GEO for Education and Training Providers: Getting Recommended During Admissions Season
Summary: Education is a textbook YMYL field, and AI surfaces treat it more conservatively than most industries — they prefer pages with a clear legal entity, verifiable numbers and accreditation backing, and they skip adjectives. So the priority is not publishing more articles but cleaning up three content layers: institution and programme entities, time-stamped admissions policies, and employment and accreditation figures with a stated basis.
Key Takeaways
- Establish the entity before the content. An institution, a faculty and a programme are entities at different levels; the
CollegeOrUniversitytype exists because "this school" and "this programme" are not the same thing to a machine. - An admissions policy needs exactly one authoritative version. When the same policy appears on five pages with five different dates, an AI answer skips it — not because it cannot read, but because it cannot tell which is true.
- Employment data must be checkable. A placement rate without a basis, a year and a sample size is noise in a generated answer.
- Admissions season is a timeliness battlefield. Policy releases, application deadlines and transfer windows go from information to misinformation within a week.
- Education is YMYL. Google's quality guidelines raise the bar for content that shapes major user decisions; bylines, credentials and verifiable sources are not optional.
1. Visibility has three layers, and they break separately
"AI does not show our school" is usually three problems stacked.
Entity layer: does the machine know who you are? Institution name, campus, faculty, programme, degree level, supervising authority — these are entity attributes. Without them, no content attaches to your name. CollegeOrUniversity exists for this layer; it is a more specific subtype of Organization, not a replacement, and it carries properties the general type cannot express.
Fact layer: can the machine confirm your numbers? Intake quota, tuition band, acceptance ratio, placement rate, faculty count — checkable facts. With a basis, a year and a source, they get cited; with only "our placement rate reaches 98%", they do not.
Timeliness layer: does the machine know which version is current? Policies change annually and prospectuses are replaced annually. If the site carries 2024, 2025 and 2026 prospectuses with no explicit current marker, an AI answer either picks one at random or takes none.
Each layer is fixed differently — markup and consistency for entities, data governance for facts, version management for timeliness. Discussing them together as "content optimization" reliably produces more articles nobody cites. Diagnose yourself with three queries: search your institution name and see who the result points to (entity layer); search "how is the placement for [programme]" and see whether the answer quotes your figures (fact layer); search "[year] admissions prospectus" and see which version comes back (timeliness layer).
Want to know how much of the ChatGPT and Perplexity answers your brand owns?
Explore TENGENCE GEO2. Entity layer: institution, faculty and programme are separate
A workable test first: search your institution's name. If the result is a third-party encyclopedia page rather than your own site, the entity layer has not taken hold.
Three markup tasks:
| Level | Type | Key properties | Common error |
|---|---|---|---|
| Institution | CollegeOrUniversity |
Official full name, sameAs (site, admissions site, official social profiles), address, supervising authority |
Mixing short and full names; sameAs holding only the main site |
| Faculty | Organization + parentOrganization |
Parent institution, contact details, programme list | Faculty and institution pages sharing one markup block, parent link missing |
| Programme | Course or EducationalOccupationalProgram |
Duration, degree awarded, learning outcomes, owning faculty | Programme page written as marketing copy with no duration or degree fields |
The most damaging issue is name inconsistency. An institution using a short name in site titles, the full name on the admissions page and a different English rendering on the English page presents three entities to a machine. We maintain a naming table: one permitted name per language and per channel, with acceptable aliases in alternateName. It sounds tedious; it is the precondition for entity merging.
One more note: claiming a knowledge panel works for institutions too. After claiming you can submit factual corrections, but you cannot influence whether the panel appears. Institutions have a specific hazard — same-named schools, especially those containing place names. One wrong sameAs entry can attach another school's social profile to your panel.
3. Fact layer: writing employment data that gets cited
Employment data is the most citable content in education and also the most easily judged untrustworthy.
Our standard requires four elements: basis (who is counted — graduates or contracted students), year (which cohort), sample size (how many were counted), source (career office, third-party report, public statistic). With all four, an AI answer can treat it as fact. Missing any one and it is more likely to be skipped as marketing language.
The negative example is typical: "Our graduates achieve a 98.6% placement rate and are highly regarded by employers." No year, no basis, no sample. The positive example reads: "Of 2,340 bachelor's graduates in the 2025 cohort, 91.2% had signed employment contracts as of 31 August 2025, per the annual Graduate Employment Quality Report published by the university careers centre in October 2025." That is a citable passage.
