In the previous article I described the research I ran to see what happens when a family asks an AI to recommend a school for their child. Some of what the AI does in between is processing you cannot see. What you can see is the pattern in its answers, and below are some thoughts on what you might do with that pattern at your school.
First, a recap of how an AI search differs from a Google search. The short version: with AI, the parent describes the child and the AI chooses the search words, and a school can be easy to find by name and still be missing when a family describes a child it would serve well.
A lot will be written in the next year or two about chasing AI to get mentioned more often. My curiosity and my research pulled on one thread. In this article I am not going to focus on the question “What can you tell me about School X?” I am going to focus on “What schools in my area are good for my child?” That is a question a parent can ask an AI and could never ask a Google search box.
In full disclosure, the research behind these recommendations was done in October 2026, and I have not been back to retest. The technology behind these assistants, and the way they search and answer, is changing quickly. What I learned is enough to make some strong working assumptions, and for each one I have tried to say what my test showed and what it did not.
I ran two kinds of tests, and it is important to know that they did not always agree. The first was by hand: I typed each question into each AI assistant the way a parent would and captured what came back. The second was automated: I set up the same questions to run through the developer versions of the same assistants. I point this out because today a parent asks these questions by typing them into a free app. Tools that ask on a family’s behalf, or that offer to test your school’s AI profile for you, are already appearing, and those tools mostly run through the developer versions. This is one example of how complex all of this is becoming.
What follows is one test, and five things the count from it points you to. Between them they cover three points in a parent’s search with AI:
- Whether your school comes up at all
- What the AI tells the family about your school once it does
- Whether the family arrives, and how you would know
First, run the test: describe the child, ten times, with and without your word.
The natural first step is to ask the AI what it knows about your school. That is the mirror test I suggested in a post in July, and it is still worth doing. What it cannot tell you is whether your school gets into the conversation when the parent asking does not yet know your name.
So this test starts from the child. You describe, in a parent’s words, a child your school serves well, and you ask the AI what kind of school would help and which ones to look at. The answer generally comes in two parts, and you want both. The description of what that child needs tells you how the AI sees the family you serve best. The list tells you whether you are on it. Then you ask again, ten times, and count.
Here is how I would run it.
Step 1. Write the parent’s description. Three or four sentences, the way a parent talks on a tour rather than the way a school writes a viewbook. Pick a child your school serves well, a real family from this year’s inquiries retold in the parent’s words, not one the team writes from the viewbook. Give a grade and a town, and say what the parent is hoping for. Leave out your kind of school: no “Episcopal,” no “Montessori,” no “classical,” and no “private” or “independent” either. The whole test depends on that omission, because the parent who most needs to find you is the one who does not yet know your word. End with the question: What kind of school would help, and which ones should we look at?
Step 2. Open a chat the way a parent would. Use ChatGPT in a private browser window, logged out, so the assistant has no memory of you. Your own account may know where you work, and it remembers earlier chats, so answers from it tell you about you rather than about a parent. If you would rather stay signed in, ChatGPT’s temporary chat does the same job. Either way, open a new chat for every run.
Step 3. Ask, and write down five things. Paste the description, read the whole answer, and record these before you close the chat. A spreadsheet with five columns is enough.
| Write down | Why it matters |
|---|---|
| Were we named? Yes or no, and where on the list | This is the count, and the only column you will add up |
| Which schools were listed, in order | Who the AI puts beside you for the child you serve best. The second of the five things below is built from this list |
| How the AI described what the child needs, and the kinds of school it suggested, in its own words | Whether it described you, and what it called you. The first of the five is built from this |
| Whether your website appears in the sources under the answer | Whether you were reached at all. Being read and not named is a different problem from not being read |
| A one-line observation of what you saw | The sentence that will help you see the pattern later |
One note on the sources. A site cited in one run of ten is noise. The ones that keep coming back are the pages the AI built its answer from, and those are the pages to go read.
Then close the chat.
Step 4. Repeat until you have ten. Same words, ten fresh chats. It may feel excessive by the third run. In my test, the same question asked ten times named the same school anywhere from 0 to 10 times, and only the clearest cases held steady. One run is a coin flip. What you need is a pattern.
