TL;DR: US tutoring and exam prep is an $18.9 billion industry of roughly 176,000 businesses that grew 0.6% in 2025 — a share fight, not a rising tide. The class of 2026 SAT tells you where the share is: 1,954,833 test takers averaged 1045 (up 16 points), but the 968,119 who tested only through a School Day averaged 909 while the 692,135 who tested on a weekend averaged 1,179 — a 270-point gap between the two halves of the same exam. 92.6% of the 2,182 colleges with confirmed policies do not require scores for Fall 2027, and yet 51% of applicants sent them anyway. Demand is being manufactured upstairs: 62% of students now use AI for homework (RAND), 81% say it is easy to cheat with it (ACT), and 46% use AI in the college search (EAB). Which means the center's product is no longer an hour — it is five documents: the diagnostic read-out, the parent progress email, the eight-week plan, the score claim, and the essay note. Every one of them is either specific or worthless, and a hedged model produces the worthless version by default. Meanwhile Google's Imposing negativity rule means your fear-based ad costs you Customer Match and lookalikes, Unreliable claims bars the improbable-result promise, and the FTC still wants substantiation behind whichever number you print. The model that writes the defensible sentence instead:
venice-uncensored-1-2 at $0.20 in / $0.90 out per million tokens — about $0.54 a month for 880 progress notes, 120 diagnostic read-outs and 60 ad variants.
What's in this guide
- The 2026 market: $18.9B, 176,000 businesses, 0.6% growth
- The scoreboard: 1.95M takers, 1045, and a 270-point gap
- 92.6% test-optional, and 51% of applicants sent scores anyway
- The December 5 clock: eight days to the last fall deadline
- Why parents are calling now: 62% of students use AI for homework
- The five documents that cannot be hedged
- Refusing vs. hedging: what we actually got when we ran the prompt
- AI detection is a liability, not a feature: 61.3%
- Google's negativity rule: your fear ad costs you Customer Match
- The claim you can defend: unreliable claims, clickbait, and the FTC
- Students are not prompts: COPPA, FERPA and the PII you paste
- The cost math at $0.20 per million tokens
- A seven-step action plan
The 2026 market: $18.9B, 176,000 businesses, 0.6% growth
Start with the shape of the business. IBISWorld sizes the US tutoring and exam-prep industry at $18.9 billion in 2026, spread across roughly 176,000 businesses — a count that has compounded at 2.2% a year since 2020 while revenue compounded at 2.6%, with the industry managing an 0.6% uptick in 2025 after two years of demand whiplash. No operator holds more than 5% of the market. That combination — a fragmenting supply base, sub-inflation growth, and a customer who buys once a year — is the definition of a share fight.
The whiplash has a specific cause, and IBISWorld names it: universities moved away from requiring test scores, exam prep took the hit, and providers diversified into general tutoring to survive. ACT, GRE and GMAT volumes fell sharply while SAT participation rebounded. Read that as a warning about business models, not about demand. The families did not stop buying academic support; they stopped buying one product and started buying an outcome, priced per month.
| Metric, 2026 | Reading | Why it matters to a center |
|---|---|---|
| US tutoring & exam prep revenue | $18.9B | Big enough to matter, flat enough that every new student is someone else's |
| Businesses | ~176,000 | No dominant player; local trust is the moat, not brand |
| Five-year revenue CAGR (2020–2025) | 2.6% | Below tuition inflation — price is carrying growth |
| 2025 uptick | +0.6% | Recovery, not expansion |
| Largest single share | < 5% | Referral and reviews beat advertising budgets |
The second structural fact is labour. Tutoring is a business where the marginal cost of the next student is a human hour, and hours are the scarcest input in every center. That is why the useful AI in this vertical is not a chatbot on the website. It is the writing that surrounds the hour: the diagnostic that has to be specific enough for a parent to act on, the progress note that has to be blunt enough to change behaviour, the plan that has to name dates, and the claim that has to survive a substantiation request.
