We checked 198 facts that ChatGPT, Perplexity and Google gave UK university applicants. The source decided which ones were right.

In October 2026 we checked 198 facts that ChatGPT, Perplexity and Google AI Overviews gave about 31 UK universities. 63% were correct, and accuracy depended mainly on which page the answer came from.

Last checked: 2 November 2026. Data: download the dataset (CSV, CC BY 4.0, DOI 10.5281/zenodo.23594680).

What we asked, and why we stopped asking "which university should I pick?"

We wrote 21 questions of the kind applicants and parents actually ask. Entry grades for named courses. Fees. Graduate jobs. Contextual offers. Whether one university is "better" than another. Each went to ChatGPT, Perplexity and Google, once, between 20 and 21 October 2026. That's 63 answers.

Then we pulled every specific, checkable fact out of each answer (a grade, a fee, a rank, a percentage) and checked it against the university's own page or official data. 198 facts in all. A fact only counted as correct if we could match it to a word-for-word quote from an official source.

We had planned to ask open questions too. "Which UK universities should my daughter apply to for medicine?" is about as real an applicant question as it gets.

It didn't work, and the reason became a finding of its own. We ran our tests from Scotland. Asked that medicine question, ChatGPT put five Scottish medical schools at the top of its list. We asked again, signed out. Same thing. When we told it to ignore our location, it gave a UK-wide list. Perplexity, asked the same words from the same place, led with Oxford, Cambridge and Imperial.

Signed in and signed out, ChatGPT also answered in Scottish school years (S4 to S6). Asked to ignore location, it listed A-levels first. The facts barely changed between the three versions. The shortlist did.

So an open shortlist tells you as much about where the person is sitting as about the universities. That makes it useless for comparing platforms, and a student in Belfast or Bristol would see something different from what we saw.

We switched to naming the universities in every question. "How hard is it to get into computer science at Imperial, UCL, Manchester and Edinburgh?" gets every asker an answer about the same four places, and every fact in it can be checked. The early open-question runs are kept in a separate pilot file and left out of every figure on this page.

The platform scores

ChatGPT comes out on top. We wouldn't read much into the order, though.

Each platform was best on some questions and poor on others. Perplexity got every UCL fact right, straight from UCL's pages, then made the worst fee error in the study. Google was the only platform that answered "which is least competitive for psychology?" with actual grades, and correctly picked Strathclyde. It was also the platform most likely to build an answer out of Reddit and Quora. ChatGPT ranked four universities by engineering graduate salary in exactly the order the official Discover Uni data gives, and also gave Durham Psychology a grade higher than Durham asks for.

In a UCAS survey of 4,485 prospective applicants, applicants and first-year students (November 2025), 73% of those using AI said they had recognised or received incorrect information from it. That question covered everything they used AI for, not just university research. Our study measures the university part directly. The more useful question is where the wrong answers come from.

The page the answer came from decided whether it was right

For each university in each answer, we recorded which page the platform drew on for each topic (grades, fees, rankings and so on). That gave 122 traceable answer-topics.

The gap between the top and bottom rows is far wider than the gap between any two platforms. When a platform found the right page, it mostly read it correctly. Dundee's rankings page is a good example: ChatGPT and Google both quoted Dundee's 2027 positions from it without a slip (Times 20th, Guardian 21st=, Complete University Guide 25th).

Third-party sites were a coin toss. Sometimes they won on merit. Asked about international fees at Manchester, Google took its fee bands from Acolyte Living, a student-housing company, because Acolyte puts every band on one page and Manchester lists fees course by course. The top laboratory fee Google gave was still lower than the figure on Manchester's own Computer Science page.

Looking official isn't the same as being official, either. Perplexity quoted graduate-outcome percentages for Exeter, Newcastle and Cardiff from universitydata.co.uk, a tidy site built on HESA data. We downloaded HESA's own 2023/24 provider table and couldn't reproduce those figures on any basis we tried, so we've marked them unverifiable rather than wrong.

Then there's the bottom row of that table, which surprised us most.

We traced all 15 of the answers in that bottom row back to the page itself. In ten, the platform hadn't made anything up. It had found a real page on the university's site, a 2022 news story or last year's league tables, and repeated it faithfully. One came out right anyway. The rest went wrong at the platform's end, or couldn't be confirmed at all: a joint-honours page read as single honours, an old prospectus site, an archive page, and Edinburgh fee bands we couldn't find on any current Edinburgh page. The ten are the ones that stuck with us. The machine did its job. The page was the problem.

When a university's own old page is the problem

Old pages rarely get deleted. Nobody links to them any more, but they're still live, and AI tools still find them.

The ones that turned up in answers:

• St Andrews' facts page still says the Guardian ranked it "second in the UK". That was the 2026 edition; the Guardian 2027 has St Andrews 4th. ChatGPT and Perplexity both repeated it as current. (In the Times 2027 table, St Andrews is 1st. The current claim is the stronger one.)
• Newcastle's employability page leads with "94% of 2020/21 graduates". Google gave it as Newcastle's current figure. HESA's 2023/24 figure is 87.6%.
• A Dundee news story from 2022 ("90%") was still Google's source for Dundee's graduate outcomes. HESA's current undergraduate figure is 90.1%, so the number happens to be close. The year is three editions out.
• Edinburgh's international fees for 2025 entry were quoted by ChatGPT as the 2026/27 fees.
• Glasgow's rankings pages give UK positions with no year attached. They don't match the 2027 tables.

Then we checked every university's rankings or "facts and figures" page, all 31, and saved a dated PDF of each one.

