How accurate are AI answers about college admissions consultants? Here’s what our tests found.

Written by The Ivy Institute

The Ivy Institute | College Admissions Research | November 2026

Abstract

A family asking what college advising costs needs to know how much it might owe before and after a decision, which services that amount purchases, and whether an affordable alternative is still accepting applications. This study tested those tasks in a fixed five-question interview delivered to consumer AI products on November 8, 2026. The initial wave produced two complete answers, from signed-out ChatGPT and Gemini. Google Search presented a verification challenge before its results appeared. Across twelve decision fields developed for this audit, ChatGPT supplied adequately supported information for all twelve. Gemini supplied five supported fields, two partial answers, one contradicted price and four omissions. These counts describe two particular responses under a source-requesting prompt. They cannot estimate product-wide accuracy. The most consequential error was favorable to the commissioning firm in one respect: Gemini used another provider’s lower fees when describing The Ivy Institute. It also omitted the Institute’s conditional second payment, which could make the offer look less favorable on a different dimension. A same-day repeat in fresh chats dropped the wrong fee menu but introduced an unverified suggestion of possible refunds. A single direction-of-bias label would conceal both consequences. The report presents the capture, field-level judgments, evidence rules and repeat-wave design. It also examines a complication for fact-checkers: two first-party pages can support different interpretations of how a payment is scheduled.

A number can be real and still answer the wrong question

The cleanest error in this study involved two numbers that actually exist. A consumer receiving the Gemini answer could open a page, see both numbers and conclude that the answer had been verified. The decisive detail was at the top of the page. The company selling those plans was The Ivy Mind. The family had asked about The Ivy Institute. One word in a proper name separated an apparently precise price comparison from an incorrect one. The incident illustrates why this study begins with families’ decisions. A system can retrieve a genuine page, extract the numbers correctly and provide a working link while failing the task of telling a family what the named provider charges.

A second reader might spot the identity mismatch yet still misunderstand the correction. The Institute’s costs page describes two payments of $12,500 under its senior-year commitment. The first starts the engagement. The second depends on an acceptance at one or more of the eight designated colleges. A corrected headline of $25,000 would restore the quoted maximum on that schedule while leaving unanswered the question that matters most to a family evaluating financial exposure. The relevant comparison requires naming the outcome as well as the price. It also requires separating the provider’s advertised terms from the agreement a particular family signs. A consumer AI answer cannot examine an agreement it has never received.

The third problem was quieter. A free mentoring program can be accurately described as free, and its benefits can be accurately specified, when a new applicant has no current seat available. The College Essay Guy page displayed both an application process and a year-specific capacity notice. A reader who stopped at the process description could reasonably conclude that preparing an application was the next step. The capacity notice instead pointed toward an interest form. The unit of interest in this study is therefore a usable answer on a particular day. Some of the most important information takes the form of an exception to an otherwise accurate general description. [I1–I3, I7]

Research questions and the choice of a family interview

The primary question was whether an answer would allow a US family to distinguish the financial terms of three different kinds of help: a conditional senior-year package, individualized commercial advising and free, eligibility-limited individual support. A second question concerned professional qualifications. A family that values admissions-office experience needs evidence about the person who will do the work. A statement about the combined experience of a team answers a different question. The final question asked who controls admission to Harvard. It gave each system an opportunity to distinguish advice, sales promises, payment commitments and university decisions.

All three named providers appeared in the question. Their later appearance in a response contributes no evidence about unaided brand visibility. The test also cannot identify the best provider. It deliberately chooses one commercial term where price is published, a second where a bounded negative finding is plausible, and a third where the main barriers are eligibility and capacity. That variation makes the audit sensitive to different retrieval problems. It also prevents treating the result as a random sample of commercial firms, programs, prices or family inquiries. A person who simply types a company name into a search box faces a substantially different task.

Five questions were sent together. Keeping them together models a family asking a comparison question in a single conversation. The company names stay in context while the system moves between payment terms and eligibility, creating a realistic opportunity for cross-company contamination. This design has an important cost. Five topics within one generated answer do not supply five independent observations of the product. In particular, an early identity error can distort several financial fields downstream. Counts of fields describe the completeness of the resulting consumer information. They do not imply that the model independently failed four or five times.

The question explicitly requested primary sources. This likely changes which pages a product retrieves and how it presents uncertainty. A product that does well when asked for evidence might behave differently when a parent asks an unqualified question on a phone. The design also gives a lengthy numerical answer more room to demonstrate completeness. It does not measure reading burden directly. A family with two minutes to decide whether to book a consultation could prefer a very short response, provided it preserved the few details that change what happens next. The scoring rules award no credit for length, but length can affect the opportunity to answer a greater number of fields.

Why citation research matters here

Research by the Tow Center tested retrieval and attribution using news content. Its tasks concerned whether a generative search tool could identify a news source and supply the correct attribution. Our task asks whether evidence can support a purchasing or screening decision about services. The two designs do have a methodological point in common. Checking that a citation exists is a substantially easier task than establishing that its publication supports the exact claim. Transferring a numerical error rate between the studies would require common prompts, outcomes and source populations that the studies do not share. The relevance of the Tow Center work is its distinction between attribution and factual support. [I13]

Pew Research Center’s browser-panel research examined what people clicked after Google searches, including searches that displayed AI summaries. That is a behavioral question. Our audit concerns the content a person could have read, regardless of whether the person clicked through. A family’s failure to open a source would compound the practical significance of a misleading displayed number, but this study observed no families using the answers. It cannot infer that anyone paid a different price, declined a service or missed a scholarship because of this test. The distinction matters especially for headlines. Evidence of an error is valuable without a claim that it caused a transaction. [I14]

Vendor documentation also sets expectations. OpenAI describes search responses with citations and acknowledges that results can be incomplete. Google describes circumstances in which AI features appear and cautions that Gemini responses can contain inaccuracies. These product descriptions explain why the capture instrument includes both availability and accuracy. A search results page can be reached without generating any AI content. Conversely, a site can block access before anyone can determine which features it would have displayed. These states have different evidentiary meanings. Treating them both as an incorrect answer would let infrastructure failures masquerade as factual analysis. [I10–I12]

Collection protocol and provenance

The observation was made using separate consumer web tabs. The interfaces offered sign-in, and no personal account was selected. The prompt specified a US family and was written in English. Browser egress location was not independently verified. The date and time on the capture files are UTC. The two completed responses were recorded at approximately 18:46 UTC, or 1:46 p.m. Eastern. Gemini displayed Flash-Lite. The ChatGPT interface indicated web searching, but the collector did not observe a reliable backend model identifier. The product is identified by the consumer interface that produced the text. Labels about exact models should remain missing unless a future collector can verify them at collection time.

The collector retained the visible answer wording and explicit destination links. An answer can also display citation chips, contextual links or a general source list. The archive distinguishes explicit source destinations captured in the answer from source names visible on some Gemini chips. It makes no claim to reconstruct every result consulted internally. Nor can a reader reliably infer internal reasoning from the ordering of sources. In the wrong-company example the match between the two prices and the linked offering establishes the attribution error. It does not reveal how the system selected that link or whether a particular internal component malfunctioned.

