The Ivy Institute | College Admissions Research | November 2026
Executive Summary
Two versions of a college essay can look almost the same length and tell quite different stories. We compared an earlier and a later personal statement from each of seven students. The middle result was a cut of 16 words. When we measured the size of each change without worrying about whether the essay grew or shrank, the middle result was 79 words.
A change in length also misses words that move or get replaced. Across the seven pairs, the middle share of matching individual words was 62.9%. For groups of three words in the same order, it was 23.2%. An opening image might survive while the student changes what the image means at the end. A parent reading a new draft can ask, “What do I understand about you now that I didn’t understand before?”
We could establish which version came later for seven students from the 2025 and 2026 application periods. Too few had both a confirmed first draft and an explicitly final version to publish a separate result for that group. We did not verify which words were ultimately submitted. The report includes the counting rules and a second calculation that leaves out an essay with an uncertain ending.
The Ivy Institute conducted and funded this study and publishes it as a provider of paid essay advising. The seven students had writing we could compare. We did not test whether advising or a particular revision improved anyone’s chance of admission. The Ivy Institute, internal materials, data, and information.
What the study asked
What changed between the two versions, and could a reader explain the difference? We looked at the actions the student described, what they made of the experience, the order of the story, repeated ideas, the selected prompt where known, and length.
We investigated six aspects of revision: specificity, reflection, structure, repetition, prompt alignment and length. The first four required reading. For prompt alignment, the necessary evidence was the exact selected prompt as well as the essay. Length could be measured directly once we established where each essay began and ended. We also asked how much wording changed beneath the net word count and how strongly our answer depended on a difficult decision about a draft ending.
The underlying comparison was within a student, within a single essay history. This design holds the identity of the writer and the broad purpose of the response constant. It does not hold the writer’s circumstances constant. Between versions a student might remember something, take on a new responsibility, receive suggestions, change their mind about disclosure, or rethink the audience. Our analysis can describe a change in the writing without deciding which of those events caused it.
For families, a narrower question is often more answerable than “which draft is better?” After reading each version, can a reader explain what the writer did and what the writer now understands? If the answers differ, the student can decide which interpretation they intend to send. A reader whose interpretation matches the student’s purpose can help the student decide when an additional editing round would have little to offer.
Which essays could we compare?
We compared one personal statement per student. Both versions had to contain enough of the same essay to read side by side, and we needed evidence of which came first. We left out outlines and suggestions that a student had not written into an essay.
The review started with 50 candidate application collections. Forty-seven appeared from their titles to concern individual student applications and were screened for earlier and later essay work. The other three did not meet that screening description. A separate search located 90 possible personal-statement or main-essay items. That search included duplicates, instructional material, professional or graduate-school writing, and several possible versions from the same applicant. The two search totals overlap and cannot be added to estimate the number of students served by the Institute.
We first looked for evidence of a shared application, then for the writer’s own extended prose. A title referring to a draft could lead to an outline, a collection of possible openings, several responses to different questions, an adult’s alternative, or a single complete personal statement. A single applicant could discuss the same idea in several places. We looked at the content and its relationship to the surrounding work before accepting a comparison. The purpose was to identify distinct stages of one response, rather than two pieces of writing that mentioned the same general topic.
We used a selected group of seven students with a traceable pair of complete or substantially complete personal statements. We selected one pair per student. A short response to a college-specific question, a set of bullet points, a rehearsal exercise or a reader’s standalone rewrite could not substitute for a personal statement. A student could have another perfectly usable essay outside the material we reviewed. Exclusion means the available material did not support this comparison. It says little about the student’s actual degree of preparation.
Within each pair, the earlier prose had to be distinguishable from instructions and comments. The later prose had to develop the same underlying response and be demonstrably later through explicit version sequence, dates or revision designations. A recurring narrative alone could establish a relationship but could not establish order. A later date alone could establish order but could not establish that two essays belonged to the same response. When different candidate endings occurred together, we recorded that ambiguity and repeated the central calculations with that entire pair omitted.