The ten: Oxford, Cambridge, Bristol, Strathclyde, Glasgow and Manchester show older editions; St Andrews, Cardiff, LSE and UCL mix them. Several of them undersell the university. Manchester's page shows positions up to 16 places lower than its 2027 ranks. Bristol's lists the Guardian at 15th; it's 12th.

It isn't only universities. When we checked on 21 October 2026, UCAS's page for parents said "For 2027 entry, your child will typically pay up to £9,790 per year". The 2027/28 cap is £10,050. UCAS updated the page to the 2027 figure after we got in touch.

Each university's details are in the report cards, with the PDF for every page we describe as out of date.

Reduced offers reported as standard offers

Technical, and easy to get wrong.

Many universities make lower "contextual" or "widening-access" offers to applicants from particular schools or backgrounds. Course pages usually give both the standard offer and the reduced one. AI answers often merged them, dropped the label, or garbled the grades.

"AAA*" is worth a second look. It isn't a real A-level combination, because grades are always written highest first. It's what you get when a range like "A*A*A* to AAA" is cut and stitched back together.

In one of our early pilot runs, Google also gave Hull York Medical School's contextual AAB as the medicine offer. Hull York's offers page labels it clearly. The answer flattened the two.

Two of the Edinburgh errors were cited to UCAS course pages rather than Edinburgh's own. Those pages show the entry range and the widening-access grades together, and the answers ran them into one. The entry details on UCAS course pages are supplied and kept up to date by providers, the universities and colleges that run the courses, not by UCAS.

Right fact, wrong year

The most common error type in the study, and the hardest to spot, because the figure is real.

England's tuition fee cap is £9,790 for 2026/27 and £10,050 for 2027/28. Students applying now are mostly starting in 2027. Answers about fees kept giving £9,790, sometimes while citing a university site that already published the 2027/28 figure. One described £10,050 as the "Year 2" fee, which is only true for someone who started in 2026.

Edinburgh and St Andrews already publish their 2027-28 fees. The answers quoted 2026/27 anyway.

League tables follow the same pattern. ChatGPT said Warwick was "1st for Economics" in the Times Good University Guide. It was, in 2026. In the 2027 table it's 2nd. ChatGPT cited a revision site for that, but Warwick's own reputation page still carries the same line in its highlights.

What makes this hard for applicants: nothing in the answer looks wrong. The number exists. It's just last year's.

"Is X better than Y?" Ask Reddit

We asked: "For economics, is LSE, Warwick, UCL or St Andrews the better choice?"

Google's AI Overview cited nine sources. Four were Quora threads, the oldest from 2016. The rest were a Reddit post, two Instagram posts, a Student Room thread and a student-housing company's "10 best" list. No university site, no league table, no official data. Its opening verdict was cited to Quora.

Forum posts can be useful. Students say things on Reddit that no prospectus would. But a confident one-paragraph verdict built from them reads like a ranking, and it isn't one. Across the "which is better" questions, forum and Q&A sites showed up on all three platforms, and Google leaned on them most.

The £9,800 mistake

We asked Perplexity what it would cost a student from England to study at Edinburgh, St Andrews, Glasgow or Aberdeen.

It said St Andrews charges students from the rest of the UK for only three years of a four-year degree.

It doesn't. Fees are payable every year, and the St Andrews page Perplexity cited says so.

The three-years-of-four policy is real. It belongs to Edinburgh Napier.

That's about £9,800 missing from a family's budget, from one sentence that sounded entirely reasonable. Google, asked the same question, got it right.

What the pages that got it right have in common

The good news is that some pages worked on every platform, and they weren't hard to build.

• Contextual-offer pages. When we asked Bristol, Edinburgh, Leeds and Glasgow about contextual offers, 11 of the 12 facts were right. Each university has one short page that answers exactly that question.
• The Department for Education's fee page on gov.uk: one page, every figure, a year next to each. Asked about England's fee cap, the platforms quoted it correctly.
• Lancaster's Psychology course page. Short. The offer is labelled "A level requirements AAB", with the entry year right above it. ChatGPT and Google both got it right.
• Discover Uni, the official course-data site. ChatGPT quoted its salary figures exactly, including £46,500 for Bath mechanical engineering five years after graduating.
• Dundee's "all entry requirements in one place" pages, which gave correct Psychology grades on all three platforms.
• Edinburgh's plain-text entry-requirements pages. Edinburgh found in 2025 that ChatGPT was giving last year's grades, because the real ones sat behind dropdown menus. It built plain-text companion pages, and its own March 2026 retest had ChatGPT giving correct grades for 9 of 11 courses, up from 3 of 11. In our runs, ChatGPT and Google both read the Psychology page correctly.

Going back through the pages behind the correct answers, very few were the ones a university would put in a brochure. A contextual-offers page. A fee table on gov.uk. A course page with nothing on it but grades and a year. The pages that get the design budget weren't the ones being read.

That matches what Axelerant found from the other direction: 34% of UK universities keep entry requirements out of the page's initial HTML, where many AI tools never look.

Named, cited... and not visited

One more thing we noticed while recording sources.

When ChatGPT names a university in an answer, the name is often underlined. Click it, and you don't go to the university's website. ChatGPT opens its own side panel about the university. In Google's AI Overviews, clicking a university's name runs a new Google search for it.

So being named in an answer doesn't mean a visit. We logged, for every answer, whether a reader could reach the university's own site in one click.

*One of Google's was reachable only through the "Website" button on a Business Profile card, not a citation.

The citations do link out. The names don't. On ChatGPT, clicking a university's name opened ChatGPT's own panel every time; on Google, it ran a new Google search. So a university can be named in an answer, cited in it, and still get no visit unless the reader goes looking for the small source link.

