The fact sheet
Tutoring in America, in numbers.
Every figure on this page has a source. Click any of them.
Part 01
The market
Who takes the SAT, who pays for help, and the public money that changes the equation.
The SAT
About 18% of SAT takers pay for private prep. 11% take a paid class and 7% hire a private tutor. The other 82% get no paid help.
6 of 8
Ivy League schools require test scores again
Six of the eight Ivy League schools require the SAT or ACT again. All eight will require scores when the announced policies take effect.
Yale Admissions→$800+
the price of a brand-name prep course
Live online SAT courses from the big prep companies start near $800. Most families cannot pay that.
Kaplan pricing→Tutoring
13 million of America’s 54 million K-12 students get tutoring. In a survey of 23,000 parents, that share jumped five points in one year.
11%
expected yearly growth of US private tutoring
Analysts expect the US private tutoring market to grow by $28.85 billion between 2024 and 2029. That is about 11% growth each year.
Technavio→$100+
the hourly price of a private SAT tutor
Experienced SAT tutors in major metro areas charge $100 to $300 an hour. The price decides which families get one.
Private Prep→Public money
Education savings accounts (ESAs) give families state money to spend on tutoring, curriculum, and online education.
18 states run education savings account programs, 21 programs in total. The six filled dots are the states where every family qualifies, regardless of income.
489K
students with a funded ESA account
About 489,000 students had a funded ESA account in the 2024-25 school year. The number grows every quarter as new programs start.
EdChoice Fast Facts→$9,572
the average ESA award in Arizona
Arizona's average ESA award is $9,572 a year. Florida's average is about $8,000. The money follows the student, not the school.
EdChoice Fast Facts→$12,790
the average private-school tuition per year
The average US private school charges $12,790 a year. The average private high school charges $16,420. Elite schools charge $60,000 or more.
Education Data Initiative→The market, added up
The totals come from simple chains of public numbers. Here is the arithmetic, so you can check it.
SAT prep
About 2 million students take the SAT each year. About 18% pay for private prep, at $800 to $2,000 each. That is a market of $300 to 600 million a year.
All tutoring
America has 54 million K-12 students. About 24% get tutoring. That is 13 million students. Assume half of them pay privately, at about $300 a month (our assumption). That is a market of $15 to 20 billion a year.
The check
Analysts estimate US tutoring and test prep at $30 to 40 billion a year. Our bottom-up estimate is smaller than theirs. A smaller estimate means we are conservative.
The school layer
4.7 million students attend private school. The average tuition is $12,790 a year. That is about $60 billion a year in tuition. ESA programs add about $4 billion a year of state money.
Sources for every input are linked in the cards above.
Part 02
The evidence
What the research says: tutoring works, unguided AI backfires, and AI built for learning holds up in randomized trials.
How to read these numbers
The studies below use two tools. A randomized trial assigns students to a program by chance, like a drug trial, so the difference between groups is causal. An effect size measures the gain in standard deviations (SD): a +1.0 SD gain moves an average student from the 50th percentile to the 84th. Most school programs that work at all gain less than +0.10 SD, so anything above +0.20 SD is a strong result. Kraft (2020) is the standard guide.
Tutoring works
One-on-one tutoring is the most consistent result in education research. The problem was never whether it works. The problem is the price of a human.
Average gain, in standard deviations (SD)
Each bar is the average gain across 263 randomized tutoring trials, grouped by how many students the program served. Small pilot programs average +0.51 SD, a very large gain for education research. Programs with more than 1,000 students average +0.16 SD, about a third as much. The authors call this decay the central problem of tutoring at scale.
98th
the percentile of the average tutored student
In Benjamin Bloom's 1984 studies, the average student who got a one-on-one tutor scored higher than 98% of students in a normal classroom. That is a +2.0 SD gain, and Bloom named the finding the 2 Sigma Problem. Later, larger studies find smaller effects, but tutoring keeps its lead.
Bloom, 1984→96
randomized trials of tutoring, one verdict
A 2020 review pooled 96 randomized trials of tutoring. It found some of the largest and most consistent gains ever measured in education research.
NBER→+0.55
SD, the effect of explaining your reasoning out loud
Across 64 studies, students who had to explain why an answer was right learned substantially more than students who only practiced. The average gain was +0.55 SD, more than double the bar for a strong result.
Bisra et al.→AI, unguided
Handing a teenager a chatbot that answers homework questions is not tutoring. The best experiments we have say it backfires.
