By the time a family fills in an enquiry form, the shortlist already exists. It was built earlier, in a search result or an assistant's answer, while a parent typed a question that mentioned a budget, a grade and a country. A university cannot sit inside that answer and argue with it. It can influence exactly two things: whether it is a source the assistant is able to quote, and whether it is the page the assistant links to.
That is a narrower job than most international offices expect, and a more practical one. This post covers what happens when a family asks a real question, where the answer comes from, why one international students page cannot serve five markets, a checklist an office can run this month, and what you can honestly measure. It is also the reason our international marketing and admissions program starts with pages and facts rather than campaigns. Every figure below carries its source and date, as of September 27, 2026. Where the mechanism is not publicly documented, this post says so instead of guessing.
What happens when a family asks a real question
Admissions prompts are long and full of constraints, because the people typing them are trying to rule options out. These are the shapes they take:
- "Is a masters in data science in Ireland worth it for a 62 percent BTech graduate with no work experience"
- "UK or Canada for an MSc in nursing with 6.5 IELTS, and which one lets me work after"
- "How much money do we need to show for a Canadian study permit in 2027"
- "Can my daughter work part time on a German student visa, and how many hours"
- "Is [university] recognized in India, and will the degree be accepted for a government job"
- "Cheapest UK universities for an MSc in computer science with a January 2027 intake"
The assistant answers in a paragraph or two, usually with a handful of options named and a few links attached. That answer does the work a counselor's first meeting used to do: it decides which three to five institutions are worth a form.
How common is this? Two dated surveys, both self reported, point the same way and disagree about scale. IDP's Emerging Futures research, fieldwork in July and August 2025 with 7,900 current and prospective students, found 54% intended to use AI to help decide their institution and 53% their subject, up from 35% and 38% in August 2024, as reported by ICEF Monitor on October 16, 2025. Intent is not behavior. INTO University Partnerships asked 1,600 newly enrolled students in the US and UK in September 2025 what they had actually done, and 17% said they used AI as part of their initial search, with higher rates in some markets: around 30% in South Korea and the Philippines, 28% in Taiwan, 25% in Vietnam, 22% in Japan and 21% in mainland China, reported by ICEF Monitor on January 8, 2026.
Read those two numbers together and the useful conclusion is not a percentage. It is a sequence. The assistant is the first filter for a growing minority and the starting point for many more, and then families move to sources they trust: the same IDP research found more than half still rely on university websites and more than a third name counselors among their most trusted sources. So the assistant rarely closes anything. It decides whose website gets opened.
Where those answers come from
The engines document their plumbing, not their judgment. That distinction matters, because most advice sold to universities right now pretends otherwise.
Google's own documentation on AI features says AI Overviews and AI Mode use a query fan out technique, issuing multiple related searches across subtopics and data sources to build a response, which lets them show a wider set of links than a classic results page. The same page states there are no additional requirements to appear, no special schema.org structured data to add and no new machine readable or AI text files to create. OpenAI documents four user agents: OAI-SearchBot to surface sites in ChatGPT search, GPTBot for training crawls, ChatGPT-User for actions a person triggers, and OAI-AdsBot for ads safety checks. Its documentation notes that a site can allow OAI-SearchBot while disallowing GPTBot, and that a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though it can still appear as a navigational link. Perplexity documents PerplexityBot, used to surface and link websites in its results and not for model training, and Perplexity-User for user initiated visits.
What none of them publishes is how a source is chosen or ordered once it is available. Anyone who tells you they know how ChatGPT ranks universities is describing a test, not a documented mechanism. Tests are still worth running, and the honest version of the claim is: these pages were quotable, and they got quoted.
What can be observed is the mix of page types that show up in answers about studying abroad, and how unstable that mix is. Semrush's study of most cited domains, published November 10, 2025, covered 230,000 prompts and more than 100 million citations across ChatGPT search, Google AI Mode and Perplexity between July 14 and October 12, 2025. It found Reddit, Wikipedia, LinkedIn, YouTube and Forbes among the most cited domains overall, and recorded Reddit's share of ChatGPT responses falling from close to 60% in early August to around 10% by mid September. That is one vendor's prompt panel over 13 weeks, not a law of nature, and the volatility is the finding worth keeping: community and third party pages carry real weight, and how much weight changes inside a quarter.