University report cards

How to read these cards

Each card covers what ChatGPT, Perplexity and Google AI Overviews said about the university across our 21 questions (63 answers), and what its own website shows.

• Named is the number of answers that mentioned the university.
• Facts checked is the number of specific claims about that university we could test against an official source.

Seven universities weren't named in any scored answer. The questions used named lists, and they weren't on any list. Their cards cover only their own rankings page.

The card scores add up to 162 of the 198 facts. The other 36 don't belong to any one card: 17 compared several universities at once (Sheffield's salary ranking is one), 16 were general facts such as England's fee cap, and 3 were about Aberdeen, which isn't in our set of 31. All 198 are in claims.csv.

Where we suggest a fix, it's one we think the web team can make in a day. Every page we describe as out of date is backed by a dated PDF or screenshot, and where we could, a public copy on the Wayback Machine, linked from the card.

Jump to: Bath · Birmingham · Bristol · Cambridge · Cardiff · Dundee · Durham · Edinburgh · Exeter · Glasgow · Imperial · King's College London · Lancaster · Leeds · Liverpool · LSE · Loughborough · Manchester · Newcastle · Nottingham · Oxford · Queen Mary · Queen's Belfast · Sheffield · Southampton · St Andrews · Strathclyde · Surrey · UCL · Warwick · York

University of Bath

CUG 9th · Guardian 8th · Named 4 · Facts checked 2 · Correct 2

What AI got right: ChatGPT quoted the official Discover Uni salary figure for Bath mechanical engineering exactly (£46,500 median five years out). Perplexity's 98% figure for civil engineering graduates was right too. Bath's rankings page shows all three 2027 tables, with links to each. [PDF 22 Oct]

One thing to fix: Bath's civil engineering outcomes page is headed "Graduate Outcomes 2022/23", but the text below says the data is for "2023/24 leavers". A small mismatch, and exactly the kind that makes an AI hedge on the year or pick the wrong one. [PDF 22 Oct]

University of Birmingham

CUG 12th · Guardian 14th · Not named in scored answers

What's working: Birmingham's league-tables page shows CUG, Guardian and Times 2027 alongside QS and THE 2027. Of the 31 universities, it's one of the cleanest rankings pages. [PDF 22 Oct]

One thing to fix: nothing from this study.

University of Bristol

CUG 17th · Guardian 12th · Times 10th · Named 7 · Facts checked 8 · Correct 6

What AI got right: Bristol's medicine and contextual-offer pages did their job. All three platforms gave the size of the contextual reduction ("up to two grades below the standard entry requirements") correctly. ChatGPT and Google gave the medicine offer correctly (AAA with Chemistry), and ChatGPT labelled the 2,240 UCAT threshold as a 2026 figure, as Bristol does.

One thing to fix: Bristol's rankings page still shows the 2026 tables. It lists the Guardian at 15th, but Bristol is 12th in the Guardian 2027 table, so the page undersells the university. [PDF 21 Oct · Wayback copy, 23 Oct]

(Neither error came from Bristol's pages. Perplexity's A*AA for medicine had no citation. Google said Bristol "automatically drops" offers for eligible applicants; Bristol's own page says they are "automatically considered for" a contextual offer.)

University of Cambridge

CUG 1st · Guardian 2nd · Times 4th · Named 6 · No checkable claims about Cambridge in the answers

What's working: the claims Cambridge makes on its undergraduate pages are still true this year.

One thing to fix: the "Why Cambridge" page says "Number 1 in CUG 2026" and "Top 3 in Guardian 2026". Cambridge is still 1st in CUG 2027 and 2nd in the Guardian 2027, so swapping in the current edition costs nothing and removes a year-old label. Cambridge has no dedicated rankings page, so this one carries the weight. [PDF 21 Oct · Wayback copy, 23 Oct]

Cardiff University

CUG 27th= · Guardian 43rd= · Named 9 · Facts checked 13 · Correct 4

What AI got right: Google gave the medicine offer exactly as Cardiff's page states it: AAA, Biology or Chemistry, plus a second science. ChatGPT gave Psychology's "AAA–ABB covering standard and contextual" correctly and described Cardiff's UCAT policy accurately.

What went wrong, and isn't Cardiff's doing: Perplexity cited Cardiff's own course page for "A*AA" when the page says AAA. On graduate outcomes, Google's lead source chip pointed to Cardiff Metropolitan University.

One thing to fix: Cardiff's rankings page contradicts itself. The table says CUG 2027 27th, while the text below says "22nd in CUG 2026", and it gives two different QS ranks. Old news releases (94% from 2014/15, 95% from 2020/21) were also what AI found for graduate outcomes. HESA's 2023/24 figure is 88.9%. [PDFs 21 and 23 Oct; the page blocks archiving tools, so there's no Wayback copy]

University of Dundee

CUG 25th · Guardian 21st= · Times 20th · Named 10 · Facts checked 11 · Correct 7

What AI got right: Dundee is the study's clearest example of a rankings page working. ChatGPT and Google both quoted Dundee's own 2027 ranks correctly from it (Times 20th, Guardian 21st=, CUG 25th), along with its Guardian subject ranks. All three platforms gave the Psychology offer (CCC, with BBC for Year 2 entry) correctly, from the "all entry requirements in one place" pages.

One thing to fix: that all-in-one page is about 190,000 characters, with countries listed A to Z. Google gave the standard offer but not the widening-access CCD, and used wording from non-UK rows, which suggests it read only the top. Putting the UK rows first would close the gap. Separately, a 2022 news story ("90%") is still Google's source for graduate outcomes. HESA's 2023/24 figure is 89.0% (all levels; 90.1% for undergraduates).

