Key findings
- Universities hold roughly 5% of AI answers about their own certificate programs. Institutional share in the certificate tier ranged from 4.0% to 12.8%, compared with 33.7% to 57.3% for bachelor’s.
- Every one of Copilot’s 431 queries returned exactly two citations. 431 out of 431. Zero variance.
- Reddit is Gemini’s single most-cited domain, appearing in 40.6% of adult-learner queries with 203 citations, ahead of every university.
- BestColleges appears in 92.4% of Grok’s answers. US News, by comparison, is worth 1.6%.
- SNHU accounts for 48.6% of every institutional citation Gemini makes, and appears in 30.6% of all Gemini queries.
- Changing one verb changes the source. Feasibility phrasing lifted Reddit presence 2.1x to 9.4x across all four engines.
- Not one .edu domain appears across all four engines. The universal core is entirely made up of aggregators, forums, government, and associations.
Two questions about evening classes, two completely different results
We asked Copilot how night and weekend class schedules work at a community college. It cited two institutions: localcc.edu and statecc.edu. Then we asked Copilot how adult learners get through evening labs when they have kids. It cited reddit.com and parentingforums.com.
Same engine. Same tier. Same prospective student. Both questions are about evening classes.
What changed was what the question asked. The first asks how the thing works. The second asks whether a person can manage it. Copilot treated those as different enough to go looking in a different part of the internet, and it never came back to the .edu domains it had cited a moment earlier.
That is the pattern this study kept running into. AI engines do not route on topic. They route on what kind of question is being asked. Two people can ask about the same program, the same tier, and the same constraint, and get answers built from entirely different sources.
The consequence for institutions is direct. When an adult learner asks the question they care about most, whether they can actually do this while working full time, universities largely drop out of the answer, and other people’s experiences take their place.
We measured that effect across 1,931 queries and four AI engines. Queries built on feasibility language routed to peer forums at 2.1 to 9.4 times the rate of other phrasings, on every engine we tested.
Why this matters now
Adult learners are the growth segment in higher education, and their discovery behavior has moved. A prospective student researching a bachelor’s completion program no longer just goes to Google, searches for the degree they are interested in, then opens ten tabs and completes their research on the university’s owned website. They ask an AI engine and get an answer.
That answer is edited, not listed. It is a synthesized verdict assembled from a curated source set, which means the question is no longer whether you rank. It is whether you are in the set.
Institutions have spent two decades optimizing for process content: how to apply, what transfers, what it costs. That content still gets cited. Our data shows it is also the smallest and least decisive layer of what an adult learner actually asks.
- Study design: what we tested and what we cut
| Dimension | Count | What we tested |
|---|---|---|
| AI engines | 4 | ChatGPT-5.6 Sol, Gemini 3.6 Flash, Copilot Smart, Grok 4.5 |
| Education tiers | 5 | Certificates and bootcamps; associate and community college; bachelor’s completion; master’s and graduate certificate; doctoral and executive |
| Constraint categories | 9 | Transfer credit; military credit; cost and aid; scheduling; employer perception; ROI; accreditation; employer reimbursement; accelerated pacing |
| Total queries | 1,931 | ChatGPT-5.6 Sol (459), Gemini 3.6 Flash (500), Copilot Smart (431), Grok 4.5 (541) |
What we cut. Perplexity was in scope. Citation collection did not complete, so we excluded it entirely rather than report partial data.
Finding 1: The certificate collapse
Institutional share of AI citations drops off a cliff at the certificate tier.
| Tier | Institutional share |
|---|---|
| Bachelor’s completion | 33.7% – 57.3% |
| Doctoral / executive | 26.3% – 57.2% |
| Associate / community | 20.3% – 53.5% |
| Master’s / grad certificate | 20.7% – 38.2% |
| Certificates / bootcamps | 4.0% – 12.8% |
Three of the four engines landed within 1.6 percentage points of each other in the certificate tier. That kind of convergence across independently built systems is not noise.
Who fills the vacuum? Course Report and SwitchUp operate as a near-duopoly, appearing as a locked pair in 19 of 25 review-primary Copilot queries. On Gemini, Reddit takes 67% of certificate queries outright.
The business problem is the timing. Certificates and bootcamps are the fastest-growing revenue line in most institutional portfolios, and institutional visibility is close to zero exactly there. Employers, not schools, have always been the final arbiter of what a short credential is worth. What is new is that the arbitration now happens inside a machine-generated answer, assembled from sources the institution does not control and cannot see.
Finding 2: The verb decides the source
The cold open was not an anecdote. Intent routing keys on sentence construction rather than subject matter.