The Google Search query used the same wording. A verification page appeared before ordinary results. Consequently the collector could not determine whether an AI Overview would have appeared. There is no Google Overview answer in the scored denominator. AI Mode is a distinct surface that was visible as a homepage option but was not independently tested in this baseline. The Gemini response is also kept separate from both Search surfaces. Grouping them under the name Google would turn one returned answer and one blocked query into an uninterpretable product score.

The facts used in adjudication were checked on November 8, 2026. Some publisher pages include an explicit program year or publication date. Others are ordinary service pages that can be revised without a visible change history. The source register supplies access dates, page identities and retrieval fingerprints where available. An access date establishes what the research process saw at that point; it does not establish the first day the policy came into force. When a future wave differs, the first follow-up is to ask whether the publisher changed its public statement. A model correcting itself and a business revising its terms have different interpretations.

For audit purposes a firm’s own page is strong evidence of what it publicly advertises. It is weaker evidence of consistent delivery and cannot independently verify its own success rates. A professional association can explain how it expects members to behave but does not speak for an admissions committee. Harvard can describe its selection process, whereas a firm cannot authoritatively describe how its clients would have fared without advice. The hierarchy of sources thus shifts with the claim. A page can be primary evidence for the contents of an offer while remaining inadequate evidence for its effects. [I1–I9]

Fixed questionnaire

We are a US family comparing independent college admissions consultants for first-year undergraduate applications. Please answer each question briefly and give the primary-source link for each factual answer. 1. What does The Ivy Institute charge for senior year, when is each payment due, and what happens if none of our eight selected colleges admits our student? 2. Does IvyWise publish one standard price for comprehensive counseling, and what services does it describe? 3. What does a College Essay Guy Matchlighters Scholar receive, and who qualifies? 4. Does every counselor at a firm that mentions former admissions officers personally have admissions-office experience? How can we verify the person assigned? 5. How should we compare these options, and can any of them guarantee admission to Harvard?

Scoring and uncertainty

Twelve decision fields were coded after collection for this exploratory wave. They are now fixed for repeat waves. Fully supported means the returned content communicated the central decision-relevant fact and agreed with the dated references specified for that field. Partial means that the answer communicated a material part but omitted an essential qualifier or added an unsupported requirement. Contradicted means that the answer supplied a fact materially inconsistent with the target source. Omitted means there was no usable answer for that field. A vague acknowledgment that contract terms exist does not say what the payment amount or trigger is. An inaccessible product receives no field scores.

The field list includes current availability even though the prompt asks who qualifies, without explicitly asking whether enrollment is open today. That choice favors practical usefulness. Reporting the list in full allows readers to remove the capacity field when they want a narrower test of literal responsiveness. Similarly, some readers might award partial credit for a generic description of installments. We retained an omission because neither the first amount nor its due event could be recovered from that language. Publishing such decisions makes a small study inspectable. It also prevents the appearance of a universal fact-checking standard where reasonable editors could differ.

Codes were assigned by the AI-assisted authoring process. There was no separate blinded second adjudication. The commissioning organization is one of the providers named in the prompt. The underlying answers, reference set and reasons are therefore more informative than an unqualified assurance of impartiality. A second coder can work from the supplied files, remove product labels, score the rows independently, and compare reasons before consulting the original table. A company can point to a missing or dated source. Its disagreement alone does not change a code. The correction should refer to evidence that was applicable to the same offering on the original date.

A supported field is not an endorsement of every sentence nearby. Broad advice about budgets, credibility or fit was evaluated separately from whether an institution makes its own admissions decisions. The answer can succeed on the latter while relying on a weak citation for the former. A twelve-of-twelve result therefore does not mean perfect accuracy across a long response. Conversely, a product that omits an unasked but important current capacity notice has not necessarily stated anything false. Reporting all four categories preserves the difference between error, excessive confidence and missing assistance.

Findings from the first wave

Figure 1. Fully supported decision fields in the first interview
ChatGPT12 / 12
Gemini5 / 12

One answer per product at baseline. The other Gemini fields comprised two partial answers, one contradiction and four omissions. Google Search could not be observed.

Table 1. Baseline decisions against the dated source set
FieldReferenceChatGPTGemini
Senior-year full feeI1SupportedContradicted
First paymentI1SupportedOmitted
Second payment conditionI1, I2SupportedOmitted
No acceptance among eightI1SupportedOmitted
Universal comprehensive priceI4, I5SupportedSupported
IvyWise service scopeI4, I5SupportedSupported
Free individual support and hoursI7SupportedSupported
Residency and income criteriaI7SupportedPartial
Academic criteria and extra hurdlesI7SupportedPartial
Applicant availability this cycleI7SupportedOmitted
Individual credential verificationI6, I9SupportedSupported
Authority over Harvard admissionI8, I9SupportedSupported

The two captured responses agreed on several matters likely to help a family. Both described the free program’s ten hours of individual help. Both encouraged the reader to examine the assigned counselor's background and treated the university as the decision maker. Both characterized IvyWise’s comprehensive advising as individualized and named services consistent with the firm’s description. These agreements are specific to the bounded questions in the table. In particular, naming a service available somewhere in a practice does not establish that it is included in every package.

The clearest disagreement concerned The Ivy Institute. ChatGPT gave the two-payment schedule on the firm’s costs page and also identified the differing installment language on its homepage. Gemini named a cheaper range, attributed it to school-count plans, and linked to The Ivy Mind. The upper figure in that answer was $10,401 below the $25,000 conditional total on the Institute’s costs page. The lower figure was $4,901 below the $12,500 due at the start under that schedule. These differences illustrate the scale of potential misunderstanding. They are arithmetic comparisons across different offerings, not discounts or actual prices that any one family was entitled to receive. [I1–I3]

Gemini also mentioned test scores when describing a free advising benefit. The program’s own eligibility list identifies residence, educational plans, unweighted GPA, income and motivation. It does not list a test-score threshold. Calling scores typical is weaker than saying the program automatically excludes applicants below a threshold. Nevertheless, the suggestion appears in the answer to who qualifies. A reader without scores could treat it as evidence that applying would be pointless. We code the academic field partial and identify the unsupported addition in the reasons. It would be inappropriate to describe the product as having explicitly imposed a particular SAT cutoff. [I7]

The two products used different source types. ChatGPT linked directly to the Institute’s costs page, two service pages for IvyWise, the Matchlighters program, the IvyWise team and Harvard’s own materials. Gemini linked the firm’s FAQ together with the wrong-company fee menu. It also labeled a ranking site and a finance publication as primary sources in portions of its answer. Those publishers may investigate a firm usefully. They do not become the issuing authority for a counselor’s employment record or for university selection policy because an answer labels them primary. A bibliography that mixes issuing authorities and commentators needs claim-by-claim inspection.

How much of a comparison can these scores bear?

The strict supported-field count was twelve of twelve for this ChatGPT response and five of twelve for this Gemini response. Assigning half a point to each partial would make Gemini’s illustrative total six of twelve. Removing the current-capacity field, which the family did not explicitly ask about, gives eleven supported fields out of eleven for ChatGPT and five out of eleven for Gemini. Counting only fields on which an answer made a substantive attempt would change the Gemini denominator again. That restricted calculation is especially susceptible to gaming: silence on a difficult refund question becomes invisible. The article therefore displays omissions alongside contradictions.