(Perplexity took Dundee's ranks from Wikipedia and was a year out. That one isn't on Dundee.)

Durham University

CUG 6th · Guardian 6th · Times 6th · Named 10 · Facts checked 10 · Correct 8

What AI got right: Perplexity and Google read Durham's course pages and got everything right: Psychology AAA, contextual BBB, GCSE Maths B/5, and Law A*AA with contextual AAB and the LNAT. Durham's rankings page leads with "6th in the UK" in all three 2027 tables. [PDF 22 Oct]

What went wrong, and isn't Durham's doing: ChatGPT said A*AA for Psychology (Durham says AAA), citing Durham's online prospectus site rather than the course page.

One thing to fix: small ones. The Law page labels its offer "Entry requirements: A*AA" without saying it's an A-level offer, and the offer sits on page 6 of 17 inside a country selector. The page also quotes graduate outcomes for 2022/23 leavers when 2023/24 data has been out since June. [PDF 22 Oct]

University of Edinburgh

CUG 19th · Guardian 11th · Named 22 · Facts checked 27 · Correct 14

What AI got right: Edinburgh's plain-text entry-requirements pages work. ChatGPT and Google both gave the Psychology range correctly (A*A*A* to AAA), with the ABB widening-access minimum and all eleven required subjects. ChatGPT and Perplexity explained the Plus Flag scheme accurately. Perplexity quoted the 2026/27 living-cost estimate to the pound.

Where it went wrong: - The top of the range got garbled into "AAA*" twice: Perplexity on Psychology, reading a joint-honours page, and ChatGPT on Computer Science. - Google called ABB "the range of grades typically needed", citing UCAS. - Google also turned "we will normally make you an offer" into "guaranteed". - ChatGPT gave 2025-entry international fees as 2026/27.

One thing to fix: the undergraduate rankings page lists no UK league tables at all. So every platform quoted QS (35th) for Edinburgh, and Google guessed "around top 5 to 19". Adding "11th in the Guardian University Guide 2027" would give AI answers a better and more accurate number to use. [PDF 21 Oct]

University of Exeter

CUG 11th · Guardian 16th · Named 3 · Facts checked 2 · Correct 0

What AI got right: Exeter's own fee page gave Perplexity the correct 2027 figure (£10,050) when it was asked about England's fee cap. Exeter's rankings page is a model for everyone else: a table with a Year column beside every guide (Times 2027 16th, Guardian 2027 16th, CUG 2027 11th), so nothing on it can be read as the wrong edition. [PDF 22 Oct]

What went wrong: asked about graduate outcomes, Google said 92% with no traceable source, and Perplexity quoted a secondary site's 89.3%. HESA's 2023/24 figure for Exeter is 87.8% (all levels).

One thing to fix: none of the answers cited an Exeter page for graduate outcomes. One short, dated sentence with the HESA 2023/24 figure, on a page about careers or outcomes, would give AI answers something official to quote.

University of Glasgow

CUG 29th · Guardian 21st= · Named 16 · Facts checked 11 · Correct 9

What AI got right: a strong record, mostly from Glasgow's own pages. The answers covered: - the medicine requirements, word for word (AAAAB in S5 plus BB at Advanced Higher, and AAA at A-level with Chemistry); - the adjusted-entry scheme, correct on all three platforms; - the Law offer; - the 2026/27 RUK fee of £9,790.

One thing to fix: Glasgow's awards and rankings pages give UK ranks with no year attached (CUG 31st, Guardian 24th, Times 22nd). Those don't match the 2027 tables (29th and 21st=), so they're from an earlier edition. Adding the edition name and year, or just updating the numbers, fixes it. The main Common Law course page also still shows 2025 entry. [PDF and screenshot 21 Oct · Wayback copy, 23 Oct]

Imperial College London

CUG 5th · Guardian 5th · Times 3rd · Named 10 · Facts checked 8 · Correct 4

What AI got right: all three platforms got Imperial's international fees right, from Imperial's own pages, labelled 2026/27 (£45,500 engineering, £42,700 Business). Perplexity gave the Computing offer and the TMUA correctly. The league-tables page shows all three 2027 tables and says when it was last updated. [PDF 22 Oct]

Where it went wrong: - ChatGPT understated the Computing offer by a grade. - Google's display dropped the asterisks, so "A*A*A" appeared as "AAA". - Perplexity said Imperial fixes overseas fees at the entry rate, which is the opposite of Imperial's policy.

One thing to fix: to answer "what are Imperial's international fees?", an AI has to combine four faculty pages. Third-party sites that list every band on one page were cited instead. A single undergraduate fees summary page would help. (Also: on the league-tables page, the "Complete University Guide 2027" link opens the Guardian 2027 table.) [PDF 22 Oct; link checked by hand]

King's College London

CUG 16th · Guardian 19th · Not named in scored answers

One thing to fix: the About page shows QS 2026 (31st) and THE 2025 (36th), and no UK league table. QS 2027 and THE 2027 are both out, so the global ranks are a year or two behind as well. [PDF 21 Oct]

Lancaster University

CUG 10th · Guardian 27th · Named 3 · Facts checked 3 · Correct 2

What AI got right: Lancaster's Psychology page may be the cleanest course page in the study. It's short, the offer is labelled "A level requirements AAB", and the entry year is right above it. ChatGPT and Google both got it right.

One thing to fix: the About page says "top 10 UK university" without naming the table. It's true of CUG 2027 (10th), so say so: "10th in the Complete University Guide 2027" can be quoted, while an unattributed "top 10" can't. [PDF 21 Oct]

(Perplexity's "least competitive" verdict rested on an unsourced offer-rate site, not on anything Lancaster published.)