We isolated a set of feasibility verbs: balance, manage, handle, cope, burnout, juggle, justify. Queries built on those verbs routed to peer forums at sharply higher rates on every engine we tested.
| Engine | Reddit presence, feasibility verbs | All other queries | Lift |
|---|---|---|---|
| Copilot | 61.4% | 6.5% | 9.4x |
| ChatGPT | 43.8% | 11.9% | 3.7x |
| Grok | 66.0% | 25.8% | 2.6x |
| Gemini | 77.1% | 36.7% | 2.1x |
Four engines, same direction, n = 44 to 48 per engine.
The implication is a language mismatch rather than a content gap. Institutional pages are written in procedural grammar: what transfers, how to apply, what it costs. Adult learners ask experiential questions: can I actually do this while working full time. Our observation that sentence construction routes queries to different sources is consistent with work on intent taxonomies and speech-act-based retrieval (Lichtenegger et al., 2026).
Finding 3: Institutions own process, third parties own judgment
Rank the nine constraint categories by average institutional presence, and the rank order becomes the argument.
| Constraint | Institutional presence | Zone |
|---|---|---|
| Transfer credit | 83.5% | Process |
| Military credit | 75.0% | Process |
| Accelerated pacing | 69.2% | Process |
| Scheduling | 68.0% | Process |
| Accreditation | 65.0% | Process |
| Cost and aid | 62.5% | Process |
| ROI | 54.2% | Judgment |
| Employer reimbursement | 43.8% | Judgment |
| Employer perception | 37.2% | Judgment |
The descent is monotonic. Every step from process toward judgment costs institutional share, without a single reversal.
Employer perception is the clearest case. Ask “do employers respect this credential,” and you get Forbes on Copilot, Reddit on Gemini, Reddit on Grok, and a ranking blend on ChatGPT. Four engines, four different answers, and not one of them is the university that grants the credential.
Employer reimbursement is the starkest. On Copilot and Gemini, queries about the $5,250 employer tuition benefit returned zero institutional citations. Not few. Zero. The territory belongs to irs.gov and shrm.org. ChatGPT-5.6 Sol and Grok 4.5 split from that pattern and did cite institutions, and at n=4 this category is directional rather than settled. The two-engine result is still worth acting on.
Finding 4: Every engine has a fixed ceiling, and complexity buys nothing
No engine scales output with query complexity. Each sits at a fixed level and stays there.
| Engine | Mean citations | Distribution |
|---|---|---|
| Copilot | 2.00 | 431 of 431 at exactly 2. Zero variance. |
| Gemini | 3.01 | 496 of 500 at exactly 3. All four exceptions in the certificate tier. |
| ChatGPT | 4.83 | Range 2 to 8 |
| Grok | 8.46 | Range 6 to 12 |
Query length correlated with citation count at 0.13 on Grok and negative 0.03 on ChatGPT. Effectively nothing.
This inverts a core SEO instinct. In traditional search, a richer and more specific query opens more surface area. Here it opens none. On Copilot, the entire adult-learner category is a two-slot auction, and no amount of breadth, specificity, or complexity buys a third slot. Our volume-invariance result is consistent with prior work on how generative systems assemble and bound their source sets (Aggarwal et al., 2024).
Four engines, four different machines
A school optimized for one of these platforms can be invisible on another.
Copilot: the adjudicator
Two citations, always. Copilot is single-source-class 87% of the time, meaning it picks a lane and commits rather than blending. Its top pairs are BestColleges plus US News (67 occurrences) and Reddit plus Quora (51). Business press and government together account for 45.4% of its citations.
The strategic read: on Copilot, you are not competing for a slot in a blend. You are competing to have your entire source category selected.
Gemini: the triangulator
Three citations, and genuinely mixed source classes. Gemini cross-references before accepting institutional claims, which shows in the numbers: it has the lowest institutional share of any engine at 21.0%, and a .org lane worth 22.0%.
Reddit is Gemini’s number one domain at 203 citations across 40.6% of queries, ahead of every university in the study.
ChatGPT: the closed roster
The most institutional engine at 41.0% .edu, and the narrowest source universe by a factor of five. Only 92 unique domains across 2,219 citations. BestColleges appears in 56.4% of queries, US News in 24.0%.
The template lock is the finding worth sitting with: 100 distinct master’s-level questions produced only 15 distinct citation sets. Volume is not opportunity.
Grok: the monopolist
Highest citation volume in the study at 8.46 per query, and it has effectively outsourced the category to one publisher. BestColleges appears in 92.4% of all Grok queries. US News is worth 1.6%. Only 12 domains are unique to Grok.
These architectural differences track with emerging work on how generative retrieval systems differ in source selection (Grossman et al., 2026).