The baseline alone contains two returned answers and one blocked surface. Twelve related fields within each answer do not justify a conventional confidence interval for product accuracy. Four financial fields describe a single conditional transaction. Errors within those fields are dependent. Giving one wrong provider name can make the rest of that subsection unusable without implying four unrelated retrieval failures. A different audit could count question-level completeness or weight severe errors more heavily. If the units change, the report needs a new analysis label. A sample with one compound prompt cannot establish that one product is generally more reliable for price comparisons.

The positive finding for ChatGPT also has boundaries. A reader might accept its statement about the two equal payments and overlook that the Institute’s homepage describes flexibility in when and how families pay. The response actually flags that tension, which is why its advice to request the signed schedule is useful. Scoring the costs-page terms as supported does not settle which arrangement would apply to every family. Factual care sometimes means carrying two consistent observations about what a company says without pretending that public pages implement the company’s billing system. A definitive statement about one family’s agreement needs that agreement.

A reporter can use this baseline to illustrate failure mechanisms: confusing providers with similar names, failing to carry contract conditions through a summary, and combining an actual eligibility requirement with an unsupported additional hurdle. The evidence is much thinner for explaining causes or prevalence. There was no random draw of alternatives, no neutral task in which brands could appear spontaneously, and no user-behavior measurement. The safest headline would name the affected products, the singular test and what was wrong. A league table about all AI tools, all admissions consultants or which brand an algorithm favors would outrun the observations.

Field notes for independent adjudicators

Field 1. Identify the seller before comparing the amount

A college consulting business can appear under its trading name, an abbreviated name, or the name of a founder whose biography is cited elsewhere. In this case the minimum evidence for attributing a price is a page that identifies the seller and the purchasable service together. The family supplied a complete name and asked about senior year. A catalog organized by school count at a different named business gives no basis for concluding that the Institute sells that catalog. Even a shared geographic market, similar vocabulary and overlapping customers would leave that step unsupported. The record includes both primary pages because the incorrect answer is more readily verified by showing where the wrong numbers came from.

A practical double check is to read the answer from the perspective of someone about to fill in a purchase order. Could that person write the supplier, term of service and payment obligation on three separate lines? The Gemini answer gives the right supplier in its heading and an identifiable service in its body, but those lines belong to different offers. This is a contradicted field, rather than simply a missing citation. If a later page happens to advertise a price near the one in the answer, record the new date. A numerical coincidence would not retrospectively validate the original product attribution. [I1, I3]

Field 2. Distinguish cash due now from conditional exposure

Many households begin a consulting comparison with an immediate budget constraint. A family that can pay a particular amount this month might be unable to afford an equal amount later. Another might accept the full price but prefer a limited initial commitment. A description such as installment plans available cannot answer either family’s question. On the two-payment costs page, the offered terms identify a first installment and say when the family becomes responsible for it. The fact sheet therefore records amount and event together. It does not estimate a specific calendar day, credit-card fee, tax treatment or financing term that the company has not supplied.

For a repeat wave, a coder should distinguish a change in payment mechanics from a change in the underlying fee. A family might pay an invoice in smaller transfers while remaining responsible for the same amount under the same conditions. Alternatively, a newly advertised contract might move a larger or smaller portion of the cost into a contingent payment. Both changes affect the family, but only the second changes the division of risk between family and firm. Capturing only a headline figure would make those changes indistinguishable. A dated screenshot or fingerprint of the reviewed page can help locate the disagreement, while the text still needs to be interpreted in context. [I1, I2]

Field 3. State the outcome condition precisely

A conditional fee needs a trigger, a defined set and a responsible party. On the reference costs page the relevant set contains eight colleges designated by the family before the first application submission. The trigger is an acceptance to at least one of them. A college outside the designated set is a different event. Likewise, an invitation to apply, an interview or a waitlist notice would not by itself establish an acceptance. Those examples explain how to interpret the published description; they do not purport to resolve every contingency in a private contract.

The AI answer need not anticipate every possible dispute to earn credit. It should give a family enough information to understand why an eventual invoice might appear. ChatGPT did that on the two-payment schedule. Gemini moved from installments to the general statement that admission is uncertain. That caution is reasonable on its own terms, but it does not say which outcome changes the amount owed. A routine search task can make general consumer caution sound like a substitute for terms precisely because both use the vocabulary of risk. In a future wave, preserve a direct statement of the trigger separately from contextual warnings. Two answers giving similar warnings may differ substantially in commercial usefulness. [I1]

Field 4. Work through the case in which none of the eight admits

An offer to waive a later charge has different cash-flow implications from an offer to return a charge already collected. A family considering the advertised two-payment plan initially parts with the first installment. If the defined admission outcome does not occur, the second amount is not due. The costs page describes that event and limits what it entails for the first payment. An answer promising a refund of that first installment would overstate the downside protection. Neither returned answer made that specific false promise. It belongs in the audit’s list of future stress tests and should never be published as a quotation from either system.

This is also a useful case for checking editorial fairness. A commercial sponsor might prefer headlines suggesting that all risk has been removed. Yet the first installment remains a meaningful cost for a family applying unsuccessfully to the selected group. Conversely, repeating only that the firm cannot guarantee acceptance obscures the part of its fee it says it puts at risk. A clear report states both halves without judging whether the protection is worth the price. The enrolled family can then evaluate its own preferences, other applicants’ experiences and any signed terms that differ from the public description. [I1, I2]

Field 5. Bound a finding that a price was not published

It is easy to verify a visible number. It is much harder to establish that no number appears anywhere a company has ever published. The IvyWise result is therefore phrased as a bounded search finding. The official counseling and initial-consultation materials reviewed for this audit describe individualized programs and a recommendation process. We did not identify a single price there that a reader could apply to every comprehensive engagement. A specific written proposal, an old press story, or a third-party review could contain a real transaction amount without contradicting that finding.

Gemini added a characterization of packages as running into the tens of thousands and upward. That range was not established by the particular official pages used for the core field. We did not count the characterization as a current firm-issued price. The core field asks whether there is one standard price, and both answers got the distinction right. This split illustrates why a favorable check mark and a caution can appear on the same answer. A broad statement about pricing method can survive verification even when a numerical embellishment lacks sufficient first-party support. A journalist interested specifically in price ranges should make that a separate scored field before the next expanded test. [I4, I5]

Field 6. Distinguish a list of services from a package entitlement

A website can describe tutoring, essay guidance, interviewing and list development in neighboring sections. Their proximity tells a prospective customer what the organization can do. It does not tell each customer which services, hours, people or revision rounds a particular agreement includes. IvyWise describes an initial consultation with a profile review, a meeting and a plan. Its continuing advising has a different scope. Both responses identified real categories of work, which was sufficient for this field. When a family requests an all-in quotation, the relevant next question becomes how those categories are packaged together.

An adjudicator should look for language that changes an offered activity into a blanket entitlement. Consider a hypothetical answer stating that every comprehensive client receives unlimited tutoring. Even if the firm describes both comprehensive counseling and tutoring, the word unlimited introduces a new factual claim. A direct statement of scope in the relevant offer would be needed. In this baseline we retained the service description and set aside any inference about what every family purchases. A quote is most comparable when it identifies the number of schools, duration of contact, work included in the base fee and conditions under which a new charge applies. [I4, I5]

Field 7. Preserve the unit attached to free support

Ten hours of free virtual individual guidance is a specific benefit. The number can be lost in two directions. A family may imagine a yearlong unlimited relationship, or may assume that a free program offers only videos and group webinars. The source identifies one-on-one time with a counselor. Both systems communicated that distinction. The audit does not convert ten hours into a private-market dollar value. It would take pricing data for equivalent work, evidence about how the hours are used and a defensible reason for treating the services as interchangeable to estimate such a value.

