most, it appeared in 2.4% of answers and accounted for 70 of 4,291 citations.
Is AI search citation stable?
No. Reddit's share on ChatGPT fell from roughly 60% of responses to 10% between August and September 2025, recovered, then dropped 86% over four days in August 2026.
About the author. Jack is the founder of Visibly and has spent over a decade in SEO and digital strategy, including a senior digital director role at a FTSE 100 company and agency work for national brands. He built Visibly to measure and improve how AI answer engines recommend brands. More about Visibly.