The scoreboard: 1.95M takers, 1045, and a 270-point gap
The College Board published the class of 2026 results on September 29, 2026. 1,954,833 students took the SAT at least once, down 2.5% from the class of 2025 — the first decline since the pandemic class of 2021. The mean total score rose 16 points to 1045, with Reading and Writing up 7 to 528 (back to pre-pandemic) and Math up 9 to 517 (still below the 2020 cohort's 523). 42% met both college and career readiness benchmarks. Participation grew in 42 states and DC, and international participation rose 13.5%.
That is the headline, and the headline is the least useful number in the report. The useful number is underneath it:
| Class of 2026, by testing route | Students | Share | Mean total | Met both benchmarks |
|---|---|---|---|---|
| School Day only | 968,119 | 50% | 909 | 18% |
| School Day and weekend | 294,579 | 15% | 1,179 | 68% |
| Weekend only | 692,135 | 35% | 1,179 | 66% |
| All takers | 1,954,833 | 100% | 1,045 | 42% |
Half the country's SAT takers test only because their school handed them a test on a Tuesday. The other half chooses a Saturday, pays, and often tests twice. The two groups differ by 270 points on the same exam, in the same year, under the same scoring rules. Whatever the causes — income, information, prep, retakes — that gap is the market. A center that can move a School-Day-only student 60 points is selling something a family cannot get from the school, and 60 points on a 909 base is a materially different story than 60 points on a 1,179 base.
The distribution matters for the promise you can honestly make. Of 1,954,833 takers, only 167,152 (9%) scored 1400 or higher; the median was 1030 and the 75th percentile 1220. A center advertising a 1500 average is advertising the top 1% of a curve it did not create. A center advertising "+118 points, n=214, completed 24 hours, 2024–2026" is advertising something an examiner could audit — and, as the policy section below shows, that is the only version that survives contact with an ad platform.
92.6% test-optional, and 51% of applicants sent scores anyway
Here is the contradiction that keeps tutoring alive. FairTest's August 25, 2026 tally covers 2,182 colleges with confirmed policies for Fall 2027: 2,021 are test-optional or test-free — 92.6% — and only 161 require scores, up by exactly one from the prior cycle. The return-to-testing story is real but narrow: among the 12 Ivy Plus institutions, only 4 remain optional (33.3%). Among the 262 ranked top-100 schools, 29 require scores.
Now the behaviour. In the 2025–26 application cycle, 51% of applicants submitted test scores, per Common App data — the first time the submitting share has been above half since 2020. Ninety percent of colleges do not require the score, and half of all applicants send it anyway. Parents are not buying compliance with a requirement. They are buying an advantage in a holistic review where a submitted 1350 beats an unexplained blank.
| Group | Confirmed policies | Optional or free | Require scores | Optional share |
|---|---|---|---|---|
| All active institutions | 2,182 | 2,021 | 161 | 92.6% |
| Ranked top 100 | 262 | 233 | 29 | 88.9% |
| State flagships | 52 | 46 | 6 | 88.5% |
| Admit rate under 30% | 104 | 84 | 20 | 80.8% |
| Ivy Plus | 12 | 4 | 8 | 33.3% |
| HBCUs | 86 | 77 | 9 | 89.5% |
One more number from that dataset is the marketing brief in a single line: at institutions that are test-optional, 62.4% of enrolled students arrived without a score (IPEDS 2024), against 27.8% at institutions that require one. A center's qualified lead list is therefore not "every junior." It is the family at a selective school whose kid is in the bottom half of the score distribution, or the family at a test-optional school who wants the score because the school down the road submits.
The December 5 clock: eight days to the last fall deadline
This is written on November 12, 2026. The December SAT is December 5; the regular registration deadline is November 20 — eight days out — with late registration closing November 24 and a $38 late fee. Scores land December 18. The next administrations are March 6, 2027 (deadline February 19) and May 1 (April 16).
Eight days is enough to change exactly one thing: the specific sub-skill the student fails. It is not enough to "do more prep." A center that answers a November 12 inquiry with a six-week curriculum has answered a different question. The correct answer names the test date, the number of sessions that physically fit before it, the two question types worth touching, and the honest sentence about what 23 days can and cannot move. That is a 200-word document per inquiry, and it is the single highest-leverage thing AI writes in this business — because it is written 40 times a week, at 9pm, by an owner who is also teaching.