We treated first and final as separate questions. A writer might explicitly identify a first draft and then produce several revisions without calling any of them final. Another might have a designated final essay and an earlier dated draft with no evidence that it was the first. Each can support an ordered comparison. Both conditions must hold to support the stricter claim. We did not promote a version to final merely because its number was highest. In our screening, we found both competing final designations and later alternatives. We also found feedback or an outline mixed with draft text. Accepting those materials without checking would change the meaning of the results.
No student was contacted for clarification. A release based only on complete first-to-final histories would have been quite short and potentially identifying. We kept that strict analysis private and report a broader, explicitly labeled comparison here. This decision changes the inference substantially. The study can describe ways these statements evolved during identifiable periods. It cannot estimate the amount of change over the entire writing process for a typical student. All seven pairs refer to main personal statements. The results do not describe changes to short supplemental answers. The Ivy Institute, internal materials, data, and information.
How change was measured
For each student, we counted the words in both versions and subtracted the earlier total from the later one. That tells us whether the essay grew or shrank. We also counted the size of the change in either direction. Looking at both measures helps a family understand why a near-zero change can hide a much larger revision.
Read the detailed method and checks
These choices have small but measurable consequences. Consider the invented wording “we’d just finished a two-hour rehearsal.” A split at every punctuation mark would treat the contraction and the compound differently from our count. A student should use the college’s own entry field to verify a limit. For this study, consistency across the two versions matters more than matching any particular platform. We did not recode an essay as finished or unfinished based on its measured count.
For each student we subtracted the earlier count from the later count. A negative result means the later version used fewer words. Separately, we calculated the absolute difference, so a reduction of forty words and an increase of forty words both count as forty. The median is the middle value after sorting. In a set of seven pairs, three observations fall on either side of the middle position, allowing for ties. The essay at the middle position for one measure need not occupy that position for another.
Paragraphs were counted as nonempty blocks of essay prose. A deliberately isolated line counts as one block even if it contains only a few words. The count therefore captures a layout choice as well as a structural one. Two short paragraphs can become a longer one with little change in substance. Conversely, a student can keep exactly the same number of paragraphs while replacing the central example. We read the sequence and function of the paragraphs separately from counting them.
To measure wording, we asked how many word occurrences in one version could be matched with the same words in the other. We ignored capitalization and treated curly and straight apostrophes alike. If a word appeared four times earlier and once later, one occurrence could be matched and three remained unmatched in the earlier version. We repeated that operation for every distinct word. Reordering an identical passage creates no unmatched words under this rule. A change from a singular to a plural can create one unmatched occurrence on each side. The method therefore addresses vocabulary and frequency, with no direct measure of meaning or sequence.
For each pair, we report three quantities derived from this comparison. The first is the number of unmatched earlier words. The second is the number of unmatched later words. The third is the smaller of those two numbers. This last quantity captures changes in word use that the net difference in length alone cannot explain. If a student simply appends a paragraph and keeps all earlier words at the same frequencies, the third quantity is zero. If a student removes a paragraph and supplies a different one of equal length, the quantity may be large even though the net length change is zero. We make no claim about how many times a student pressed a key or edited a sentence.
A worked calculation
Take two invented sentences: “We painted the wall after lunch” and “We repaired the wall before lunch.” Both contain six words. Four occurrences can be matched: we, the, wall, lunch. Two words from the first sentence have no counterpart in the second, and two words in the second have no counterpart in the first. The net length difference is zero. Nevertheless, the action and its timing have both changed. A reader checking the revision with the student would reasonably ask which statement is accurate.
Now consider “I called her after practice” and “After practice I called her.” After normalization, every word can be matched. The difference in emphasis may matter even though the lexical measure returns zero. These two examples explain why a count and a close reading belong in the same study. Neither proves that an essay became stronger. Each can detect a kind of change that another measure overlooks.
What we looked for in each draft
Two drafts can use almost the same words and still leave a reader with a different impression. The questions in this table tell you what we looked for. A larger score means the change affected more of the essay. It does not mean the later draft was better.
