In my research this fall, the biggest difference I found between a family searching Google for a K-12 school and a family asking an AI assistant is who chooses the search words. On Google, the parent does. With AI, the parent describes the child, and the AI decides what to search for.
That one change touches much of what schools have learned about being found online. This article covers six ways AI search differs from the Google search most of us know how to work. A second article covers what I would do in response. I kept them separate so the findings can stand on their own.
I can show the Google half. I looked at five years of Google search data for 65 ways a parent might phrase a private school search. The phrasings that register are the ones you would expect: “best private schools in” a city, “private school tuition,” “private schools near me.” Every phrasing that describes a child, “is private school right for my child,” “private school for my son,” “private school for ADHD,” was too rare for Google to report any volume at all.
The AI half is newer, and I will tell you what I know and what I believe. At NACAC in October, EAB read out the prompts students actually type when they research colleges: “I have a 3.6 GPA, what schools are competitive for me?” They describe themselves. I believe K-12 parents will do the same thing with their children, and I believe the reason they rarely did it on Google is that nothing useful came back when they tried. An assistant takes the description and comes back with a list of schools. So I tested what happens when a parent does.
I tested it with eight private schools I have no connection to. For each one, I put myself in the seat of that school’s enrollment or marketing director, asked the questions a parent would ask, the way a parent would ask them, and watched whether the AI named that school. Below, “the school I was testing” always means that one. The full method is in the box below for anyone who wants it. The short version: the free versions of ChatGPT and of Google’s AI Overviews, the AI answer that now sits at the top of a regular Google search, logged out and in a fresh chat every time; and for some tests the developer versions of ChatGPT, Gemini and Claude, which let me repeat a question ten times and read the web searches the AI ran before it answered.
If you know how to get your school found on Google, you should know that AI is changing what happens behind the scenes between a parent’s question and the list of schools they see. Here are the six differences I found.
1. The AI chooses the keyword, not the parent.
When a parent describes a child, the AI does something the parent never sees. It reads the description and decides what kind of school that child needs. Not a particular school, a kind of school: a progressive school, a Montessori program, a gifted program, a Catholic school. Then it searches the web for that kind of school in the town the parent named, and whatever that search returns becomes the parent’s shortlist.
I could watch this happen in the developer version of ChatGPT, because it shows the searches it ran. One of the eight schools in my test was a Catholic elementary school. I wrote a parent describing a five-year-old whose family wanted kindergarten to be about more than early reading, and asked what kind of school would help and which ones to look at. ChatGPT turned that into searches for “progressive kindergarten,” “Montessori,” and “social-emotional learning.” It never typed the word Catholic, and in 10 conversations with that parent, the Catholic school I was testing was named 0 times. Then I changed one word in the parent’s question, from “which schools should we look at” to “which Catholic schools should we look at,” and left everything else the same. The school was named in 10 of 10.
I ran the same experiment at all eight schools, adding each school’s own word: Catholic, Episcopal, boarding, progressive, independent. Across the eight, the school I was testing was named in 48 of 80 conversations without its word and 70 of 80 with it. Almost all of that swing came from two of the four faith schools, the Catholic elementary school and the Episcopal high school. The other two faith schools were already named 9 and 10 times out of 10 without the word. The small progressive school barely moved, 9 of 10 to 8 of 10. The boarding school went the other direction: adding the word “boarding” made it slightly worse, 4 of 10 to 2 of 10. The word is a key for some schools, not all.
The two Catholic schools show when it is. The Catholic high school was named 10 of 10 times either way. The parent had asked for a son who would “show up for others,” and the AI already knew that school as the one built around service, so it named it without ever searching for the word. The Catholic elementary school had no such luck. A kindergarten that grows “heart and character” belongs, in the AI’s memory, to progressive and Montessori schools, and a parish school is one of dozens with a saint’s name. The word matters when the school’s kind is the only distinctive thing a parent’s description could reach. It stops mattering when the school is already known for the very thing the parent asked about.
Here is why this matters for anyone who has worked on search. On Google, a school could learn which words parents typed and write its pages for those words. With AI, the words that decide the search are often ones the parent never typed. The AI reads the child, picks a kind of school, and searches for that. If the kind of school it picks is not what you call yourself, your pages may never come up in that search, however well they are written. And even when they do come up, the AI can read them and still leave you off the list. That happened twice at the boarding school: its own website was among the pages ChatGPT read before answering, and the school was not named.
2. Your website still wins when you are named. Discovery moved upstream.
When I asked what a school costs, both ChatGPT and Google’s AI Overviews went straight to the school’s own website in 16 of 16 answers and got the tuition right in 15 of them. When I asked which schools a family should look at, the school’s own site was among the sources in 9 of 24 answers.