The November inquiry is not asking for a curriculum. It is asking whether 23 days is enough. Answer that in the first sentence or lose the family to the center that did.
Why parents are calling now: 62% of students use AI for homework
The demand signal has shifted from "my kid is behind" to "my kid is cheating and it is working." RAND's American Youth Panel found homework AI use among middle school, high school and college students rose from 48% in May 2025 to 62% in December 2025 (n=1,214), driven by middle and high schoolers. 67% of students now agree that AI use for schoolwork harms critical thinking, up from 54% ten months earlier. Their top uses: better explanations of assignments (38%), brainstorming (35%), looking up facts (33%), drafting or revising writing (33%). Only 45% consider getting direct homework answers to be cheating; 78% say using AI to understand an assignment is not.
The parental version of that survey is worse. Pew (published February 2026, n=1,458 teens and parents) found 54% of 13–17 year olds had used a chatbot for schoolwork, with 10% using it for all or most of it. The Thomas B. Fordham Institute's fall 2026 survey of more than 2,000 students found roughly three-quarters misuse technology during school hours, about 1 in 5 use AI without teacher permission, and 26% admit using it during quizzes or tests.
And the thing that turns this into a tutoring purchase: ACT's January 2026 research brief (n=4,903 high school students registered for the June 2025 national ACT) reports 81% agreeing it is easy to use AI to cheat on assignments, 71% agreeing grades no longer accurately reflect performance because of it, 69% saying AI gives applicants an unfair advantage in admissions, and only 12% saying AI was involved in their ACT preparation. That last number is the whole pitch: the tool that inflates the GPA does nothing for the one number that is taken under supervision. ACT notes the obvious reason — test scores are "highly unlikely to be affected by AI."
Upstream, EAB's February 2026 survey of more than 5,000 high school students found 46% now use AI in the college search, up from 26% in spring 2025, and 18% removed a college from consideration based on what an AI answer said about it. If your center is not in the answer an AI gives about "SAT prep near me," you are not in the consideration set. That is the argument for published, structured, specific pages — not for keyword stuffing.
The five documents that cannot be hedged
Everything above converges on the same operational claim: the center sells five written artifacts, and each one has a legal or commercial requirement for specificity. This is where a safety-tuned model fails, and it fails differently in each case.
| Document | What the hedged default produces | What the file actually needs |
|---|---|---|
| 1. Diagnostic read-out | "Needs improvement in math; practice consistently." | The named sub-skill, the error pattern, the two question types, the score delta and the n behind it |
| 2. Parent progress email | "Every student progresses at their own pace." | The number, the attendance fact, the consequence, the date of the next decision |
| 3. Eight-week plan | "Build a consistent study routine." | Session dates against Dec 5 / Mar 6, assigned drills, and what is deliberately not covered |
| 4. Score claim (ad, landing page) | "Boost your score!" | Average gain, sample size, completion condition, period — substantiable if challenged |
| 5. Essay feedback note | A refusal, or a rewrite in the model's voice | Line-level critique of the student's own draft, naming the sentence that is doing no work |
1. The diagnostic read-out
The diagnostic is not a score. It is a decision: which 12 of the 400 possible sub-skills get the next 16 hours. A generic model writes "focus on algebra" because it does not know that the student loses points exclusively on linear-systems word problems where the units change mid-sentence. The prompt has to carry the item-level data, and the model has to be willing to rank deficiencies rather than distribute encouragement evenly across them.
2. The parent progress email
This is the one that costs centers their renewals. The email has to say: eight sessions, score down 20 in math, three of the last four homework sets skipped, not on pace for the 1300 target, and here is the change we are making Thursday. A model that inserts "every student learns differently" has just handed the parent a reason to do nothing. The sentence that changes behaviour is the one that names the cause and attaches it to a date.
3. The eight-week plan
A plan with no dates is a syllabus. Against a December 5 test there are 23 days; against March 6 there are 114. The document has to spend the days, name the drills, and state the opportunity cost — "we are not touching geometry; at 1040 it is not where the points are." Saying what you will not do is the part models omit, because it sounds like a limitation.