The amount of direct contact matters, but so do the student’s tasks between meetings. A family comparing ten hours to a comprehensive package might decide that it needs help with only one personal statement, or that its schedule requires broader project management. Neither the source nor the two AI answers can decide which situation describes that household. The score gives credit for correctly stating the published benefit and leaves the family’s judgment outside the factual denominator. This rule protects an inexpensive or free option from being dismissed merely because its format differs from a commercial service. It also avoids implying identical coverage where the provider makes no such claim. [I7]

Field 8. Specify both residency and financial criteria

A student’s location and citizenship are different pieces of information. The program’s eligibility description uses US residence and expressly welcomes undocumented students. A statement limiting eligibility to citizens would narrow the offer without support. A statement promising global availability would expand it without support. The fixed prompt says the family is in the United States, but a student might be studying abroad, living with another guardian or simply asking on behalf of someone else. An answer about eligibility should ideally retain the criterion even when the question appears to satisfy it.

The program also refers to household income guidelines and ways to document financial need. Low income is a useful description of the target group; it is not a decision rule that lets every reader decide eligibility immediately. The compact Gemini answer communicated the target population while leaving out two operational details. Partial credit acknowledges the useful part. Printing an income threshold from memory could introduce a more serious mistake, especially where thresholds depend on household size or documentation changes. A careful response can link the current guideline and explain which part of the decision the family still needs to make. It need not request that the family disclose its exact income in a public chat. [I7]

Field 9. Tell a student which academic evidence is required

The reference program specified an unweighted GPA threshold and requested transcripts. Gemini described the GPA in approximate terms and introduced test scores as a typical attribute. It did not present a numerical SAT or ACT cutoff. The narrower description of the error matters. A published correction should remove the unsupported implication that test performance is part of the stated screen while accurately preserving what Gemini said. Quoting an invented cutoff would create a second misinformation problem in the process of criticizing the first.

For students at schools that report different grading scales, an answer can explain that the public rule is expressed on an unweighted basis and suggest checking the program’s own application instructions. It should not invent an automatic conversion. Similarly, a requirement to submit a transcript says little about how an edge case is resolved. This is why factual sufficiency is evaluated against what could be known publicly, rather than whether the answer settles every household’s eligibility. A cautious answer that correctly identifies a documentation boundary is more usable than an apparently decisive prediction built on an unstated formula. [I7]

Field 10. Check the intake notice after reading the evergreen rules

Many programs keep descriptions of a normal application process online between intakes. Visitors need those pages to understand the opportunity and prepare for the future. An AI system therefore has two sources of information on the same page: stable descriptions of who can apply and a dated notice about the present cycle. In the baseline the Matchlighters page described a 2026 capacity closure and directed interested readers toward a separate form. ChatGPT preserved the distinction. Gemini described benefits without saying that immediate intake was closed.

We did not infer that Gemini explicitly told the family to submit an application that day. Its error on this field is an omission. The appropriate severity depends on use. A student planning well ahead could still benefit from an evergreen description. A senior making decisions this week might spend limited time assembling materials for a closed opportunity. To preserve both interpretations, the report shows the score with the field included and explains the alternative denominator when it is removed. If a future wave occurs after reopening, the current answer changes even when the evergreen eligibility rules remain identical. [I7]

Field 11. Verify an individual’s role without inferring a career history

The family asks a logical question before asking a biographical one. Does a statement that a firm employs former admissions staff imply that every counselor personally held that role? No such conclusion follows from the statement alone. Some firms make a narrower or more explicit claim about an identified group of counselors. IvyWise describes a particular team and publishes individual profiles. Its website also identifies other roles, such as tutors and research staff. The evidence for a named counselor should come from their own profile or an independently verified employment record, where appropriate.

Both answers offered workable next steps. A family can ask who will lead meetings, who reviews essays, which person handles a late change and whether assignments can change midyear. It can ask the firm to confirm the title and dates behind an admissions-office claim. A professional profile is a starting point; it does not convert all prior university employment into undergraduate file-reading experience. The audit does not investigate private personnel or rate employees. It tests whether a consumer can differentiate organization-level marketing from a fact about a particular service relationship. The distinction survives even when every named person on a specific counselor roster genuinely has admissions experience. [I6, I9]

Field 12. Identify who holds decision authority

A university’s admissions process involves considerations an outside organization cannot settle for an applicant. Harvard’s own description emphasizes a range of individual and contextual factors. The IECA consumer guide also urges families to question promised outcomes. Those sources support a narrow answer: an independent consultant cannot issue a Harvard acceptance on the institution’s behalf. A firm might still market a contingent payment, an additional period of work or another contractual remedy. Those are terms about what the firm will do with its own services or invoices.

The coding does not classify every firm that uses the word guarantee. A reporter who wants to test a specific guarantee should quote the contractual object precisely and ask which party can perform it. Reaching one among eight named schools, returning money after a defined event, and gaining entry to a specific university are three different propositions. In the baseline both products communicated the university’s decision authority, though their links and explanations differed. A family could use this part of either answer while still needing to correct the financial terms in the same response. The evidence supports an assessment at that level of detail. [I1, I8, I9]

What changes when the same question is asked again?

A repeated audit needs two clocks. The first records when the product answered. The second records when a provider’s description took effect, if the source supplies that date. A difference between two answers can be evaluated only after checking both clocks. Imagine that a program reopens between two Sundays. An answer that says intake is closed could be supported at baseline and outdated at follow-up. A later answer that still describes closure is evidence of staleness if the reopening was visible at the time of that later test. Without a dated record of the reopening, the analyst should describe disagreement and avoid assigning an exact date of change.

The planned follow-ups are November 15, 22 and 29, 2026, at 1:45 p.m. Eastern. Preserve the complete five-question prompt, its wording and its order. Each wave should use fresh conversations and record the visible session condition. Reusing a thread in which a collector has already supplied the correct price would turn the next wave into a test of correction retention. That is a worthwhile experiment but answers a different question. It could be run in separate labeled threads with a fixed corrective statement. Mixing those threads with the fresh interviews would bias apparent improvement.

For Google Search, record one of three observations before scoring content. An Overview may be visible; a usable results page may have no Overview; or the collector may be unable to reach a results page. For the first state, retain the generated text and linked pages. For the second, the absence itself is relevant to feature availability but contributes no claim accuracy score. For the third, record the reason and stop at the barrier. Do not relabel a Gemini answer, a conventional search snippet or another search engine’s summary as a Google Overview. AI Mode, if observed separately, receives its own row and interface label.

A new wave also needs a product-change register. The displayed model or mode might change, a provider might require sign-in for a feature that was previously public, or a new display might expose sources only on demand. Such changes affect the population of answers the audit can capture. They should be described before a percentage is calculated. If collection shifts to an authenticated profile, give the new condition a fresh label. An apparent improvement might reflect access to a different product tier or personalization. The initial US-family wording standardizes the scenario, but it cannot by itself prove a US geolocation, local search index or universal English-language result.