The full conversation showed the same thing. I had ChatGPT walk a parent through the whole search, one chat per school: where do we start, what kind of environment would suit her, which schools, tell me about this one, how should we visit. At the second step, in 7 of the 8 conversations, the AI described an environment that fit the school I was testing, small classes, time outdoors, a faith community, whatever that school says about itself. At the third step, the shortlist, the school was missing in 5 of the 8. When I then asked about the school by name, all 5 times the AI found real alignment with the child it had just helped me describe.
One more thing about those shortlists. The schools on them were often not the ones 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. The faith-based kind is what dropped out. A school’s familiar list of rivals may leave out most of what the AI is actually putting in front of the family.
The website you have worked on for years is doing its job once a family knows your name. The work that is new sits one step earlier, at the moment the shortlist forms.
3. One test is one card from the deck.
A Google ranking is something a marketer checks once and trusts for a while. An AI shortlist does not work that way. When I asked the same question ten times in ten fresh conversations, the same school came up anywhere from 0 to 10 times. Only the clearest cases held steady: two schools were named every time, one never, and the rest landed somewhere in between on any given day.
If you have asked an AI about your school once and felt reassured, or alarmed, you drew one card.
4. Other people’s pages are now your voice.
The AI assembles its answer from whatever it finds online, and much of what it finds about your school is not yours. Google’s AI Overviews quoted a wrong tuition in 9 answers across 6 of the 8 schools, and every one of those traced to somebody else’s page: a luxury real estate blog, a stale directory listing, a school in another state with the same name, and in one case the school’s own viewbook from an earlier year, still sitting on a flipbook site. Real estate and relocation sites showed up among the sources at 7 of the 8 schools. A school with the same or a nearly identical name leaked into at least one answer at 4 of the 8, nine answers in all.
The same thing can happen through a rival. In one comparison, a school’s own facts were all correct, but the other school’s tuition was understated by nearly $24,000, apparently taken from a school with the same name somewhere else, and the AI concluded that the other school was considerably cheaper. One case, so I hold it lightly. It means a comparison can go wrong for you through a page about someone else.
Most AI answers show their sources somewhere, but not always where a reader will see them. In my test, Google’s Gemini app showed none unless I opened a separate menu, and some of ChatGPT’s citations carried no web address at all. On Google, the sources are the results page itself, and the parent chooses among them. In an AI answer, they sit under or behind a finished paragraph, few readers open them, and in my test they were a different set from what a Google search would show: review sites, Reddit threads, real estate pages, and schools with your name in other states.
5. Not all assistants search the same way.
With Google there was one algorithm to understand. Now there are several assistants, and they behave differently. In my test, ChatGPT and Google’s AI Overviews shared only 34 of the 195 websites they cited for the same schools, about 17 percent. Gemini sometimes followed a single directory’s list in that directory’s order.
The biggest difference was how hard each one searched before answering. In the developer versions, ChatGPT ran about two web searches per question. Claude ran about four and read the pages, and it named the school I was testing in 77 of 80 conversations without the parent’s word, and the two faith schools ChatGPT had missed in 17 of 20. Gemini searched least of all, about once, and often answered from what it already knew. The harder the assistant searches, the less the parent’s label matters. The assistant most families use sits in the middle.
One more difference sits underneath. They do not all search the same index. Gemini and Google’s AI Overviews use Google; ChatGPT and Claude are reported to use other indexes. Ranking well on Google says nothing about how you appear in the others.
6. Rewriting your website for the machine has no measured payoff.
You will hear advice to rewrite your website for AI, and it is a reasonable thing to want. I went looking for the evidence behind it and could not find it. The two academic studies that tested distinctive wording for AI search found no lift. Google says plainly that no special markup or optimization is needed to appear in its AI features. In my own test, the schools the AI found most reliably with no label from the parent were the ones that are distinctive in fact: the small progressive school in a rural town, which ChatGPT named 9 of 10 times without the word and 8 of 10 with it, and the boarding school, which the assistants that searched harder named every time.
I did not test website changes, so I cannot tell you what on a page moves an answer. What I can tell you is what the answers were built from, and that is where the next article starts.
What comes next
Those are the six differences I found. Each one points to a step in a family’s search that now happens inside the AI’s answer, out of sight of both the family and the school. In the second article, I look at these results and consider what a school can actually do with them, in how it reaches and engages families who now start their search in a conversation with AI.
Part 2, on what I would do in response, is coming soon.