The qualitative analysis asked how extensively the treatment of each feature changed. We used three ordered categories. A score of 0 means the earlier treatment was substantially retained. 1 indicates a local change in wording, emphasis, or arrangement. 2 indicates a change large enough to alter the role of a passage or what a reader could infer from it. These values describe scope, not quality. We used “unassessable” when the necessary evidence was missing. A single analytic review produced the codes. We did not estimate agreement between independent reviewers.
| Feature | Question asked while comparing the passages | Example of a consequential change |
|---|---|---|
| Specificity | Which observed actions, settings, choices and consequences can the reader identify? | A developed incident replaces a group of summarized examples. |
| Reflection | What meaning does the student give the experience, and what supports that meaning? | A later action shows how an earlier interpretation changed. |
| Structure | What function does each part serve, and how does the sequence affect understanding? | The middle of the essay begins doing a different job even though the opening image returns. |
| Repetition | Does a repeated element add meaning or perform a job already done? | Repeated explanations give way to a single demonstration, or an extra conclusion duplicates an earlier one. |
| Prompt alignment | Is the selected question known, and can both versions be assessed against the same requirements? | A previously unanswered part of the selected question receives a direct answer. |
| Length | How many words were gained or lost under a consistent counting rule? | Report the signed and absolute difference for the same student. |
Specificity does not mean accumulating proper nouns. A student can name a dozen places and leave the reader unsure what happened in any of them. Another student might name no place and still convey a precise decision. We counted a detail as meaningful when it changed what a reader could discern about the scene, the writer’s action or its consequence. A change that removed detail could receive the same scope code as an addition. The code indicates that the detail did different work.
Reflection concerns the writer’s interpretation of an experience. We looked for a relationship between an event and a belief, choice, question, or later action. A final sentence that merely repeats an existing lesson in smoother prose receives a different judgment from a paragraph that gives the writer a new reason for acting. A more forceful claim is not intrinsically a deeper one. Sometimes a writer revises a sweeping conclusion into a smaller claim they can defend more readily.
Structure concerns relationships among parts. We read a draft once in sequence, then asked what a reader knows after the opening, where the explanation changes direction, and what the closing contributes. Paragraph breaks offered clues. They could not settle the judgment. Moving a transition from the end of one paragraph to the beginning of the next may clarify the sequence without changing the architecture. Giving a previously minor event a full scene can change the center of the essay without changing its chronological order.
Repetition required particular caution. A recurring phrase can provide continuity while its meaning evolves. A repeated explanation can make the same assertion twice with no additional evidence. We asked what the second occurrence makes possible. If removing the occurrence leaves the reader with the same understanding, it may be redundant. If its meaning depends on what came between the two occurrences, it may be doing important structural work. This is an interpretive distinction for which reasonable readers can disagree.
Prompt alignment cannot be inferred from a response simply resembling a familiar personal statement. An essay about an intellectual question might plausibly fit more than one prompt. We did not have verified selected-prompt information for the seven pairs sufficient for a consistent comparison, so no alignment distribution is reported. A general assignment to write a personal essay establishes the genre but does not identify the question the student actually selected. For a future study, the student’s selected prompt should be captured at each observation, since students can also change the prompt during revision.
What changed in the seven essays?
The middle earlier draft was 670 words, and the middle later draft was 645. To learn what happened to a particular writer, we compared each student with themselves first. That produces the reported middle change of 16 fewer words. Subtracting 645 from 670 would answer a different question.
| Measure | Median | Denominator |
|---|---|---|
| Earlier essay length | 670 words | 7 essays |
| Later essay length | 645 words | 7 essays |
| Later minus earlier, calculated within each essay | −16 words | 7 pairs |
| Absolute change in length | 79 words | 7 pairs |
| Earlier paragraph count | 8 paragraphs | 7 essays |
| Later paragraph count | 7 paragraphs | 7 essays |
| Earlier word occurrences without a matching word | 263 words | 7 pairs |
| Later word occurrences without a matching word | 257 words | 7 pairs |
| Smaller unmatched count within each pair | 221 words | 7 pairs |
Source: The Ivy Institute, internal materials, data, and information.