For an international office, the practical translation is a page inventory. If the answer to a family's question currently lives somewhere else, and you have no page that could carry it, you are not in the running to be quoted.
| What the family asks | Where the answer usually lives now | Page you need to own |
|---|---|---|
| Total cost for one year, including living costs | Agent blogs and aggregator profiles, often out of date | A fee page per course level and market, in text, dated |
| Whether their qualification and grades qualify | Forum threads and consultant WhatsApp messages | Entry requirements by country qualification, stated plainly |
| Visa, funding proof and work rules | Government pages, which is correct and where you should point | A market page that states the rule, links the official source and says what you require |
| Whether the degree is recognized at home | Regulator lists, news, discussion threads | A recognition and accreditation page naming each body and linking it |
| What happens after graduation | News coverage of policy changes | An outcomes page with a stated method and date, and no promises |
| Whether students like them are happy there | Reddit, YouTube, alumni groups | Nothing you can own, and everything you should be present in honestly |
The last row is the one most offices skip. You cannot publish your way into a forum thread, but staff and alumni answering real questions under their own names is a legitimate, slow form of corroboration.
Why one international students page cannot answer five markets
Most university sites have a single international page, written to be true everywhere, which makes it specific nowhere. The reason that fails is not tone. It is that the fact deciding a family's question is country specific and dated, and it moves:
The pages and the questions they close
- United Kingdom. Graduate route permission drops from two years to 18 months for applications made on or after January 1, 2027, with PhD and doctoral graduates unaffected, per UKCISA's student update of October 14, 2025.
- Canada. A single applicant studying outside Quebec must show CAD 23,448 a year, excluding tuition and transportation, for applications received on or after September 1, 2026, per IRCC.
- United States. The rule replacing duration of status with a fixed admission period was blocked by a federal judge on September 14, 2026, the day before it was due to take effect, as reported by the Harvard International Office.
A family in Hyderabad and a family in Lagos are not asking the same question, and neither of them is asking your generic question. A market page is what lets you be specific without lying to someone else: one page per source market and study level, holding the entry requirements for that country's qualifications, the funding proof that market's applicants must show, the intake and deadline that applies, the equivalence of local grades, an explicit link to the official rule, and an "as of" date with a named owner.
This is also where the previous cycle's lesson applies. Our post on the 2027 intake reset for Indian students traced how quickly three countries changed rules in a single admissions cycle, and the pages that survived that were the dated ones.
What an international office can actually control, this month
None of this requires a new platform. It requires someone with edit rights and a week.
What has to survive the handoffs
- Get the facts out of PDFs and into HTML. Fees, entry requirements, intakes and deadlines belong in page text, not in a prospectus download or a JavaScript only fee calculator. It is not publicly documented whether each assistant's fetcher executes JavaScript, so write the facts into server rendered HTML and treat anything else as decoration.
- State entry requirements plainly, per qualification. "Equivalent to a UK 2:1" is not usable by a family holding a four year BTech with a percentage score. Give the local qualification, the score you need and the English test scores, in a list a person can check themselves.
- Put an answer paragraph near the top of every course and market page. Forty to sixty words that answer the page's question directly, before the marketing. Google is explicit that no special optimization is required for AI features, so treat this as editorial discipline rather than a trick: a page that states its answer in one paragraph is easier to quote correctly, and easier for a parent to read.
- Make one fact true everywhere. Audit the fee, the intake date and the entry requirement across your course pages, your international pages, your agent portal and your aggregator profiles. Contradictions are how a wrong number becomes the quoted number, and aggregators outrank you on some of these queries.
- Check that nothing is blocking you. Confirm robots.txt allows Googlebot, OAI-SearchBot and PerplexityBot. Check the Search generative AI control in Search Console settings is set to include, since it is a property level switch that can exclude your content from AI Overviews and AI Mode. Check your templates for nosnippet, max-snippet and data-nosnippet, which Google's robots meta tag documentation describes as also limiting how content is used as a direct input for AI Overviews and AI Mode.