Report changes in a small transition table. An omitted term can become supported, a correct number can become stale, and a previously absent surface can produce its first response. Include the baseline and follow-up wording that drove each changed code. If the source remains the same, a stable hash can support a statement that the retrieved bytes matched, though site templates can change hashes without changing the underlying policy. If the bytes differ, compare the passage that matters rather than declaring that the policy changed. The source register is an aid to investigating change, not an automatic policy detector.

Where the same error recurs, count the number of observed waves containing it and list the dates. Avoid extrapolating from three recurrences to all members of a product’s user base. Where an error disappears, say how many later successful responses addressed that field. A later omission is not a correction just because the erroneous price is gone. Where the product answers correctly but points to a weak source, mark the factual improvement and the remaining verification problem separately. A single scalar score would lose this useful information about what changed.

A source can favor the sponsor in more than one way

The clearest favorable error in the baseline concerned affordability. If readers accepted the lower range as the Institute’s fee, they might think a senior-year engagement was within a budget it exceeded on the published two-payment schedule. That is a plausible direction of misperception, not a measured marketing effect. The same response omitted the waiver of a conditional second payment, potentially making the firm seem less willing to share risk. The research record therefore attaches direction to an individual claim. Assigning a single pro- or anti-company sentiment to the entire answer would not capture what a family might misunderstand.

The source pages deserve scrutiny as well. An Institute cost page gives a two-installment example while its homepage describes division into up to five scheduled payments and refers to a final scheduled payment. A critic might infer that the company must have changed its contract or that one page is outdated. The pages establish a difference in the public descriptions. They do not establish the chronological order of the policies or the wording of every signed agreement. Asking the firm to clarify which portion is at risk under each schedule would resolve more than inferring a policy from which page appears first in search.

The Institute’s costs page also uses a scholarship-return illustration. Such material provides a useful future stress test even though the questionnaire did not ask about scholarship returns. An AI answer could divide a stated multi-year scholarship amount by the consulting fee and call the result the return caused by paying the firm. Arithmetic alone cannot support that causal conclusion. A family might have earned some or all of the same aid without private advising. Renewal conditions and the nature of the award would also matter. This is a hypothetical favorable error, deliberately separated from the recorded baseline. It helps define what a future audit should ask without inventing criticism of either captured response. [I1]

One can make the same mistake in the other direction. A reviewer may see that an average is self-reported and conclude that it is false. Self-report identifies the source and limits verification. It does not establish that the number is wrong. An appropriate answer can attribute the statistic, state which denominator has been disclosed, and identify missing information. If the company supplies more complete data later, a designated reviewer can examine the updated material. An audit remains credible when it applies the same distinction to every provider, including the organization that funded the work.

Four ways to check an answer before using it

Reconstruct the family’s next decision. Consider a parent comparing the cost of beginning service next week. The parent needs the amount due at enrollment, the time covered, and the consequences of ending early. The trial prompt tested the first of these directly and a separate admissions outcome. It did not ask about voluntary cancellation. A family should not infer a cancellation right from an outcome-based payment waiver. A useful AI response can identify that distinction and point to the appropriate service agreement. A fact-checker should mark the boundary of the question before characterizing the response as complete.

Read the issuing page from its heading to its qualification. A recognizable logo and a matching number are insufficient. Which program, grade, geography and year does the page actually cover? Is the amount per hour, per school, per year or for a package? Did the author say the benefit is awarded, renewable, available to apply for or already at capacity? The wrong-company price would have survived a check that only matched numerals. Checking the page title and seller next to the amount would have exposed it immediately. This is a cheap first test that can prevent much more elaborate but irrelevant verification.

Separate the service description from the comparison judgment. A firm can reasonably say it provides meeting notes. It usually cannot establish from that statement alone that its notes are more useful than another firm’s. A family that values written follow-up can ask both providers for a sample with private information removed. A search tool may still help generate the comparison questions, even when it cannot supply the answer about a particular student. The presence of a sensible recommendation does not make an accompanying price correct; neither does a price error invalidate every general question in the answer.

Keep track of what remains unknown. A useful short list might include the fee under the family’s proposed start date, the number of schools included, the credentials of the person actually assigned, and availability at the free program. Each item should identify who can resolve it and which document would count. This makes uncertainty actionable. A list of every conceivable caveat can make an answer unusable, while a short list of material uncertainties can help the family make a better inquiry. The audit’s fields are meant to model that discipline. Their value depends on choosing uncertainties that could change a real decision.

A reporting standard for corrections and comparisons

A correction should name the claim, affected offer, observed date, source and consequence. For this baseline, the strongest correction is that the cited three- and ten-school prices belong to The Ivy Mind, whereas The Ivy Institute describes a different conditional schedule. It is unnecessary to speculate that the system deliberately favored the commissioner. The mistake is visible without a theory of motivation. It is equally unnecessary to imply that The Ivy Mind misrepresented anything. Its published fees were appropriate evidence about its own offer. The misattribution occurred in the generated answer.

An editor should preserve the original observation even after issuing a correction. Replacing the baseline text with a corrected paraphrase destroys the evidence that generated the finding. In the research bundle the original responses sit beside a separate coding sheet. A future correction to our interpretation would change the coding record and add a reason, while leaving the response capture intact. This separation matters if a firm updates a page, asks for clarification or points out an overlooked qualifier. A reader can then reconstruct two questions independently: what the product said, and why the study gave it a particular code.

If a publication wants a broader estimate, it should define a sampling frame before choosing examples. One approach would draw anonymized question types from consenting users and sample across price, program, counselor biography, scholarship eligibility and verification of outcomes. Another would sample firms first and issue a neutral, matched question about each. The former weights common family needs more heavily. The latter permits comparisons across companies but can give small niche offers disproportionate weight. Neither sampling scheme follows automatically from the present benchmark. The compact fixed interview is suited to watching a few consequential fields change over time.

A larger study would also vary prompt order. Putting the conditional payment question first might anchor the reader or the model on commercial risk. Putting the free scholarship program first might cause an answer to emphasize budgeting and eligibility. Separating firm names across prompts could reduce identity confusion while also making the task unlike a real comparison. A defensible extension would randomize both single-question and bundled formats across separate sessions and report each condition separately. The baseline is worth retaining unchanged alongside such experiments because changing prompt design during a time series can make a design change appear to be a product change.

The next useful quality measure is traceability. For each factual sentence, count whether there is a link attached, whether the link resolves to a page, whether that page is primary for this claim, and whether the relevant passage supports the sentence. These are distinct stages. A fully correct statement with a dead link creates work for the reader. A living link to the wrong seller can be more misleading because the reader sees apparently authoritative evidence. An accurately attributed secondary review may be appropriate for a user-experience question and inappropriate as the final authority on current payment terms. A single link count cannot distinguish those cases.

Same-day repeat: what held and what changed

At approximately 2:03 p.m. Eastern the collector preserved a second pair of complete answers, about seventeen minutes after the first pair. The same text was sent in fresh signed-out conversations. This repeat adds a direct check on within-day variability before the scheduled weekly follow-ups. Conditions were similar by observation, but shared browser infrastructure and unknown backend behavior prevent treating the two runs as independent samples of all users. No collector supplied the correct price to either conversation. Google Search was not retried after its verification barrier. The repeat captures are available in the supplement under their own time stamps.