Subtracting the two column medians would give a reduction of 25 words. That answers a question about the centers of two collections of texts. The paired estimate of a 16-word reduction answers the question about change for the same essay. When a relatively short first draft becomes longer and a relatively long first draft becomes shorter, the order of students can switch between the two columns. The student with the middle length before revision need not have the middle length afterward. For this research question the within-essay calculation is the relevant one.
The 79-word median absolute change captures how far drafts moved in either direction. It is compatible with the smaller signed change because shortening and lengthening move in opposite directions on the signed scale. The median later draft had seven paragraphs compared with eight earlier. Those two figures also summarize separate distributions. They do not imply that each writer combined a pair of paragraphs, or even that the student at the median in one column occupied the median in the other.
The lexical comparison gave a different picture from length alone. The median within-pair count of 221 unmatched word occurrences in both directions means that at the middle of the distribution, at least that many earlier occurrences lacked matches and at least that many later occurrences lacked matches. It does not mean a researcher identified 221 individual substitutions. A change in a single passage can generate many unmatched words, and an important reversal can reuse precisely the same words. Also, subtracting the two median unmatched counts does not recover the median signed length change. The underlying subtraction identity applies to each pair before summarization.
For a family wondering whether a revision deserves another reading, this distinction has a practical consequence. Two drafts can fit into the same space on screen and nonetheless tell the reader different things. Looking only at whether the text now meets a limit misses those changes. A family can make a more informed decision by checking separately whether the required length has been reached, what the later version now claims, and whether the student wants to stand behind that claim.
What close reading added
| Feature | Median scope on the 0–2 scale | Readable pairs |
|---|---|---|
| Specificity | 2 | 7 |
| Reflection | 2 | 7 |
| Structure | 2 | 7 |
| Repetition | 2 | 7 |
| Alignment to the selected prompt | Unassessable | Selected prompt insufficiently established |
Source: The Ivy Institute, internal materials, data, and information. A median of 2 means a consequential change at the middle of the distribution. Several features can change in the same passage. To protect participants, the public report gives median codes and withholds sparse category counts and all student-level combinations.
Reading the prose showed why the wording moved. Across the set there were revisions in which an example became an extended demonstration, revisions in which several scenes were condensed, and revisions that kept a familiar opening and sequence while adjusting the meaning of the explanation at the end. Taken together, the scope scores indicate that the middle pair for each feature involved more than local copyediting. The medians do not rank features by importance. Nor should four scores from the same essay be treated as four independent observations.
Specificity and reflection often need to be read together. When a writer adds an action, a reader has a new basis for interpreting a claim. When a writer removes an example, the remaining examples may bear more explanatory weight. A revision may introduce fewer distinct events and nonetheless convey more about the one event it develops. For that reason the study did not give points for every new detail. It asked whether the revised evidence changed the reader’s understanding.
The structure comparisons included both continuity and change. A statement can retain its organizing image throughout revision and still become substantially different if the student changes which events that image connects. Conversely, rearranging paragraph breaks can improve pacing while leaving the argument recognizable. Reading for paragraph function helped avoid attributing the same interpretation to those two cases. It also kept a longer draft from automatically receiving a higher structural score.
Repetition sometimes arose at the point of closure. A writer can finish an essay, reconsider how the last line should sound, and leave two possible conclusions together. A reader then faces a choice: treat both as the writing under review or infer which one the student intended to keep. We retained visible essay prose when there was no secure basis to remove it, then tested how the main results changed when we omitted the uncertain pair entirely. This sensitivity check matters more than arguing that one reading of an ambiguous ending must be correct.
Feedback was visible alongside some work. Its presence made it possible to recognize a question a reader had raised and compare that question with later writing. A later version sometimes addressed the concern. That sequence does not identify who supplied the new material, whether the student agreed with the reader, or what the student would have written otherwise. We did not attribute a percentage of words or any changes in quality to an adviser. The study design supplies no comparison group for that purpose.
Sensitivity to an uncertain ending
When the pair with the potentially overlapping ending was omitted, the analysis retained six students. The median within-pair length change was −13.5 words, and the median absolute change was 74.5 words. The median smaller unmatched count was 189.5 words. Half-word medians occur here because the average of the two middle observations is used for six pairs. No individual essay contained half a word.