4. The score claim
The defensible version is arithmetic: average gain, n, the completion condition, the period, and the score band the sample actually represents. "+118 points across 214 students who completed the full 24-hour program, 2024–2026, mean starting score 1012" is a claim a regulator, a platform reviewer and a skeptical parent can all check. The indefensible version is "+300 points guaranteed," which is also, as the policy section shows, precisely what Google's Unreliable claims rule names.
5. The essay feedback note
Colleges have been explicit that they want the student's own writing; several now ask applicants to disclose AI use. A tutor's legitimate product is critique of the student's draft: which paragraph is a summary of a summary, which sentence tells the reader what to feel, which detail is the only one that could have come from this student's life. The illegitimate product is a rewrite in the model's voice — which is also the version that trips detectors designed for exactly that register. The line is not "AI or no AI." It is "does the student's sentence survive, improved, or is it replaced?"
Refusing vs. hedging: what we actually got when we ran the prompt
It would be tidy to claim that filtered models refuse these tasks. They mostly do not, anymore. We ran the parent-email prompt — 1040 diagnostic, math down 20, three of four homework sets skipped, 1300 target, "no hedging" — against a current mainstream model, and it complied in 115 output tokens. The result was correct, and useless in three specific ways: it left [Parent Name], [Your Name] and [Center Name] as bracket placeholders, it opened the consequence with the softener "For context:", and it contained no date and no next decision.
That is the actual failure mode in 2026, and it is worse than a refusal because it looks like work. A refusal is visible and teaches you to change tooling. A compliant-but-inert draft gets pasted into an email client by a tired owner at 9:40pm. What the file needed was: the parent's name pulled from the CRM, the specific date of the next session, the consequence clause ("we will hold Thursday 6pm and reassess after the December 5 test"), and no transitional padding.
Where an unfiltered model earns its keep
- Ranking deficiencies instead of balancing encouragement across them
- Writing the consequence sentence a parent will act on
- Critiquing a student's own essay line by line without rewriting it
- Explaining why an answer is wrong in the student's own words, at their reading level
- Turning 400 sub-skills into a 16-hour plan with dates and trade-offs
Where any model still needs a human gate
- Anything a student said about self-harm, abuse or home life — route to a person, not a draft
- Diagnosis language ("test for dyslexia") — medical framing is out of scope
- Reproducing or closely paraphrasing live test items — copyright and test security
- Final score claims — the number comes from your data, never the model's
- Anything sent to a minor without a parent in the thread
The practical rule: the model writes the document, the database supplies every fact. Names, dates, balances, attendance, score deltas and n come from your practice-management system as variables. If the model can invent a number, it will eventually invent a number, and a hallucinated score claim in a parent email is a substantiation problem disguised as a typo.
AI detection is a liability, not a feature: 61.3%
The Stanford study that every tutoring center should have read before promising "AI-free" work: Liang et al. tested seven GPT detectors on 91 TOEFL essays written by non-native English speakers. The detectors flagged 61.3% of them as AI-generated on average, and all seven flagged the same 18 essays (19%) as machine-written. On the US-born students' essays, false positives were near zero. The paper's conclusion is the operational one: these tools systematically misclassify the writing of people who learned English as a second language.
Now price that risk in a center whose fastest-growing customer segment is exactly that family. A detector verdict is not a measurement; it is an accusation with a confidence score attached, and the failure lands on a paying family who did the work. If you sell "we verify AI-free essays," you have sold a promise your instrument cannot keep — and unlike a bad ad line, this one gets resolved in a public review.
What to do instead: keep drafts and version history, coach from outline to final in-session, and evaluate writing by what a reader can verify (specificity, sentence variety, whether the detail could only come from this student) rather than by an opacity score. Use a detector, if at all, as a flag that triggers a conversation with the student — never as evidence in a message to a parent.
Google's negativity rule: your fear ad costs you Customer Match
Most tutoring marketing advice still says "lead with the pain." In Google Ads in 2026 that advice has a specific, checkable cost, and almost nobody in this vertical has noticed it.
