AI Names Universities and Cites Student Housing Instead

We asked Google’s AI Overview which are the best private universities in Madrid, five times, with Spain and Spanish set explicitly on every run.

Universidad Pontificia Comillas was named in all five. Good news for Comillas.

Now look at where the recommendation came from. Eight domains were cited across the runs. The three present in every one are two student residences and a guidance platform. Not a single university website, in any run.

We repeated it in Paris and the same thing happened, with a French language school, a relocation agency and a tutoring firm among the stable sources.

Disclosure: EchoWi sells software that measures this, and some of the institutions in this data are customers of ours. We are not saying which: the results are published exactly as they came back, including the part that does not flatter anyone, and who pays us does not change what the search engine returned.

The short version

1. Universities get named and never cited. Zero university domains among the sources in Madrid.
2. The stable sources are businesses selling around the degree: halls of residence, tutoring firms, guidance platforms.
3. Paris repeats it with different companies playing the same roles.
4. When a university site does appear in Paris, it appears exactly once and split across subdomains. That held in both sessions we measured.
5. Being named without being cited is fragile, because it rests on what other people publish about you.

What we measured

The Madrid answer named Comillas, CEU San Pablo, Universidad Europea de Madrid, Nebrija and CUNEF. Every one of those institutions publishes a large, well-maintained website. None of those websites was used to build the answer.

Who the sources actually are

Madrid, cited in every run: smartresidences.es, a student accommodation company; a second student residence; and a study-guidance platform.

Paris, every domain the three runs cited:

Four university domains appeared, all of them Université Paris Cité, and none reached more than one run of three: u-paris.fr, crl.u-paris.fr, physique.u-paris.fr and iut.univ-paris-diderot.fr. Four subdomains of one institution, each cited once.

That is not presence, it is dispersion. The university has enough content to be found and it is spread so thin that no single page accumulates any authority.

Read that list again. A relocation agency is helping decide which university a student is told to consider.

That is not a failure of the model so much as a description of what exists to be read. A relocation agency writes “the best universities in Paris and where to live near each one”, because that article sells apartments. A university publishes a course catalogue, an admissions page and a research news feed. Only one of those two answers the question.

Re-measured on 7 August: universities appear, one run at a time

We ran the Madrid question four more times on 7 August 2026, same market, same language.

One part of the finding changed and the rest held. It is no longer true that no university website is cited: two were, ufv.es and universidadeuropea.com. But each appeared in exactly one run of four, and neither reached the stable set. The three domains present in all four runs are still student accommodation and a course-guidance platform.

The university we track was again named in every run while its own site was never cited, which is precisely the distinction this article is named after.

One thing we did not expect: Instagram was cited in three of the four runs. No university site reached two.

Both measurements are registered with their dates and run counts. The 6 August reading was not an error. It was a different photograph.

Paris held too

We ran the Paris question four more times on 7 August 2026, same market, same language.

Five of the six domains stable on the second day were already stable on the first: the language school, the tutoring firm, the relocation agency, the course directory and Dailymotion. Wikipedia rose from two runs of three to four of four. The three university domains on the second day, u-paris.fr, iut.univ-paris-diderot.fr and assas-universite.fr, appeared once each and none reached the stable set.

This is what separates a finding from a coincidence. What changes between runs is the long tail. Who holds the answer up does not, and across two cities and two sessions no university holds it up.

Why you get named and not cited

Two different mechanisms, and they matter differently.

Being named comes from the model’s training and from everything published about you. Comillas appears in rankings, press coverage, alumni pages, forums and directories accumulated over decades. That is why the answer knows the name.

Being cited comes from having a page that answers the question being asked. The question is comparative: which are the best. A syllabus does not compare. An admissions page does not compare. A research announcement does not compare. There is nothing on a university website for a retrieval layer to lift when the question is which one to choose.

We have now found the same split in four unrelated sectors. In flight booking, every stable booking domain was a metasearch engine and no transactional site reached the stable set. In consumer electronics, the one retailer cited publishes comparison guides across brands it sells. A retrieval layer can use a page that compares. It cannot use a page that only presents.

The fragile part, and it is not obvious

Being named without being cited feels fine. The institution appears, the marketing director sees the brand in the answer, everyone moves on.

It is the weakest position in the answer, for a reason worth stating plainly: your presence depends entirely on what third parties publish about you, and you do not control any of it.

Three things follow.

• You cannot fix a wrong description. If the answer says something outdated about your fees, your intake or your programmes, the correction has to happen on somebody else’s page.
• You move when they move. A residence company redesigning its blog, or dropping the article that mentions you, changes your visibility without anything changing at your institution.
• A competitor can outrank you by writing. Any institution that publishes genuine comparative content is competing for the citation slot, which is currently occupied by companies that are not universities at all.

The gap worth finding first

One detail in the Madrid data is the most useful thing this method produces.

smartresidences.es was cited by Google. Its article opens by naming a university that the answer never mentions.