Step 5. Add your word and run ten more. Change one thing. The last clause becomes “and which Episcopal schools should we look at,” or whatever your kind of school is called. Everything else stays the same. At the two schools in my test where the word mattered most, that one change moved the count from 0 of 10 to 10 of 10.
Step 6. Read the two counts together. You now have a rate, not a verdict. “Named in 3 of 10 without our word, 9 of 10 with it” is a finding your head and your board can act on, and “the AI doesn’t know we exist” is not. Four results are possible, and each points somewhere different.
- Named most of the time, with or without the word. The AI connects the child you serve to your school on its own. Your attention goes to what families hear once you are named, which is 3 and 4 below.
- Rarely named without the word, usually named with it. Families who already know your kind of school will find you. Families who describe the child and not the category will not. The work is translation, which is 1 below.
- Rarely named either way, and your site is not in the sources. The AI’s searches are not reaching your pages. Look at the search phrases it shows and the kinds of school it suggested. That is the language it went looking in, and your pages were not written in it. The work is the same translation, 1 below.
- Rarely named either way, and your site is in the sources. It read you and chose other schools. Twice in my test, ChatGPT read a school’s own website before answering and still left the school off the list. Read what it said about the schools it chose, and compare it with what your pages say about the same child. My test cannot tell you what tips that choice.
Your counts may land in the middle. In my test two schools were named ten times out of ten without their word, two were almost never named, and the other four fell between. A middle result is still the first real number anyone at your school has had for this.
If you want to go deeper, this test pairs with the one I suggested after NACAC, asking the AI what kind of family would not be happy at your school. One shows how the internet describes you. This one shows what the AI thinks the family you serve is looking for, and whether that search leads to you.
The test is the investment. Each of the four results sends you to a different place among the five things that follow, and nobody needs all five at once. The count tells you which one earns the next block of your time.
What the count points you to
1. Explain your kind of school in the words families use.
If the first test shows the AI describing your school and calling it by a different name, the gap is in the translation between how families talk about their children and how the AI files you. That gap is a messaging job, and it is one admissions offices are good at.
In my test, the AI never once offered a faith-based or boarding school as a kind of school to consider, at any of the 5 schools where that was the kind, when the parent described the child and named no label. The Catholic high school in my test was found anyway, 10 of 10 times, because the parent’s wish for a son who would “show up for others” is the thing that school is already known for. The Catholic elementary school was not, because “heart and character” at kindergarten belongs, in the AI’s memory, to progressive and Montessori schools. The difference between the two is whether the school’s distinctive practice is already attached, in the public record, to the words a parent would use.
So, for each kind of family you serve well, write the paragraph that connects their concern to your concrete practice. What advisory actually means. What “structure” looks like in your middle school. What faith participation involves, and who is welcome. How a quiet, capable kid gets noticed in your classrooms. These are the pages a parent reads on a tour night anyway. They are also the only kind of page that could answer a search the AI runs in its own words.
I want to be straight about the limit. I did not test whether writing those pages raises how often an AI names a school, and the studies I have read did not test it either. What they tested was unusual wording for its own sake, and that showed no lift. Explaining your category in family language is not that. It is traditional admissions messaging, worth doing for the tour night alone. The honest thing to say is that its effect on AI visibility is unproven.
2. Write down who the AI lists instead.
When the AI omits a school from a shortlist, it lists other schools in its place, and those schools are the real alternatives in that family’s mind. In my test they were often not the schools a private school thinks of as competitors. When the parent described a child and named no kind of school, the AI named public, charter or magnet schools in 23 of 32 answers, alongside private schools of other kinds.
So keep the list. Every time you run the test, record the schools that appear, the ones that appear first, and the reasons the AI gives. After ten runs you will have something no competitor analysis gave you before: the set of options the AI is actually putting beside you, for the child you serve best. Bring it to the admissions team. The trade-offs a family is weighing when they walk in, a magnet program against your tuition, a charter’s lottery against your advisory, are the conversation your tour guides are already having. Now you know which ones.
A caution. I did not judge whether the schools listed instead were good fits for the child, so I cannot say the AI’s list was worse than one with your school on it. The list is information about the family’s choice set, not a verdict on the AI.