Google's personalized advertising policy splits restricted categories in two. Access to opportunities (US and Canada) covers exactly three: consumer finance, employment and housing. Education is not in that list — which means a tutoring center may still use age, gender, parental status, marital status and ZIP code targeting, and may still run Customer Match. That is a real advantage, and it is the opposite of what the housing and employment verticals face.
But sensitive interest categories are the other list, and one of them is Imposing negativity, defined as "using a negative perspective or bias to promote any content category," with examples including "suggesting negative outcomes for users if they don't take specific actions." For anything in a sensitive interest category you cannot use advertiser-curated audiences at all: Customer Match, your data segments, lookalike segments and audience expansion are all unavailable. Predefined Google audiences remain allowed.
Your "your child is falling behind and will not get into a good college" campaign is not just a taste problem. It is a sensitive-interest classification that silently disables the customer list you spent three years building.
Two more targeting facts that change the funnel. Users under 18 are not eligible for personalized advertising of any kind, including serving from predefined audiences — so a campaign aimed at the student is a campaign aimed at nobody. And advertisers are barred from uploading customer information collected from viewers of child-directed content. In practice: market to the parent, run the retargeting on the adult, and keep the student-facing surface as a booking page rather than an audience.
The commercial implication is the useful one. Fear copy and retargeting are substitutes for each other. If you want to keep Customer Match, lookalikes and your data segments, the ad has to describe the outcome you produce rather than the disaster you prevent — "1040 in October, 1180 in March, here is the 8-week plan" rather than "don't let your child fall behind." Same offer, same price, and one of them keeps your audience tooling.
The claim you can defend: unreliable claims, clickbait, and the FTC
Three policy layers govern what a score claim may say, and they point the same way: the number or nothing.
Google's Misrepresentation policy bars "making inaccurate claims or claims that entice the user with an improbable result (even if this result is possible) as the likely outcome a user can expect." Read the parenthetical twice: a 400-point gain that one student achieved once is still barred as an implied expectation. The same policy's clickbait section bars ads that use negative life events — death, accidents, illness, arrests, bankruptcy — to "induce fear, guilt or other strong negative emotions to pressure the viewer to take immediate action," and its dishonest-pricing section covers the fake "was $2,400, now $1,200" anchor that this vertical loves.
The FTC supplies the substantiation floor: an objective claim needs a reasonable basis before it runs, and "results may vary" does not cure an unsubstantiated headline. Testimonials and reviews sit under the Endorsement Guides (16 CFR Part 255) and the Consumer Reviews and Testimonials Rule (16 CFR Part 465), which since 2024 has prohibited fake reviews, purchased reviews, review gating, and undisclosed insider testimonials — with civil penalties for knowing violations. Parent-influencer posts about your center are endorsements and need the disclosure.
| Line you want to run | Problem | Rewrite that survives |
|---|---|---|
| "Guaranteed +300 points or your money back" | Improbable-result claim; also a refund obligation you cannot administer | "Average +118 points, 214 students who completed 24 hours, 2024–2026, mean start 1012" |
| "Don't let your child fall behind" | Imposing negativity; disables Customer Match and lookalikes | "1040 in October. 1180 in March. The 8-week plan is on this page." |
| "#1 rated tutoring in the city" | Unsubstantiated superlative; needs a cited, dated source | "4.8 stars across 212 Google reviews, last 12 months" |
| "Was $2,400, now $1,200 (today only)" | Dishonest pricing if the higher price never ran | "$1,200 for 24 hours, billed after the third session" |
Students are not prompts: COPPA, FERPA and the PII you paste
Tutoring is one of the few small-business verticals where the data in the prompt belongs to a child. Two federal regimes apply, and they are usually both in play.
COPPA (16 CFR Part 312) governs personal information collected online from children under 13: verifiable parental consent before collection, purpose limitation, a retention policy, and deletion on request. A student portal, a practice app or a scheduling form that a 12-year-old fills in is in scope. The amended rule treats biometric identifiers — including data derived from voice or face — as personal information, which matters if you record sessions.