The name was in a page the model read, considered good enough to cite, and it still did not survive into the summary. That single fact rules out three explanations at once: the institution’s site is crawlable, its content is readable, and it was judged relevant enough for the surrounding query. What remains is that nothing on its own pages answers the question that decides an application.

That is a content brief, and content briefs get delivered. It is a far better place to start than a visibility percentage nobody can act on.

We are not naming the institution. It did not ask to be an example, and we would rather tell a university before we tell the internet.

What a university should measure

Five questions cover an enrolment, and none of them contains your name. A query that already names you measures nothing.

The fifth is the one most institutions skip and it is worth the most. If you recruit in Latin America, Asia or the Gulf, the question is asked in that language from that country, and the answer is built from sources there. Our three-market study found that stable source sets in the United States, Spain and France did not overlap at all. Measuring your international recruitment from your own campus gives you a reassuring picture of somewhere else.

And check there is an answer at all before you buy anything to watch it. Measuring five categories in Spain in August 2026, two returned no AI Overview across thirteen runs. Universities do get one, in both countries we tested, which is why this study exists.

What to publish, and why nobody in higher education wants to

The uncomfortable part of this finding is that the fix is obvious and institutionally difficult.

The pages that win the citation compare. So the content that would take that slot from a relocation agency is content that names your peers and says where each one is stronger. Universities do not publish that, for reasons that are entirely understandable: comparison invites complaint from the institutions you rank, from faculties who feel misrepresented, and from a marketing committee that has to sign it off.

Four things that are comparative, defensible, and already exist somewhere inside most institutions:

• Employment outcomes with numbers and dates. Graduate destinations by subject, salary bands, time to first role. Most universities collect this for accreditation and publish it as a PDF nobody can extract.
• Total cost against the alternatives. Fees plus housing plus the years to completion, against the public option and against the nearest private one. Prospective families build this spreadsheet anyway; the difference is whether you build it for them.
• Subject-level honesty. “If you want research-led theory, us. If you want a placement in your second year, them.” An institution that says that once is more useful than one that claims everything.
• Why a student left, or chose elsewhere. Uncomfortable, and the single most extractable form of comparison there is.

None of that requires a bigger website. It requires publishing the comparison you already make in the admissions interview, on a page. The companies currently occupying your citation slot are doing exactly that, with less information than you have, because they are willing to write it down.

What this study does not show

• Nothing about academic quality. That AI names an institution says nothing about how it teaches, and being absent says nothing either.
• Sixteen runs is not a survey. Five in Madrid and three in Paris on 6 August, four and four on 7 August, one surface throughout. The pattern repeating in two cities on two days is what makes it interesting, not conclusive.
• One surface. This is Google’s AI Overview. ChatGPT, Perplexity and Gemini assemble answers differently and may cite differently.
• We cannot show that fixing it works. A July 2026 review of 45 GEO studies found no technique with a demonstrated causal, stable, cross-platform effect on AI citations. Anyone promising otherwise is ahead of the evidence, us included.
• Customer data is in here. Some of the institutions measured are customers of ours, and we are not saying which. Their results are published unedited, and you should weigh the study knowing a commercial relationship exists somewhere in it.

Common Questions About This Study

Why does AI name universities but cite student housing?

Because naming and citing come from different places. The name comes from decades of accumulated coverage in the model’s training data and across the web. The citation comes from a page that answers the specific question, and the question is comparative. Residences, tutoring firms and guidance platforms publish comparisons because those articles sell rooms, classes and leads. Universities publish catalogues and admissions pages, which do not compare and therefore have nothing to lift.

Is being named enough?

It is better than absence and it is the weakest position in the answer. Your presence rests on what third parties publish about you, so you cannot correct an outdated description, you move whenever they move, and any institution that starts publishing genuine comparative content can take the citation slot from businesses that are not universities.

What should a university publish instead?

Content that compares, honestly, including where you are not the right choice. Employment outcomes with figures and dates. Cost broken down against the alternatives. Subject-level comparisons against named peers. That is uncomfortable for an institution and it is exactly the shape the answer needs, which is why companies selling accommodation currently own it.

Does this apply outside Spain and France?

We measured those two and are not extending the claim. What we can say is that the split between pages that compare and pages that present has now appeared in four unrelated sectors and three countries, so it would be surprising if higher education elsewhere behaved differently. Measure your own market rather than assume ours transfers.

How many runs do I need for my own institution?

More than one, always. Answers vary between identical requests: three runs of one question in Spain returned eleven cited domains, of which four appeared exactly once. For a mention rate near 30%, a hundred runs gives you a range of roughly 21% to 39%, so aggregate across a set of questions and compare wide time windows rather than chasing weekly movement.

Can I see this for my own university?

Yes, and the first version costs you an afternoon rather than a subscription. Ask the five questions above, in your market’s language with the country fixed, several times each, and write down which institutions get named and which domains get cited. Then open each cited source and search it for your name. If your name is in a source the answer used and not in the answer, you have found the gap this study is about.