Researchers followed 26,811 students after they started to use AI. Homework scores rose 18%, but closed-book exam scores fell 20% within six months. This is a warning sign, not yet proof of cause.
worse
exam scores after the AI was taken away
High school students who practiced with unrestricted ChatGPT looked better during practice. When the AI was removed, they scored worse on exams than students who never had AI at all. A version with tutoring guardrails removed the harm but added almost no gain.
Bastani et al., PNAS→64%
of US teens use AI chatbots
64% of US teens use AI chatbots, and about 3 in 10 use one daily. Only 51% of parents know their teen uses them.
Pew Research Center→59%
of teens say AI cheating is common at their school
59% of teens say students at their school regularly use AI to cheat. The default use of AI in school today is the escape hatch, not the tutor.
Pew Research Center→AI, built for learning
The same technology, engineered around how learning works, produces real gains in randomized trials.
50→60
the average student's percentile after 8 weeks
In a randomized trial with 1,763 students in Sierra Leone, a deliberately constrained AI math program moved the average student from the 50th to the 60th percentile in 8 weeks. That is a +0.26 SD gain, above the bar for a strong result.
Google DeepMind→194
Harvard students in a head-to-head trial
In a randomized trial with 194 Harvard students, a carefully designed AI tutor taught more physics in less time than one of the best physics courses in the country.
Kestin et al., Scientific Reports→62→66
the pass rate when human tutors got an AI assistant
In a field trial with 1,800 students, an AI assistant raised the pass rate of tutored students from 62% to 66%. The least experienced tutors improved the most.
Wang et al.→+4
points of A and B grades from check-in messages
In two randomized trials with 2,483 students, two to three targeted check-in messages a week raised the share of A and B grades by about 4 points. The tutor that reaches out first wins.
NBER→Part 03
The vocabulary
The terms behind the research. Each definition is one breath long, and each ends with how StudyHall applies it.
Cognitive load theorySweller, 1988
Your working memory holds only a few items at one time. When a problem overloads it, the small steps drop first, and learning stops.
In StudyHall: We drill the basics until they are automatic. This frees your working memory for the real reasoning.
Spaced repetitionthe spacing effect
Practice spread over days beats the same practice crammed into one sitting. Memory grows most when you review right before you forget.
In StudyHall: The daily quiz brings each missed skill back on a schedule, right before it fades.
Retrieval practicethe testing effect
When you pull knowledge out of your head, you remember it better than when you read it again.
In StudyHall: Everything in StudyHall is a question. There is no passive review.
Interleaving
Practice that mixes related skills beats practice in tidy blocks. It feels messier. It sticks better.
In StudyHall: The daily quiz mixes skills in the same proportions as the real SAT.
Desirable difficultiesBjork
Practice that feels harder, with spacing, testing, and mixing, produces learning that lasts longer. Practice that feels smooth often does not.
In StudyHall: The quiz feels effortful on purpose. We never optimize for easy.
Zone of proximal developmentVygotsky, 1978
Learning happens in the band between what you can do alone and what you cannot do at all. Too easy teaches nothing. Too hard teaches panic.
In StudyHall: We pick each question to sit just past what you can do alone, and we hold you in that band.
Scaffoldinggradual release
Support that fades as your skill grows. The tutor shows you, then works with you, then watches you, then times you.
In StudyHall: Click, our tutor, tracks a stage for every skill: I do, we do, you do, then timed. Support fades as you earn it.
Mastery learningBloom
You demonstrate each block of a skill before you move to the next one. School cannot run this way, because the calendar outranks the student.
In StudyHall: Skills unlock when you demonstrate them, not when the calendar says so.
Self-explanation
Students who explain why an answer is right learn much more than students who only practice. Saying the reasoning exposes it, so a tutor can fix it.
In StudyHall: This is the whole voice loop. You cannot skim a conversation that you are in.
Socratic feedbackwhat expert tutors do
The best human tutors almost never say wrong. They ask the one question that makes the student find the error. This protects the student's confidence.
In StudyHall: Click argues with you instead of answering you. The answer key stays on our server, so Click cannot cave and reveal it.
Step-based tutoringVanLehn, 2011
Software that works through every step of your reasoning gets close to a human tutor. Software that only grades your final answer gets less than half of that effect.
In StudyHall: We find the step where your reasoning broke, not just the letter you picked.
Misconception
A wrong turn with a name, not a random error. The same few wrong turns appear across thousands of students, and each one has a fix.
In StudyHall: Click keeps a short list of your active misconceptions and retires each one when you beat it.
Tutoring works, and most families are priced out of it. That is the whole problem we exist to fix.
For the learning science behind StudyHall, read the manifesto. Numbers checked August 2026.