University of Leeds

CUG 18th · Guardian 26th · Named 6 · Facts checked 5 · Correct 5

What AI got right: a perfect record. All three platforms explained Access to Leeds correctly (up to two grades lower, AAA to ABB), from Leeds' own pages. ChatGPT and Google gave the Law offer (AAA) correctly.

One thing to fix: the About page gives only "Top 100" in QS 2027, with no UK league-table rank. The Law course page mixes editions, with CUG 2026 and Times 2026 tiles next to CUG 2027. [PDFs 22 Oct]

University of Liverpool

CUG 22nd · Guardian 30th · Not named in scored answers

What's working: the rankings page shows all three 2027 UK tables plus QS and THE 2027. [PDF 21 Oct]

One thing to fix: nothing from this study.

London School of Economics

CUG 3rd · Guardian 3rd · Times 2nd · Named 4 · No checkable claims about LSE in the answers

One thing to fix: the About page says LSE was "ranked top in the UK for 2025 and 2026 by the Times and Sunday Times Good University Guide" but gives no 2027 UK position, though LSE is 2nd or 3rd in all three 2027 tables. AI answers that quote the page will repeat last year's claim. [PDF 21 Oct · Wayback copy, 23 Oct]

Loughborough University

CUG 8th · Guardian 9th · Times 13th · Named 6 · Facts checked 6 · Correct 4

What AI got right: ChatGPT gave all three 2027 ranks correctly, Google gave CUG 8th, and ChatGPT quoted the Discover Uni salary for mechanical engineering exactly. Loughborough's achievements and "why study here" pages now show the 2027 tables. [screenshots 21 Oct]

One thing to fix: Google's "520+ acre campus" matches Wikipedia's infobox (523), while Loughborough's own figure is 440 acres. Correcting the Wikipedia infobox, with a citation to the university page, is the fix. On 20 Oct Google was also still quoting older 11th–12th ranks, which likely came from a version of the achievements page from before it was updated.

University of Manchester

CUG 24th · Guardian 20th · Named 7 · Facts checked 6 · Correct 4

What AI got right: Perplexity and Google gave the Computer Science offer correctly (A*AA, with A* in Maths). Perplexity got the MBChB international fee (£39,900 for years 1–2) and the "up to 7% a year" rise policy right.

Where it went wrong: ChatGPT understated the CS offer and claimed it had dropped since 2026. Google's international fee bands came from a student-accommodation site, and its top lab fee was below what Manchester's own CS page says (£39,700).

One thing to fix: the rankings page still shows CUG 2026 (28th), the Times 2025 (27th) and the Guardian 2026 (36th). In the 2027 tables Manchester is 24th (CUG) and 20th (Guardian), so the page undersells it by up to 16 places. The Times row also links to the Complete University Guide's website. [PDF 22 Oct · Wayback copy, 28 Sep]

And a second, for the fees question: Manchester has no single page of international fee bands; fees are listed per course. A summary page would give AI an official source to cite instead of a housing marketplace.

Newcastle University

CUG 35th · Guardian 82nd · Named 3 · Facts checked 2 · Correct 0

What's working: the UK rankings page shows all three 2027 tables. [PDF 21 Oct]

One thing to fix: the employability page still leads with "94% of 2020/21 graduates". Google quoted it as Newcastle's current figure. HESA's 2023/24 figure is 87.6% (all levels). Updating that one line, with its year, fixes the most visible error about Newcastle in the study. [PDF 21 Oct · Wayback copy, 23 Oct]

University of Nottingham

CUG 26th · Guardian 51st= · Not named in scored answers

What's working: the rankings page lists all three 2027 UK tables with their positions (Times 24th, CUG 26th, Guardian 51st). [PDF 22 Oct]

One thing to fix (small): the section heading says "A top 20 UK university", directly above a table where the 2027 UK league tables put Nottingham 24th to 51st. The top-20 claim comes from QS and THE's UK placings (16th, 19th). Naming the source in the heading ("top 20 in the UK, QS World University Rankings 2027") keeps the claim and makes it quotable. The same table still lists THE 2026, though THE 2027 is out.

University of Oxford

CUG 2nd · Guardian 1st · Times 5th · Named 9 · Facts checked 4 · Correct 3

What AI got right: Perplexity and Google both had Oxford first in THE's world rankings for the 11th year, and Oxford's QS 2027 position, correct. ChatGPT's "23,000+ applications" matches Oxford's own admissions statistics.

One thing to fix: the facts and figures page still says "number one in The Best UK Universities 2026" and shows CUG and Guardian 2026. Oxford is 1st in the Guardian 2027, so updating the page only adds to the claim. Also, asked an undergraduate question, all three platforms cited Oxford's postgraduate "Why choose Oxford?" page. [PDF 21 Oct · Wayback copy, 23 Oct]

Queen Mary University of London

CUG 40th · Guardian 54th · Not named in scored answers

One thing to fix (optional): the About page shows no league-table positions at all. That's a choice, not an error. If Queen Mary wants AI answers to quote a current rank rather than a third party's guess, one dated line would do it. [PDF 21 Oct]

Queen's University Belfast

CUG 23rd · Guardian 43rd= · Named 12 · Facts checked 17 · Correct 14

What AI got right: the best record in the study. All three platforms quoted QUB's international fee schedule exactly (£22,400 / £26,600 / £39,200 / £47,000). The Law (AAA) and Medicine offers were right, as were CUG 23rd and QS 2027. Google also found the WUSCA 2026 University of the Year award, which the other platforms missed.