3. Publish a complete, dated cost page.
The best news in my test was about price. Asked what a school costs, both AI assistants went to the school’s own page in 16 of 16 answers and got the tuition right in 15. The school’s own website is the source the AI trusts for the number.
That number was the easy part. The family’s real question is “How affordable are you?” and often “How affordable are you next to the schools we are also looking at?” I asked the cost question for each of the eight schools in both ChatGPT and Google’s AI Overviews, sixteen answers in all. In three of Google’s eight, the tuition was right and something a family would act on was wrong: a financial aid total from an earlier year, textbooks described as included when the upper school charges for them, and a boarding price that left out meals. ChatGPT’s eight had none of these. A tuition that is right and an affordability picture that is wrong is still a wrong answer for the family.
So publish one dated page on your website, the same page a family reads on a tour night, that separates tuition, required fees, optional costs, and what aid looks like for a family at a few income levels, and says which school year it describes. Do not write a page for the machine. The price test confirms the AI reads the one you write for families. Two rules for that page. Compare equivalent information: the same grade, year, day or boarding status, required fees, and first-year versus recurring costs. And leave unknown costs unknown rather than guessing. A page like that answers the question families most often get a wrong answer to.
One more thing from the same test. By the last step of the conversation, “how should we visit,” the AI wrote a usable visit guide in 8 of 8 cases, with questions for the child and things to watch for on campus. Families who reach that step are well served. A cost page a family can read, and a visit page that answers the questions it tends to raise, meets them there.
4. Audit the comparison on both sides, and the pages you do not own.
The natural question to ask an AI is “what does it say about us?” That question misses the way a school most often got hurt in my test, which was through a page about someone else.
Google’s AI Overviews quoted a wrong tuition in 9 answers across 6 of the 8 schools, and every one traced to a page the school did not write: a luxury real estate blog, a stale directory profile, a same-named school somewhere else, and an undated figure in a Wikipedia box. A school with the same or nearly the same name leaked into answers at 4 of the 8 schools. And in one comparison, the focal school’s facts were all right while the rival’s tuition was understated by nearly $24,000, so the AI’s verdict was that the rival was considerably cheaper. One case, and I hold it as one case. The lesson stands anyway: a comparison has two schools and a conclusion in it, and checking your own facts covers one third of it.
Three things to do. Ask the free AI apps “how does our school compare with [the school families most often weigh against you],” and check both schools’ facts and the verdict, not just yours. Search your tuition and your basic facts with your own site excluded from the results, and read what a parent would be told from everyone else’s pages. Then fix or retire what you find, starting with the old PDFs you forgot you published. Directories and review sites will take a correction. A namesake school in another state will not, but a line on your own page, “we are the one in [town], not the one in [state],” gives the AI something to disambiguate with. I did not test that last idea; it is the cheapest thing I can think of to try.
5. Count the family, not the mention.
Everything above measures whether an AI names your school. That is the only claim my test supports, and it is not the thing you care about. A mention can be wrong, qualified, or unpersuasive. A family can hear about you from an AI and then arrive through a tour sign-up or an ordinary Google search, and your dashboard will record the last step and miss the first.
The fix is older than AI. Add “an AI assistant” to the “how did you hear about us” question on your inquiry form, and make sure the answer survives into the enrolled record, not just the inquiry, so that in May you can count AI-first families who enrolled and not just families who inquired. Then ask on tours. Did an AI help you discover us, compare us, or rule someone else out? What did it say? Families are usually glad to tell you, and what they say is the only direct evidence of the thing all of this is about.
I would resist treating a better mention rate as enrollment growth until that column exists. The schools I tested were named more often with their category word than without it. Whether a single family chose differently because of it, nobody knows yet, including me.
Closing
I am glad I ran this. I went in expecting to confirm a hunch and came out with a much clearer picture of what happens between a parent’s question and the list of schools they see, and with this test and the five things I would do differently at a school tomorrow. I also know how fast this is moving. The assistants I tested in October will not behave the same way in the spring, so I have put a date on my calendar to run this again in a few months, broader if I can, and I will report what changed.
Part 1: Six Ways AI Search Is Not Google Search.
How this was made. Linda ran the tests and made the decisions; AI helped design the test, run the developer-version repeats, score the answers, and draft this text.