FERPA (34 CFR Part 99) applies the moment you contract with a school: you become a "school official" with a legitimate educational interest, you may not re-disclose, and the district's data agreement governs what you can do with records. Pasting a student's IEP status, grades or discipline notes into a public chatbot is a disclosure to a third party. It may also be a breach of your contract regardless of what the vendor's privacy page says.
The working rules that survive both regimes: send a student ID, not a name; send the sub-skill and the item-level data, not the narrative of the family's divorce; keep prompts out of any vendor that trains on inputs unless the DPA says otherwise; and keep retention short enough that you could describe it in one sentence to a parent. Every one of these is a template decision, which means it belongs in your prompt library, not in each tutor's judgement at 9pm.
The cost math at $0.20 per million tokens
A single-location center with 220 active students, a weekly progress note per student, 120 diagnostic read-outs a month and 60 ad or email variants, on the live Venice rate card checked today: venice-uncensored-1-2 at $0.20 in / $0.90 out per million tokens, and deepseek-v4-flash at $0.138 / $0.275. gemma-4-uncensored sits at $0.1625 / $0.50, and the role-play variant at $0.50 / $2.00.
| Document type | Monthly volume | Tokens in / out | venice-uncensored-1-2 | deepseek-v4-flash |
|---|---|---|---|---|
| Weekly progress note | 880 | 700 / 350 | $0.40 | $0.17 |
| Diagnostic read-out + plan | 120 | 1,400 / 700 | $0.11 | $0.05 |
| Ad / email variants | 60 | 500 / 400 | $0.03 | $0.01 |
| Total per month | 1,060 | 814k / 416k | $0.54 | $0.23 |
Per document, that is $0.00046 for a progress note and $0.00054 for a parent email on the uncensored model — roughly 46 cents per thousand notes. The comparison that matters is not model-to-model; it is against the alternative, which is one tutor spending 12 minutes a note. At 880 notes a month, that is 176 hours of tutor time. The margin question in a tutoring center in 2026 is not whether AI writes the note. It is whether the note is specific enough that the parent renews.
A seven-step action plan
- Split the lead list by testing route. School-Day-only students are the 270-point gap; weekend takers are the retake market. Different offer, different first sentence, different price anchor.
- Build the five-document template library before you buy a model. Diagnostic read-out, parent progress email, eight-week plan, score claim, essay note. Templates are where specificity gets enforced; a prompt written per-user will drift back to padding.
- Wire the facts, not the prose, to your practice-management system. Names, dates, attendance, score deltas and sample sizes are variables. The model writes sentences around them and never invents a number.
- Audit every ad against the negativity rule. Any line that implies a negative outcome for inaction is a sensitive-interest classification that disables Customer Match, lookalikes and data segments. Rewrite to the outcome you produce and you get the tooling back.
- Put a number, an n and a period in every score claim. Then keep the underlying spreadsheet, because "unreliable claims" and the FTC ask the same question and neither accepts "results may vary" as an answer.
- Stop selling AI detection. A 61.3% false-positive rate on non-native English writers turns your fastest-growing segment into an accusation. Coach from outline to final and evaluate the writing on what a reader can verify.
- Answer the November inquiry in 200 words with the date in it. December 5 registration closes November 20. The center that names the date, the sessions that fit and the honest delta wins the family before the other center opens on Monday.
None of this requires replacing your scheduling software, your CRM or your website. It requires a model that will write the specific sentence when specificity is the compliance requirement, that will rank deficiencies instead of distributing encouragement, and that costs less per month than one hour of tutoring. The market grew 0.6% last year and the average score moved 16 points; the operators who win are not the ones with more students in the funnel. They are the ones whose diagnostic names the sub-skill, whose email names the date, and whose claim names the sample.
Write the read-out the parent acts on
RawDialog runs venice-uncensored-1-2 and deepseek-v4-flash at $0.20 / $0.90 per million tokens — no filters sanding your progress notes into "every student learns differently," no per-seat pricing, no training on your student data. Bring your template library; keep your practice-management system as the source of every number.
Related reading on this blog: the Venice AI uncensored review, the developer API guide, uncensored AI for childcare centers, coaches and course creators, and UGC video ad creatives.