The concentration problem: who actually controls the answers
Roughly 13 domains determine what an AI engine says about higher education.
| Domain | Average presence |
|---|---|
| bestcolleges.com | 46.8% |
| reddit.com | 24.3% |
| usnews.com | 14.3% |
| bls.gov | 10.1% |
| studentaid.gov | 7.6% |
| coursereport.com | 7.0% |
Here is the number that should stop you: not one .edu domain appears across all four engines. The universal core is aggregators, forums, government, and associations. Every one of them.
Named-brand concentration is just as tight. On Gemini, three institutions account for 72.1% of all institutional citations: SNHU at 48.6%, WGU at 14.0%, Harvard at 9.5%. SNHU alone shows up in 30.6% of every Gemini query in the study. ChatGPT concentrates differently but concentrates all the same: WGU 16.2%, ASU 11.4%, SNHU 10.2%.
The takeaway is not “protect your domain authority.” It is that a regional public university competing for Gemini’s institutional slot is competing against SNHU for a category SNHU already half-owns. That is a different problem requiring a different strategy. The pattern echoes findings that language models reflect and amplify existing citation bias rather than correcting for it (Algaba et al., 2025).
The exception: how WGU owns one question across two engines
One school in this study owns a question.
WGU appears in 10 of 16 accelerated-pacing queries on ChatGPT and 7 of 16 on Gemini. Two independently built platforms, both of which learned the association between competency-based pacing and WGU by name.
The lesson has a hard edge. WGU did not win that association with content marketing. It won because the institution is structurally built around competency-based progression, and the engines learned an association that happens to be true. Entity-level ownership of a constraint is achievable. It comes from the product, not the copy.
The Adult-Learner Intent Model
Four query types, four different source owners. Use it to audit your own content.
| Query type | Linguistic marker | Who wins | Example |
|---|---|---|---|
| Procedural | does X count, how do I apply | Institutions, 67% to 100% | “Do my military credits transfer?” |
| Comparative | best, top, versus | Aggregators | “Best online bachelor’s completion programs” |
| Evaluative | worth it, respected, ROI | Business press and forums | “Is this degree worth the cost?” |
| Experiential | balance, manage, cope, burnout | Reddit and Quora, 2.1x to 9.4x lift | “How do I balance this with full-time work?” |
Adult learners ask disproportionately in the evaluative and experiential registers. Institutional content is written almost entirely for the procedural one.
Seven recommendations for AI optimizations tied to data for the higher education industry
1. Rewrite for the experiential register
The evidence: feasibility verbs lift Reddit presence 2.1x to 9.4x, on four engines out of four.
Publish day-in-the-life content, weekly hour budgets, and honest answers to questions like what happens when work travel hits finals week. Here is the test: if every page you own would only ever be retrieved by a procedural query, you are invisible to the largest share of adult-learner intent.
2. Treat Reddit as owned-channel infrastructure
The evidence: Reddit is Gemini’s number one domain at 40.6% of queries, holds 92% presence on employer-perception queries, and 66% to 77% on experiential ones.
This means genuine, non-promotional participation in the subreddits where your programs get discussed. Accuracy stewardship, not astroturfing. Correcting a wrong tuition figure is the job. Posting marketing copy will get you removed and deserves to.
3. Non-degree visibility is a publisher-relations problem
The evidence: 4.0% to 5.6% institutional share in the certificate tier across three engines.
You cannot content-market into this tier. Course Report, SwitchUp, Career Karma, and BestColleges are the distribution channels. Placement on those platforms is the work, and it belongs to whoever owns partnerships rather than whoever owns the blog.
4. Defend the process layer, because it is the only thing you still own
The evidence: transfer credit 83.5%, military credit 75.0%, accreditation 65.0%.
Machine-readable transfer tables. PLA, CLEP, and DSST equivalency documentation. Accreditation in structured data. Transparent tuition calculators. This layer is winnable and currently held. Losing it would be an unforced error.
5. Own one constraint by name, the way WGU owns pacing
The evidence: WGU in 10 of 16 and 7 of 16 accelerated-pacing queries on two independent platforms.
Identify the single friction you genuinely solve better than anyone, and build an entity-level association around it. The warning matters more than the recommendation: this worked for WGU because the operating model matches the claim. If the claim isn’t structurally true, the engines will not learn it, and your competitors will correct the record for you.
6. Allocate by platform architecture, not equally
The evidence: four engines with four different citation economies.