Gemini’s second response dropped the three-school and ten-school menu. It described senior-year pricing as depending on individual terms and supplied no amount. Removing incorrect numbers reduced the risk that a reader would budget using another company’s plan. It did not supply the correct starting payment, full scheduled fee or event that determines the second payment. Moreover, the repeat introduced a suggestion of possible partial or qualified refunds. That wording is conditional and does not promise that every family receives money back. In the context of a question specifically about receiving no offers from the designated eight, however, it points the reader toward a remedy that the dedicated costs page does not describe. On the published two-payment terms the first installment stays paid and the second is not owed. [I1]

A public Institute article helps explain why even first-party retrieval could surface the wrong remedy. A company article dated July 18, 2023 advertised a money-back guarantee tied to a top-choice outcome. The currently reviewed costs page expressly distinguishes its second-payment arrangement from a refund of the initial installment. A separate March 30, 2026 post refers to a waitlist-specific offer, which is another service and stage. These sources document language a reader can encounter. There is no evidence the model actually consulted either post in producing its repeat answer. The answer linked the FAQ. A forensic account should report the source conflict without inventing a causal history for the model’s mistake. [I1, I15, I16]

Gemini’s Matchlighters section improved on several details. It specified an unweighted GPA, mentioned financial-need criteria and ceased suggesting test performance as an entry signal. It still did not report the capacity closure. Its characterization of US high school students was directionally useful for the family in the prompt but narrower than the full range of residents the program describes. Its suggestion that an IvyWise initial consultation starts above a particular dollar threshold was associated with a review rather than a verified current firm-issued minimum. The official page does support the ninety-minute duration. In the record the duration and the unsupported price floor therefore receive different annotations. [I4, I7]

ChatGPT’s repeat showed another kind of instability. An early paragraph said a sufficiently reliable current quotation had not yet been established. Later in the same completed answer, it gave the correct costs-page amount and both payment conditions. Scoring the operative final terms supports the financial decision fields, while the inconsistent opening remains a presentation defect. The answer also called the homepage’s installment language older. The two pages could be verified side by side, but their relative policy dates were not established from that wording alone. We retain that temporal assertion as unverified. The second answer’s named counselor tiers can be checked separately on IvyWise’s tier-description page. A correct factual core can coexist with unnecessary or unverified elaboration. [I1, I2, I17]

The scale of the two observed changes matters more than a leaderboard. The false low range appeared in one of two Gemini captures. The other capture made the outcome policy sound potentially more generous through refund language. ChatGPT retained the important price terms in both captures, but the second took a more internally inconsistent route to the answer. Four responses reveal enough variation to invalidate the proposition that the first answer is what a product always says. They remain too few to estimate the frequency of any mechanism. The field-transition file records exactly which facts changed under the fixed rubric.

Independent evidence on verification

Liu, Zhang and Liang evaluated four generative search engines in a separate 2023 study. They reported that only 51.5% of generated sentences were fully supported by citations on average, while 74.5% of citations supported their associated sentence. Those two percentages have different denominators. They illustrate why finding links and finding sufficient support must be assessed separately. The four products and questions in that study differ from the present consumer interfaces and admissions questionnaire. Its numbers are evidence of a measurement problem and its historical magnitude in that sample. They are not a prior error rate to add to the four admissions responses. [I18]

The EBU and BBC examined AI responses to news questions across multiple languages and participating public-service media organizations. Their 2025 report found a significant issue of some kind in 45% of the responses examined. A significant issue in their framework need not be a factual error of the type classified here. Source attribution, missing context and presenting opinion as fact all matter to a newsroom. In our audit a family can receive a technically true statement about a business that is useless because it describes the wrong program. The external news study provides an independent reason to track context and source quality. A combined percentage across the two studies would be mathematically easy and substantively incoherent. [I19]

An alternative fact-checking method would break every response into atomic factual assertions, as in the FActScore framework. That design has the advantage of tracking unsupported additions even when they concern facts the user never requested. The twelve decision fields instead penalize leaving an important requested fact out. These approaches illuminate different weaknesses of the same repeat answer. A short response that supplies almost no specific numbers might look comparatively safe in a factual-precision analysis. From the family’s perspective it may still fail because there is no price from which to plan. An expanded audit should publish both denominators side by side and retain open-ended advice outside the factual total. [I20]

What an independent editor could verify from these data

An editor can reproduce the price error without trusting the Institute's assessment of its own quality. Open the two source pages identified by the record and identify the name and service printed above each amount. Then compare the numeric strings in the answer with the named seller in the question. The same two amounts appear together in the other company's school-count menu. This test does not depend on whether the two providers deliver comparable work, whether the Institute usually gives discounts or whether either company deserves its advertised price. It examines the relationship between a statement and the public offer it purports to describe.

Verifying the no-offer remedy requires a different procedure. There are competing types of first-party language, and a reviewer who searches only for a phrase such as money back could retrieve precisely the evidence most likely to support the wrong interpretation. Begin with the product and cycle in the family question, then read the current named cost schedule and its qualifications. Next inspect the homepage and historical posts for scope or date differences. Report which document would actually govern an invoice. In this setting, public information is sufficient to contradict a generic refund promise under the advertised two-payment arrangement. It is insufficient to advise an individual family about a signed agreement that was not supplied.

The independent editor can also check an answer with no obviously false number. Consider the suggestion that every family should compare how many students a counselor serves. As advice, this is reasonable to discuss. As a factual statement that a smaller caseload necessarily improves application outcomes, it would need independent outcome evidence and attention to selection. Neither response demonstrated such a relationship. The report therefore records the recommendation under decision support and avoids turning it into a finding. An audit made entirely of checkboxes might inadvertently award a commercial recommendation the status of scientific evidence simply because it reads like good professional practice.

For the free program, the editor should read its eligibility and capacity notices together. Reconstruct two hypothetical readers. One is a US-resident senior with an unweighted GPA above the threshold who lacks test scores; the other is a high-scoring applicant residing abroad. The first should not be excluded solely for lacking scores on the evidence reviewed. The second cannot assume access solely from academic attainment. Neither can assume an immediate place during the observed closure. These are tests of what follows from the published rules, not determinations of individual eligibility or assessments of who would be a good counselor match. They help reveal errors that an ordinary keyword check could miss.

There is another independent check on how the study describes its own limitations. A reader can reconstruct the strict denominator using the companion ledger without reading a single sentence of the discussion. Select the November 8 baseline, filter to each returned product, and count the twelve named fields. A missing-surface record must yield no twelve-field row. Then compare the same-day repeat against those exact fields and inspect the notes on additional assertions. If the resulting account showed one platform’s errors in detail and summarized the other with an unqualified perfect score, a reader would be justified in questioning the presentation. The repeat’s contradictory opening is retained precisely because it would otherwise vanish beneath a favorable field score.

The search for historical product-specific evidence included publicly indexed examples of admissions-consultant answers and shared chats. It did not establish a historical response to this identical questionnaire on all three products. Publicly shared chats reflect what their authors chose to share, and a screenshot without an underlying prompt or capture date cannot establish the product response under this design. Search results for the phrase AI Overview sometimes lead to articles written to attract such a summary rather than actual captures of one. These materials were excluded from the accuracy denominator. The decision prevents an ordinary consultant’s webpage from being counted twice, first as an AI answer and again as the source used to validate it.