One thing to fix: QUB's own pages disagree with each other. The study page says "Top 200 (QS 2026)" and 91% graduate outcomes; the rankings page says "Top 175 (QS 2027)"; the Vice-Chancellor's statement says 93%. Perplexity repeated the older study-page figures. Bringing that page into line with the others is the fix. [PDFs 21 Oct]

University of Sheffield

CUG 13th= · Guardian 15th · Named 3

What AI got right: ChatGPT ranked the four universities by mechanical-engineering graduate salary in the same order as the official Discover Uni figures. Sheffield's is £33,000 at 15 months. The rankings page shows all three 2027 tables. [PDF 22 Oct]

One thing to fix: nothing from this study.

University of Southampton

CUG 15th · Guardian 18th · Not named in scored answers

What's working: the rankings page shows all three 2027 tables. [PDF 22 Oct]

One thing to fix: the Computer Science BSc page's summary box says "Typical Offer AAB" with no entry year and no mention of A-levels. The only year visible on the page is a "2026-27 Course Description Document". The 2027/28 requirements sit on a separate tab. In our early open-question runs, ChatGPT overstated Southampton's CS offer by a grade or two. Putting "2027 entry" and "A-levels" next to the offer would help. [PDF 22 Oct]

University of St Andrews

CUG 4th · Guardian 4th · Times 1st · Named 13 · Facts checked 15 · Correct 6

What AI got right: ChatGPT gave Times 2027 1st and CUG 2027 4th correctly. Google got the fee position right for Scottish students, and correctly said English students pay every year.

What went wrong, and isn't St Andrews' doing: Perplexity told English applicants that St Andrews charges for only three of the four years, an error of about £9,800. It cited St Andrews' own page, which says fees are payable every year. Three-of-four is Edinburgh Napier's policy.

One thing to fix: the facts page still says the Guardian ranked St Andrews "second in the UK" (the 2026 edition). ChatGPT and Perplexity both repeated that as current. The Guardian 2027 has St Andrews 4th, while the Times 2027 has it 1st. Leading with the current editions keeps the strongest claim and drops the stale one. [PDF 21 Oct · Wayback copy, 15 Oct]

University of Strathclyde

CUG 39th · Guardian 17th · Times 9th · Named 8 · Facts checked 4 · Correct 2

What AI got right: Google was the only platform to answer "which is least competitive for Psychology?" with the actual grades, correctly identifying Strathclyde (ABB–BBB) as the lowest of the four. ChatGPT gave the Discover Uni mechanical-engineering salary correctly.

One thing to fix: the Psychology course page shows two different versions of the entry requirements and doesn't say which entry year they're for. ChatGPT declined to give a grade at all. Separately, the "Why Strathclyde" page shows the Times 2026 award but not Strathclyde's Times 2027 top-10 place. [screenshot and PDF 21 Oct · Wayback copy of "Why Strathclyde", 23 Oct]

University of Surrey

CUG 21st · Guardian 39th= · Times 25th · Named 3 · Facts checked 2 · Correct 2

What AI got right: ChatGPT and Google both gave the Psychology offer (AAB) correctly.

One thing to fix: the course page says "Typical offer AAB" without saying it's for A-levels, and the About page gives only "QS 2027 top 250", with no UK table. Neither caused an error this time, but both make the page harder to quote. [PDF 21 Oct]

University College London

CUG 13th= · Guardian 13th · Named 6 · Facts checked 5 · Correct 3

What AI got right: Perplexity read UCL's pages and got everything right: the Computer Science offer (A*A*A, contextual A*AB), the TARA test, and UCL's own admissions figures (4,006 applications, 805 offers, 228 places). ChatGPT found the Scottish Advanced Higher requirement, which sits behind a tab and dropdown.

One thing to fix (small): the rankings page has a 10-year table with the 2027 Guardian and Complete University Guide positions (13th in both), the kind of dated record AI answers can quote. The Times Good University Guide row stops at 2026, though, and the only Times item shown is the "University of the Year 2024" tile. Adding the 2027 Times position would finish the job. [screenshots 28 Oct]

University of Warwick

CUG 7th · Guardian 7th · Times 7th · Named 4 · Facts checked 1 · Correct 0

What's working: the reputation page has an "at a glance" table with all three 2027 UK tables (7th in each), and the economics department's key-statistics page shows its 2027 subject positions. [PDF 22 Oct]

What went wrong: ChatGPT said Warwick was "1st for Economics" in the Times 2026, which was true then but not now (2nd in 2027). It cited a revision site. Warwick's own reputation page says the same thing, though: its departmental highlights still read "Economics: Ranked 1st in the Good University Guide 2026".

One thing to fix: update that highlight to the 2027 position, or label it clearly as last year's. [PDF 22 Oct · Wayback copy, 23 Oct]

University of York

CUG 20th · Guardian 46th · Not named in scored answers

What's working: the About page shows CUG 2027 (20th). [PDF 22 Oct]

One thing to fix (optional): it's the only UK table shown, which is a choice. Nothing in the study depends on it.

A checklist for university web teams

We checked 198 facts that ChatGPT, Perplexity and Google AI Overviews gave applicants about 31 UK universities. When an answer came from a university's current official page, it was right 79% of the time. When it came from the university's own out-of-date page, it was right 1 time in 15 (7%).

So most of what follows is about your own pages. Nothing here needs a new platform or an "AI strategy". Most items are an afternoon's work for someone with CMS access.

Each item below says what to do, what we saw, and a quick way to check your own site.

1. Put the year right next to every number

Fees, grades, rankings, outcomes. Not in the page header or the metadata, but beside the figure.

What we saw: when we checked, UCAS's parents' page said "For 2027 entry, your child will typically pay up to £9,790 per year". The 2027/28 cap is £10,050. (UCAS has since updated it.) Google AI Overviews showed a student-housing site's 2025-26 fee table as 2026/27. Surrey's course page says "Typical offer AAB", and Durham's Law page says "Entry requirements: A*AA", without saying which qualification either one is for.