On Copilot, you are in a two-slot auction, so win the source category (business press and government are 45.4% of its citations). On Gemini, your claims will be checked against Reddit, so they need corroboration you do not control. On ChatGPT, BestColleges at 56.4% and US News at 24.0% are the gatekeepers. On Grok, BestColleges at 92.4% is nearly the whole game, and US News at 1.6% is not worth a line item.
Equal allocation across four engines is the one strategy the data rules out.
7. Reclaim employer reimbursement
The evidence: zero institutional citations on Copilot and Gemini for $5,250 employer tuition-benefit queries (n=4, directional).
Publish the mechanics. Employer billing processes, clawback-clause guidance, HR-facing documentation. This is high-intent, high-conversion territory with near-zero institutional competition on two engines. It is the cheapest opportunity in this entire study.
The three trust layers in an AI answer: institutions own the smallest one
Process, judgment, and experience. Three layers, three different owners, and the split is far from even.
- Process is institution-owned: What transfers, how to apply, what it costs. You hold this at 62% to 84%.
- Judgment is third-party-owned: Is it worth it, do employers respect it, what is the ROI. Ranking publishers hold this, led by BestColleges at 46.8% average presence and US News at 14.3%. For ROI, BLS holds this; bls.gov appeared in 10.1% of every query in this study.
- Experience is peer-owned: Can I actually do this? Reddit holds this at up to 77%.
Higher education is optimized for the first layer, which is the smallest and least decisive of the three. And the gap widens as the credential gets shorter, which is exactly where enrollment growth is heading.
The logical extension of this data is not that institutions should try harder to control the answer. It is that the answer is assembled somewhere else, from sources institutions do not own, and the schools that do well over the next five years will be the ones that stop trying to control it and start participating where it gets assembled.
Methodology, limitations, and how to cite this study
What we ran
We built a query set spanning five education tiers and nine adult-learner constraint categories, then executed it across four AI answer engines: ChatGPT (459 queries), Gemini (500), Copilot (431), and Grok (541). Total: 1,931 queries.
For every response, we captured the full citation set, the domain of each cited source, its position in the citation order, and the tier and constraint category of the originating query.
How we classified sources
Every cited domain was assigned to one class. Institutional (.edu domains and official institution properties), review aggregator (BestColleges, US News, Course Report, SwitchUp, Career Karma), community (Reddit, Quora, industry forums), government (bls.gov, studentaid.gov, irs.gov), association and nonprofit (.org), and business press and independent editorial (Forbes, trade publications, niche blogs).
Institutional share is the percentage of citations in a given cut that fell into the institutional class.
How we classified intent
Queries were coded into four registers based on linguistic construction rather than subject matter: procedural, comparative, evaluative, and experiential. The experiential cut keys on a defined feasibility verb set: balance, manage, handle, cope, burnout, juggle, justify. The comparison in Finding 2 is that verb set against all other query constructions within the same engine, at n = 44 to 48 per engine.
Limitations
Perplexity was excluded. It was in the original scope. Citation collection did not complete, so we removed it entirely rather than report partial data.
Named-institution findings rest on two engines. Copilot and Grok record institutions as generic placeholders rather than named entities. Every claim in this study involving SNHU, WGU, ASU, or Harvard comes from Gemini and ChatGPT only. Grok shares throughout are directional.
Three categories are small-n. Military credit (n=4), employer reimbursement (n=4), and transfer credit (n=9). We report them because the direction is consistent and the practical implication is worth acting on, and we label them directional every time they appear. Do not treat them as settled.
These systems are non-deterministic. The same query can return a different set of citations on a different day. Our findings describe distributions across nearly two thousand queries, not guarantees about any individual response. The strongest results here, like Copilot’s 431 out of 431 and the monotonic constraint ranking, are the ones least sensitive to that variance.
This is a point-in-time snapshot. Retrieval behavior changes without announcement. Treat the specific percentages as a baseline to re-measure against, not a fixed property of these engines.
Higher education only. Nothing here should be extrapolated to other verticals. The source ecology is different everywhere else.
References
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). GEO: Generative engine optimization. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 5–16). Association for Computing Machinery. https://doi.org/10.1145/3637528.3671900
Algaba, A., Mazijn, C., Holst, V., Tori, F., Wenmackers, S., & Ginis, V. (2025). Large language models reflect human citation patterns with a heightened citation bias. In Findings of the Association for Computational Linguistics: NAACL 2025 (pp. 6844–6879). Association for Computational Linguistics. https://doi.org/10.18653/v1/2025.findings-naacl.381
Lichtenegger, E., Urman, A., & Hannák, A. (2026). A new taxonomy of web search: A user-centered framework for search intent in the AI era. In Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (Article 692, pp. 1–16). Association for Computing Machinery. https://doi.org/10.1145/3772318.3791050