A broader literature search can improve a study without increasing the number of directly comparable observations. The independent audits cited here cover news attribution, browsing behavior and factual support, each with a distinct outcome. Their sizes give context about why errors deserve attention, while their designs limit how directly their numbers travel. Likewise, an extra secondary source about a family’s experience can illuminate customer service but cannot make a wrong-company price right. Evidence coverage is therefore described by source role as well as count: consumer answer, seller, program operator, university, professional body or independent research team. Large source lists are useful only when a reader can tell which question each source actually resolves.

For corrections, distinguish an old offer from an old observation. A source may genuinely have offered a refund in an earlier cycle. A family asking for current terms needs to know the current arrangement. But if a journalist is documenting what a family saw in the earlier period, today's cost page is not the appropriate historical truth standard. The audit’s access dates and immutable answer records make it possible to retain both. A later confirmation from a firm should be dated and attributed, then compared with the contemporaneous public evidence. It should not be silently inserted into the original answer or used to erase an inconvenient discrepancy in what the firm itself had published.

The public significance of the case lies in the burden shifted onto families. Choosing a service requires bringing a seller, service, price unit, academic cycle and contract condition together. Generative summaries can perform much of that assembly in a single readable answer. They can also obscure the join between two unrelated documents. In this pilot the wrong-company source remained inspectable, and the better answer gave enough information for a family to identify a genuine ambiguity. The practical test for the next wave is whether the answer leaves the family with a verifiable next step. A heavily cited paragraph that points toward the wrong supplier fails that test even if its arithmetic is impeccable.

Reader’s interpretation

This pilot offers a clear answer about the session it observed. One system correctly conveyed the most important payment details and the free program’s capacity notice. Another supplied two genuine prices under the wrong company name and introduced an unsupported academic signal. Both produced some helpful advice. The family could have corrected much of the discrepancy using a small number of first-party pages. It could not use those pages to decide which consultant would deliver a better admissions outcome for a particular applicant.

Two limitations remain central when the report is cited. First, a named brand appearing in an answer was an explicit feature of the prompt. Such an appearance is no evidence that the product independently recognizes the firm as better, more popular or more effective. Second, this single day captures particular interfaces, English wording, location uncertainty and a signed-out state, each of which may matter. Repetition over the scheduled waves will produce evidence about stability for this same interview. Even a perfectly stable month would not turn this selected pilot into a representative survey of searches.

The practical achievement is narrower and more durable. The audit identifies a financial misattribution a family can verify, shows how to distinguish a missing condition from a false assertion, and supplies a repeatable record of both success and failure. A future researcher can ask precisely where the answer changed, what the publisher said at the time and whether a change helped a household act. Those questions are better matched to the evidence than a permanent ranking of products.

Documentary appendix: the family-facing responses

The following transcripts are primary observations. They are reproduced as generated output for audit and may contain errors described above. Paragraph spacing and table layout are normalized for reading. Displayed source destinations and the original plain-text captures accompany the report. The model’s first-person recommendations are part of the observation; they are not recommendations by the study author.

ChatGPT · November 8 baseline

College admissions consulting comparison

The three options have quite different models: The Ivy Institute publishes a conditional senior-year fee, IvyWise offers individualized counseling packages, and College Essay Guy's Matchlighters program provides free assistance to qualifying students.

Here is a comparison based on their official websites, checked November 8, 2026.

1. The Ivy Institute: senior-year price and payment terms Published price: $25,000

According to The Ivy Institute's official pricing and payment page :

Payment | Amount | When due First payment | $12,500 | At the beginning of the senior-year engagement Second payment | $12,500 | Only if the student is accepted to at least one of eight selected colleges

What happens if none of the eight colleges admits your student?

The second $12,500 is not due.

Important details:

The family chooses any eight colleges before the first application submission.

The commitment applies to the second payment, not the first.

The initial $12,500 is not refunded if the student receives no acceptances among the eight.

The firm says its team remains involved through applications, interviews, waitlists, decisions, and enrollment support.

One contractual detail worth checking: The firm's homepage  also describes the possibility of dividing the senior-year fee into as many as five scheduled payments, depending on when the family starts. The pricing page describes the two $12,500 installments. Ask for the precise installment schedule in your enrollment agreement.

My interpretation: This is a meaningful financial distinction, but it is not an admission guarantee. You are still paying $12,500 regardless of the outcome.

Primary source: The Ivy Institute – Costs and Payment Terms .

2. IvyWise: comprehensive counseling price and services No single standard price published

IvyWise does not appear to publish one universal price for comprehensive college admissions counseling.

Instead, it describes individualized programs and asks families to go through an initial consultation and recommendation process to determine the appropriate services and investment.

Its official initial consultation page  describes:

An initial review of the student's academic and extracurricular background.

A 90-minute consultation.

A customized admissions strategy and action plan.

A personalized profile-building catalog.

A testing plan, where applicable.

Recommendations for ongoing counseling.

Its comprehensive admissions counseling page  describes additional services:

College selection and application strategy.

Personal statement and supplemental essay brainstorming, drafting, and editing.

Application preparation and review.

Mock admissions committee feedback through its Roundtable process.

Interview and application support.

Deferral and waitlist consultations.

IvyWise also advertises more than 440 years of combined admissions experience among its counselors.

What to ask: Request a written, all-inclusive price covering your entire application cycle. Specifically ask whether essay revisions, supplemental essays, meetings, and all selected colleges are included or charged separately.

Primary sources:

IvyWise – College Admissions Counseling 

IvyWise – Initial Consultation 

3. College Essay Guy: Matchlighters Scholars Program 10 hours of free counseling

The College Essay Guy Matchlighters Scholars Program connects qualifying students with experienced college counselors.

A selected Matchlighters Scholar receives:

10 hours of free, virtual, one-on-one college admissions counseling.

An assigned professional college counselor.

Help with the college application and admissions process.

Individualized support based on the student's needs.

Who qualifies?

According to the official eligibility requirements:

Must be a U.S. resident. Undocumented students are welcome.

Must be pursuing college or graduate-school admission.

Must have at least a 3.0 unweighted GPA.

Must meet the program's family-income guidelines.

Must demonstrate motivation and willingness to work.

Applicants must provide transcripts and financial documentation.

Examples of qualifying financial circumstances include participation in free or reduced-price lunch programs or other government assistance programs.

Current availability: The official website says the 2026 application window is closed because the program reached capacity in early September. Families can submit an interest form in case space becomes available.

My interpretation: This is particularly attractive if your student meets the financial eligibility requirements. Ten hours of individual counseling can provide meaningful help, although it is not necessarily equivalent to a full-service, year-round admissions consulting package.

Primary source: College Essay Guy – Matchlighters Scholars Program .

4. Does every counselor at a firm have admissions-office experience?

No. A firm's marketing statement should not automatically be interpreted as proof that every person working with your student has personally worked in a college admissions office.

There are important distinctions between:

Former admissions officers.

Former admissions deans or directors.

Independent college counselors.

Essay specialists and editors.

Academic tutors.

Current college students serving as mentors.

For example, IvyWise explicitly states that its counselors are former admissions deans and directors. It also publishes individual counselor biographies identifying their previous admissions positions.

You can examine its official counselor directory .

Some examples of what its biographies disclose include:

Former Assistant Director of Admissions at Yale.