Quick check: search your fees page for "£". Does every figure have a year within a line of it?

2. Update your rankings page the week each league table comes out

CUG publishes in June, the Guardian and the Times in September. Delete superseded editions, or label them clearly as past years.

What we saw: 14 of the 31 universities show 2027 league tables on their own rankings pages. Ten show last year's figures or mix them with new ones, and seven give no dated UK rank at all. ChatGPT and Perplexity both quoted St Andrews' Guardian 2026 "second in the UK" as current. It came straight from the university's own facts page. The Guardian 2027 table has St Andrews 4th.

Quick check: open your rankings or "facts and figures" page today. Is every edition 2027?

3. Retire old stat pages or stamp them clearly

Old press releases and "why study here" pages keep getting cited long after anyone at the university has forgotten them.

What we saw: - Newcastle's employability page leads with "94% of 2020/21 graduates". HESA's 2023/24 figure is 87.6%. - Google answered a Dundee question with a 2022 news story. - Queen's Belfast's own pages give both 91% and 93%. - Durham's Law course page still quotes graduate outcomes for 2022/23 leavers (87%), though HESA published the 2023/24 results in June 2026. - Old or wrong own pages were the least accurate source in the study: answers that used them were right 1 time in 15 (7%).

Quick check: search your site for last year's league-table names and for "2020/21" or "2021/22". Each hit should either be updated or carry a date in its first line.

4. Keep entry requirements in plain text, not behind dropdowns

If the grade only appears after someone clicks a country selector or opens a tab, AI tools may never see it.

What we saw: Edinburgh found in 2025 that ChatGPT was quoting the previous year's grades. It built plain-text companion pages for entry requirements, and its own March 2026 retest had ChatGPT giving correct grades for 9 of 11 courses, up from 3 of 11. In our runs, ChatGPT and Google both read those pages correctly. Axelerant's audit found 34% of UK universities keep entry requirements out of the page's initial HTML.

Quick check: view the page source (Ctrl+U) and search for your typical offer, e.g. "AAB". If it isn't there, many AI tools won't find it either.

5. Put the UK answer first on long pages

What we saw: Dundee's all-in-one entry-requirements page is about 190,000 characters, with countries listed A to Z, so the UK rows come near the end. Google gave the right grades but missed the widening-access offer and used wording from the non-UK rows. On Durham's Law page the UK offer appears on page 6 of 17 when printed, after the module list and inside a "Select Country/region" box.

Quick check: how far down the page is your UK A-level offer? If it's below the fold on a laptop, move it up.

6. Spell out "standard offer" and "contextual offer" in words

What we saw: reduced offers came back as standard ones. Hull York Medical School's contextual AAB was given as the offer. Edinburgh's widening-access grades were presented as the typical requirement twice. Grade ranges got mangled into things like "AAA*", and Google's display turned "A*A*A" into "AAA".

Fix: one labelled line each, e.g. "Standard offer: A*A*A. Contextual offer: AAA." Avoid putting both in one sentence joined by "or", and avoid ranges without a label. The same goes for the entry details you supply to UCAS, since its course pages show what each university provides.

7. Answer the whole question on one page

If applicants ask "how much are international fees at X?", have one page that answers exactly that, with every band on it and a year next to each.

8. Title pages the way people ask

Third-party pages that got cited had titles shaped like the question: "Law degree requirements", "University of Surrey admissions statistics", "[University] tuition fees 2026". Official pages were titled "Law M101" or just "Psychology".

Fix: "Psychology BSc: entry requirements for 2027 entry" says more than "Psychology" to a person and to a machine.

9. Make the header agree with the entry year

What we saw: Bath's civil engineering outcomes page is headed "Graduate Outcomes 2022/23", and the text underneath says the figures are for "2023/24 leavers". Glasgow's main Common Law course page still shows 2025 entry, with the 2026 version on a different URL. Mixed signals give an AI tool a reason to pick the wrong year.

Quick check: on your five most-visited course pages, do the H1, the start date and the requirements section all name the same year?

10. Point undergraduate questions at undergraduate pages

What we saw: asked an undergraduate question, all three platforms cited Oxford's postgraduate "Why choose Oxford?" page.

Fix: if a page is for postgraduates, say so in the title and first line. Link the undergraduate equivalent near the top.

11. Check you aren't blocking the tools you want to be read by

Some sites block AI crawlers in robots.txt or at the hosting level, sometimes without the web team knowing. A page that a crawler can't fetch can't be quoted correctly, however good it is.

Quick check: run your domain through our AI crawler checker. There's more on what each crawler does in Is your website blocking AI?

12. Ask the AI about yourself, four times a year

Pick 5–10 questions applicants actually ask, naming your university, and run them on ChatGPT, Perplexity and Google. Check each figure against your own page.

Good times to do it: - January, before the UCAS deadline. - June, after CUG publishes. - September, after the Guardian and the Times. - October, in the Oxbridge and medicine application window.

Use a fresh, signed-out session and name your university in the question. ChatGPT tailors open "which universities…" shortlists to your location, so a question asked from your office won't show what a student in another city sees.

Checking an AI answer before you rely on it

If you're an applicant or a parent, five checks catch most of what we found:

1. Find the year next to every number. If there isn't one, it may be last year's.
2. Click the source. Is it the university, UCAS, gov.uk or Discover Uni?
3. Check the date on the university's page too. Its own old pages were the least reliable source in the study.
4. Separate the standard offer from the reduced one. Read the course page for both.
5. Ask twice, or ask a second tool. If the numbers differ, check the course page before you trust either.