There is a counter-intuitive conclusion here: higher is not more citable. A "100% placement" figure is often treated as an implausibility signal in generated answers, because it cannot be reconciled with other public data. The reason is simple — a perfect score is hard to cross-check against intake quotas, education-authority statistics and third-party employment reports. When it does not reconcile, the system does not audit which figure is wrong; it simply stops treating your line as a stable source. Credibility outranks polish.
How to verify: take a group of institution pages, record citation behaviour before the four elements are added, then re-run the same prompts two weeks after. The elements are degree level, location, programmes offered with duration, and admissions contact. Differences concentrate on specific prompts ("where is it located", "which programmes") and are usually invisible on broad prompts.
4. Timeliness layer: admissions calendar and version control
Admissions content has a very short life. Mainland China undergraduate admissions, for example, run roughly four months from prospectus release to the end of enrolment, with three information types carrying the highest time sensitivity: policy changes, application deadlines and transfer windows. A week stale equals misinformation.
We schedule clients on a four-phase cadence:
- T-60 days: entity and policy skeleton. Update institution/faculty/programme markup, publish the year's admissions policy overview, and mark older versions explicitly as historical.
- T-30 days: programmes and data. One page per admitting programme with duration, degree, outcomes, prior score bands and placement destinations.
- T-0 to T+30: high-frequency updates. Application, examination, results release and preference submission each get a page, updated with
dateModifiedat publication. - T+90 days: review and archive. Convert this season's pages into "prior-year reference", keep the URLs and data, and link back from the new pages.
Step 4 is routinely skipped and is precisely what prevents versions from contradicting each other. An old page that is neither deleted, archived nor marked gets retrieved alongside the new one and produces self-contradictory signals.
Archiving has a second benefit: it separates this year's data from last year's. When a generated answer cites prior-year data it usually carries a year qualifier; if your page merges three years into one table, a machine can easily pick the wrong row. One table per year, cohort written in the header, pages linked in sequence — rather than one table edited every year.
5. Four misconceptions and the boundaries
Misconception one: pasting the whole prospectus as an article. A prospectus is a legal document; an AI answer struggles to extract citable passages from it. Split it into policy highlights, a timeline of milestones and a Q&A, keeping the legal text as the source page.
Misconception two: publishing Chinese only. Recruiting international students without English pages means English-language questions cannot reach you at all. English pages are not translations; they are rewritten around English query habits — "tuition fee", "admission requirements" and "scholarship" carry intents that differ from their Chinese equivalents.
Misconception three: treating ranking as the only selling point. A ranking comes from a third party; restating it adds no factual increment. What adds increment is ranking plus the specific indicator, the year and a source link.
Misconception four: substituting student testimonials for data. Testimonials are opinions and cannot replace checkable figures. Our ratio: at least one number per claim, at most two sentences of testimony.
Boundaries matter too: admissions policy is issued by the authority; an institution can change presentation, not policy. Platforms' conservative treatment of YMYL content will not disappear with optimization; what you can do is make "checkable" cheap for them.
FAQ
Q: A vocational training provider has no degree-granting authority. Should it still use CollegeOrUniversity?
Do not force it. Use EducationalOrganization or Organization for the entity and Course for programmes; misusing the type misleads machines about your degree level.
Q: Can placement rates live only inside an image?
No. Numbers inside images are unreadable to machines and useless for citation. Put them in body text with a source link.
Q: Should an admissions policy get a new URL each year?
Yes. Mark the old page clearly as "applies to 2025 admissions, expired" and link it to the new one. Reusing one URL and overwriting it makes historical citations point at content that no longer matches.
Q: How should multi-campus institutions handle entities?
Each campus gets its own address and contact details, expressed with subOrganization or a separate Place. Never cram all campuses into one address field.
Q: How long before AI citations appear in education?
Count in quarters. Entity and markup fixes usually move knowledge panels or brand-query results within 6 to 12 weeks; content citations wait for the next index and evaluation cycle, so actions completed during one admissions season often show fully in the next.
Related Reading
- Multi-regional site architecture for cross-border recruitment
- Healthcare site search and intent recognition
- The enterprise GEO playbook
Data Sources
- CollegeOrUniversity — schema.org type definition — the properties that distinguish an institution from a generic organisation.
- Search Quality Rater Guidelines — Google — why YMYL content is held to a higher evidentiary standard.
- About knowledge panels — Google — what an entity owner can and cannot change.