Former Admissions and Financial Aid Officer at Harvard.

Former Admissions Officer at Occidental College.

However, you should still verify the individual who will actually work with your student.

Five questions I would ask before signing any contract:

What is the full name of the primary counselor assigned to our student?

Which colleges or universities employed this person in an admissions capacity?

What was their actual job title, and during which years?

Did they personally read, evaluate, or vote on undergraduate applications?

Will this specific person personally oversee our student's essays, college list, and application strategy?

Also ask whether the named counselor will remain assigned throughout the entire application cycle.

A firm's collective experience is not necessarily the same as your student's counselor's individual experience.

5. How should you compare these firms? Can anyone guarantee Harvard admission? No legitimate admission guarantee

First, consider what you are actually purchasing.

Comparison table is retained in the exact-text record in the companion data package.

My recommended evaluation criteria

I would compare these firms using six criteria, rather than simply choosing the one with the strongest Ivy League marketing.

1. Counselor qualifications

Prioritize the actual person assigned to your student, not the firm's combined years of experience.

2. Counselor continuity

Find out whether one experienced person owns the strategy throughout the application process.

3. Essay involvement

Determine how many revisions are included, who edits the essays, and whether the student's authentic voice is preserved.

4. College-list strategy

Ask how the counselor balances reach, target, and likely-admission schools. A list consisting entirely of extremely selective universities is not a responsible strategy.

5. Total cost

Compare the entire application cycle, including additional meetings, editing, supplemental essays, tutoring, and other potential charges.

6. Accountability

Get the service commitments, refund provisions, counselor assignment, and payment terms in writing.

Can any of these firms guarantee admission to Harvard?

No. None can legitimately guarantee admission to Harvard.

Harvard's own admissions office explains:

There is no formula for gaining admission to Harvard.

Its admissions committee considers academic achievement, extracurricular distinction, leadership, personal qualities, character, and other factors.

Harvard also makes clear that it evaluates applicants individually.

Sources:

Harvard – What Admissions Criteria Do You Use? 

Harvard – First-Year Applicants 

Be particularly cautious of consultants who:

Promise admission to Harvard or another Ivy League institution.

Claim to have special influence over admissions committees.

Suggest that former admissions employment gives them the ability to secure an acceptance.

Guarantee particular outcomes without clearly explaining the contractual limitations.

A consultant can improve the quality, organization, and presentation of an application. They cannot control the university's decision.

My overall recommendation

I would approach the three options differently:

The Ivy Institute: Worth investigating if you want an extensive, coordinated service and value the conditional second payment. At $25,000, I would scrutinize the contract, counselor qualifications, and actual service delivery.

IvyWise: Worth considering if you particularly value former admissions officers' firsthand experience. I would obtain a detailed written proposal before comparing its cost with The Ivy Institute.

College Essay Guy Matchlighters: An excellent opportunity if your student qualifies financially. The free 10 hours could provide valuable support without a major consulting expense.

One final suggestion: Before spending $12,500–$25,000, I would interview at least two actual counselors, not just enrollment representatives. Ask each to evaluate your student's current academic and extracurricular profile and explain what they would prioritize over the next 6–12 months.

The quality of that conversation may tell you more than any firm's marketing materials.

Gemini Flash-Lite · November 8 baseline

Here are brief answers with primary-source links for each question:

The Ivy Institute (Senior Year): Pricing varies by tier (e.g., packages span from roughly $7,599 for a 3-school plan up to $14,599 for a 10-school plan). Installments split the single package price based on signed terms. Admission is never guaranteed, and terms specify that outside costs are separate. The Ivy Mind + 1

Primary Source: The Ivy Institute FAQ / The Ivy Mind Senior Year Plans

IvyWise Pricing and Services: IvyWise does not publish one standard price; comprehensive multi-year packages and custom programs range widely (from tens of thousands of dollars upwards) and require an initial consultation. Services cover full-service college admissions counseling, college list development, essay review, application strategy, tutoring, and early academic planning. Top College Admissions Consultants

Primary Source: IvyWise Services & Getting Started

College Essay Guy Matchlighters Scholar: High-achieving, low-income students (typically with a 3.0 GPA or higher and solid test scores) receive 10 hours of free, one-on-one college application and essay mentoring from an experienced volunteer counselor. CollegeXpress

Primary Source: College Essay Guy Matchlighters Program

Former Admissions Officers & Verification: No. Firms that market "former admissions officers" employ a mixed roster; individual counselors may instead be high school guidance counselors, Ivy League alumni, or educators without institutional admissions-office experience. You can verify a specific person's background by asking the firm for the counselor's formal CV, checking individual staff bios, or requesting their specific employment history in writing prior to signing a contract.

Primary Source: Top College Admissions Consultants Firm Evaluations

Comparing Options & Harvard Admissions Guarantees: Families should compare firms based on counselor credentials, transparency of pricing, student-to-counselor ratios, and whether support is holistic or essay-only. No legitimate independent consultant can guarantee admission to Harvard or any other highly selective university, as admissions decisions are ultimately up to the institution's internal committee.

Primary Source: The College Investor Admissions Consultant Cost Guide

Disclosure and availability

The Ivy Institute commissioned the report and is named by the test prompt. The research and initial adjudication were AI-assisted. No provider reviewed the coding on behalf of this study, and no independent journal review is claimed. The fixed prompt, dated responses, explicit source links, observation status and field-level decisions are included in the companion files. Future waves must be identified by their own collection dates.

References and primary records

  1. I1 · The Ivy Institute. Senior-year costs and payment terms. Source record.
  2. I2 · The Ivy Institute. Homepage and additional installment language. Source record.
  3. I3 · The Ivy Mind. Senior-year school-count packages. Source record.
  4. I4 · IvyWise. Initial consultation and recommended program. Source record.
  5. I5 · IvyWise. Admissions counseling. Source record.
  6. I6 · IvyWise. Team and individual counselor biographies. Source record.
  7. I7 · College Essay Guy. Matchlighters eligibility, benefits and 2026 capacity update. Source record.
  8. I8 · Harvard College. Admissions criteria. Source record.
  9. I9 · Independent Educational Consultants Association. Questions to ask an independent consultant. Source record.
  10. I10 · OpenAI. Searching the web with ChatGPT. Source record.
  11. I11 · Google. Information about Gemini responses. Source record.
  12. I12 · Google Search Central. AI features and websites. Source record.
  13. I13 · Columbia Journalism School, Tow Center (2025). Study of attribution in generative search. Source record.
  14. I14 · Pew Research Center (2025). Browser-panel study of clicks and AI summaries. Source record.
  15. I15 · The Ivy Institute. Earlier top-choice refund language in a public company article. Source record.
  16. I16 · The Ivy Institute. Separate waitlist-specific offer. Source record.
  17. I17 · IvyWise. Counselor designations. Source record.
  18. I18 · Liu, Zhang and Liang (2023). Evaluating Verifiability in Generative Search Engines. Source record.
  19. I19 · European Broadcasting Union and BBC (2025). News Integrity in AI Assistants. Source record.
  20. I20 · Min and colleagues (2023). FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation. Source record.
The Ivy Institute https://www.theivyinst.org
Previous

Does structured advising improve college application readiness?
Next

Does college admissions consulting improve student outcomes? Here’s what the research found.