The full guide, with examples from the study: How to check what ChatGPT tells you about a university.

What the platforms said

We sent OpenAI, Perplexity and Google the specific errors described on this page on 26 October 2026 and asked for comment by 30 October.

OpenAI responded with background information, which it asked us to paraphrase rather than quote. It said that, like any language model, ChatGPT can make mistakes, partly because it has to interpret what the user means. It said its latest model, GPT-6, gives more accurate and better-contextualised answers when it searches the web, and that improving factual accuracy is something the whole industry is working on. It encourages people to open and check the sources ChatGPT provides, and pointed to its guidance on how ChatGPT searches the web and shows sources.

Google's press agency asked for screenshots, transcripts and the AI Overview text for each item so that Google could investigate. We sent them on 27 October. Google had not commented by the time of publishing. We'll add any response here.

Perplexity did not respond by 30 October.

UCAS replied on 27 October about the points on this page that involve it. Its parents' fees page now shows the figure for 2027 entry. The entry information on UCAS course pages is maintained by providers, the universities and colleges that run the courses. And the survey question behind the 73% figure was asked of all respondents using AI, so it covers their use of AI across a range of topics, not only when researching higher education.

We also wrote to every university named in a report card with a suggested fix, a week before publishing. None asked for a correction. Where a university has updated a page since our check, the card says so.

Method, limits and data

When and how. All runs were made on 20 and 21 October 2026, from Scotland.

• ChatGPT: signed in, with Temporary Chat on (no memory) and search on. A new chat for every question. Runs used the default model for a Temporary Chat at the time. ChatGPT didn't show which model version answered, so we haven't recorded one. OpenAI was rolling out GPT-6 in October 2026, so some or all of our runs may already have used it.
• Perplexity: signed out, in a private browser window. (Earlier signed-in tests carried details from the account into supposedly neutral answers, so we didn't use signed-in Perplexity for any scored run.)
• Google: signed out, private window, google.co.uk. An AI Overview appeared for all 21 questions.

Each question was asked once per platform, pasted exactly as written.

Scoring. Each fact was given one outcome:

Only facts about the universities named in the question were scored. If an answer added other universities, we noted them but didn't score them.

Evidence. Nothing on this page is described as out of date unless we hold a dated PDF or screenshot of the page. Where we could, we also saved a public copy to the Internet Archive's Wayback Machine, linked from each report card. One page, Cardiff's rankings page, couldn't be archived: its bot protection blocked both the Wayback Machine and archive.today, so the dated PDFs (21 and 23 October) are the evidence. We also found our own page-reading tool returning cached, older versions of pages six times, so every page cited here was checked in a browser.

Limits. - One run per question, per platform. AI answers vary from one ask to the next; in our early runs, ChatGPT gave two different grade requirements for the same course within an hour. - A snapshot of October 2026. Platforms and pages change. We'll re-run the core questions in January 2027, before the UCAS deadline. - 31 universities: the 24 Russell Group members plus seven others. Not every university was named in a scored answer. - Asked from Scotland. Named-list questions remove most of the location effect, but not necessarily all of it.

How this fits with other work. UCAS's November 2025 survey found that 73% of respondents using AI had recognised or received incorrect information from it, across everything they used AI for (Jo Richards, Wonkhe, March 2026). We measured accuracy on university questions specifically, and traced where the wrong answers came from. Edinburgh ran its own test of 100 searches in March 2026 and found 79% accuracy, with Perplexity lowest; our results point the same way. St Andrews studied ChatGPT mentions in November 2025, and Axelerant audited how readable university sites are to AI. Ours checks individual facts about 31 universities on three platforms, and traces each one to the page the answer used.

The data. Every scored claim, every answer and the full rankings-page audit are available as CSV files under a CC BY 4.0 licence, with the pilot runs in a separate, labelled file.

• Dataset on Zenodo (DOI: 10.5281/zenodo.23594680)
• Files: claims.csv (198 rows), runs.csv (63), queries.csv (21), rankings_page_audit.csv (31), pilot_claims.csv (37), pilot_runs.csv (8)

FAQ

Is ChatGPT accurate about UK university entry requirements? Mostly. In our October 2026 checks, ChatGPT was right on 72% of the facts it gave about 31 UK universities. Its errors included giving Durham Psychology a grade higher than Durham asks for, and quoting last year's fees as this year's. Check the course page for your entry year before relying on any grade.

Which AI was most accurate in this study? ChatGPT (72% of facts correct), then Perplexity (61%) and Google AI Overviews (55%). The source mattered more than the platform: answers that drew on a university's current page were right 79% of the time, and answers that drew on its own out-of-date pages 7%.

Why do AI answers quote last year's league tables? Often because the university's own website still shows them. 17 of the 31 universities we checked had a rankings page showing last year's tables, mixed editions or no dated UK rank. AI tools repeat what those pages say.

How can I check an AI answer about a university? Look for a year next to every figure, click through to see whose website the source is, and confirm grades and fees on the university's course page for your entry year. Our full guide has eight checks.

Can universities fix what AI says about them? Partly, yes. The most common errors traced back to pages universities control: old rankings, undated figures, and entry requirements hidden behind dropdowns. Edinburgh's own retest after rebuilding its entry-requirements pages saw ChatGPT's correct answers rise from 3 of 11 courses to 9 of 11. Our checklist for web teams lists 12 fixes.

Much of what went wrong in these answers wasn't invented. It was a real page, a year old, or a fee for the wrong intake, repeated with total confidence. That's fixable, mostly by universities, and mostly in an afternoon. Until it is, the course page for your entry year is still the only answer worth trusting.