- Earn third party corroboration. Regulator and accreditation listings that match your claims word for word, aggregator profiles with your current numbers, alumni and staff answering questions in public under their own names, and news coverage of things you actually did. Our guide to AEO for higher education covers the program schema and recognition pages in more detail.
- Name an owner and a review date. Every page in the set gets one person and a date it must be checked. This is the control that makes the other six survive a change of rules.
How to measure it, without overclaiming
Reporting on AI surfaces is improving and still partial. Know exactly what each tool gives you.
Search Console shows impressions, not questions. Google introduced a generative AI performance report in June 2026, and its help documentation states the report covers AI Overviews and AI Mode, counts an impression as a time links to your site were shown in a generative AI feature, treats two results from the same site in one feature as a single impression, excludes Search Labs experiments, and was extended to all websites worldwide as of August 31, 2026. It reports impressions with page, country and date, and it does not include clicks or queries.
Clicks are counted, but not separable. The performance report documentation states that clicking a link to an external page in an AI Overview or in AI Mode counts as a click, that an AI Overview occupies a single position with all its links assigned that position, and that position in AI Mode follows the same methodology as a results page. Those clicks sit inside the Web search type totals, so you cannot filter an AI only view of them.
GA4 separates assistants from Google's own AI. Google Analytics added an AI Assistant channel on May 13, 2026, described in the channel group documentation as users arriving from sources like ChatGPT, Gemini, Deepseek, Copilot or Grok, with medium set to ai-assistant. The same documentation defines Organic Search as including Google's AI Overviews and AI Mode, so Google's AI surfaces will never appear in the AI Assistant channel. Because the channel depends on a referrer matching a list, a visit that arrives with no referrer will not be classified there, which is why your assistant traffic is almost certainly larger than the channel says.
Keep a prompt log, because nobody will give you one. Write 40 to 60 questions in the words your families use, split by market and study level, and test them monthly across the assistants you care about. Record three things per question: whether you appear, whether the facts quoted about you are correct, and which page was linked. Wrong facts are the actionable finding, because they usually trace back to a page of yours or an aggregator profile you can fix.
An impression in an AI feature is not a shortlist place, and a shortlist place is not an application. Report them as three separate things, or you will end up defending a number nobody can connect to an enrollment.
Then close the loop on your own side. Ask one question on the enquiry form about where they first heard of you, keep the wording stable for a year, and count enquiries, applications and deposits by market and program against the last cycle. Imperfect self reported attribution that is consistent beats a dashboard that changes definition every quarter.
Reporting may widen. On June 3, 2026 the UK Competition and Markets Authority imposed a publisher conduct requirement on Google that obliges it to give publishers effective controls over the use of their search content in generative AI, to explain how that content is used, and to provide clear and detailed metrics on user engagement with their content in search generative AI features. What that produces, and when, is not yet something to plan against.
What this changes about consultant and agent relationships
Briefly, because it changes less than people claim and earlier than they expect. The assistant now answers before the consultant does, so families arrive at a counseling call with a shortlist, a rough cost figure and two or three half correct facts already in hand. The consultant's job shifts from introducing options to correcting a picture, and the corrections they make will come from whatever they have to hand.
That makes two things operational. Your dated pages become the reference both sides quote, which only works if a consultant can find the fee and the entry requirement in ten seconds without logging in. And your consultant brief becomes a same day job: when a rule changes, the page and the brief move together, or your network will keep selling last cycle's answer. The operating model behind that, including how a governed consultant network and a direct admissions desk work together, is set out on the program page linked at the top of this post.
Where Lemniscate fits
Lemniscate Growth's higher education practice runs the unglamorous half of this: the page inventory, the market pages, the fact audit across your site and your aggregator profiles, the prompt log, and the measurement that tells you which of it moved applications. If you want to see where your university currently stands, book a university strategy session and we will walk through the questions your source markets are asking, what assistants currently say about you and which pages are missing.
This post summarizes documentation, survey research and government rules as of September 27, 2026. It is not immigration advice, and immigration rules change: check the official source for each market before you publish a page that quotes one.