Both analyses tell a similar story about scale. The typical signed difference was small compared with the movement of word occurrences in both directions. The recalculation does not solve the selection problem, verify submission, or independently test the reading rubric. It measures the consequences of one explicit editorial decision. A reader who would have excluded the ambiguous version from the outset can use the six-pair results for that reason. The Ivy Institute, internal materials, data, and information.
How much wording stayed in the same order?
Imagine keeping most of the words in a sentence but changing their order. A count of separate words would show little movement. Counting groups of three words makes some of that rewriting visible. Across seven pairs the middle match was 62.9% for individual words and 23.2% for three-word groups.
Read the detailed method and checks
For each pair we used the shorter version as the denominator. That lets us ask what fraction of the available word occurrences in the shorter text could be matched across the two. The numerator counts each occurrence only as often as it occurs in both texts. The denominator for the three-word calculation is the length of the shorter text minus two. We joined the body paragraphs in reading order. A group could therefore span a paragraph break. The punctuation and line breaks themselves did not determine whether a group could match.
Individual word occurrences: 62.9%
Consecutive groups of three words: 23.2%
| Unit matched between versions | Median share matched, all seven pairs | Median share matched, six-pair sensitivity analysis |
|---|---|---|
| Single word occurrences | 62.9% | 66.2% |
| Consecutive three-word groups | 23.2% | 24.0% |
Source: The Ivy Institute, internal materials, data, and information. Each percentage was calculated for one pair first. The table shows the medians of those percentages, rounded to one decimal place.
Many words remained available in both versions even when their neighbors changed. A personal statement can return to the same places, people and actions while the writer changes the sentences that connect them. Common words also match across quite different sentences. As a result, finding the word “the” in both texts adds little evidence that a particular phrase survived. Requiring three consecutive matches makes that coincidence less likely. It also makes the measure sensitive to small, routine edits.
Consider the invented line “By Friday the new shelf finally stood level.” Insert “wooden” before “shelf.” Nearly every individual word still matches. Several groups of three change around the insertion, although the reader understands essentially the same event. There is one act of editorial decision-making and several changed groups. This is why we report the matching measure as a description of the writing. Translating the percentage into “percent rewritten” or “percent original” would give a false sense of precision.
Now suppose the student keeps that whole sentence but relocates it from the introduction to the final line. Most word groups match. The structural effect could be considerable. At the beginning, the reader knows the repair will succeed and reads to discover how. At the end, the same sentence resolves a question. The numerical comparison understates this kind of change, while a reader examining the sequence can describe it readily. The two analyses complement each other particularly well when an essay is assembled around a recurring image.
The choice of denominator also matters. Suppose one writer keeps every word of a short draft, adds several scenes, and finally returns to the original ending. A comparison using the shorter version could approach complete overlap because nearly all the earlier writing survived. A comparison using the longer version would show a smaller share because the large additions have no earlier counterparts. Each figure could be correct. We used the shorter denominator to describe common material that fits within both versions and report earlier and later unmatched counts separately for readers interested in expansion and removal.
These analyses would be inappropriate as a test of who wrote an essay. Two drafts written entirely by the student can share relatively few word groups if the student retells an event from another point of view. Two drafts with extensive outside comments can share almost every word if the student decides the initial formulation was right. Knowing how much language changed gives a reviewer a reason to examine a passage. It cannot identify the writer or the process that produced it. The study drew no inference about the use of artificial intelligence.
Why a median can be modest when the work was substantial
Imagine sorting seven cards by how much their essays grew or shrank. A card for a large cut goes toward one end. A card for a large addition goes toward the other. The middle card can show a small cut even when the surrounding cards represent considerable change. Sort those same cards instead by the size of the change, folding additions and removals onto the same scale. A different card can move into the middle. This is what the signed and absolute summaries ask the reader to distinguish.
There is a further distinction between the size of a change and the fraction of an essay affected. A 50-word change to a 200-word answer occupies more of the available space than the same change to a long personal statement. Restricting this sample to personal statements reduces that particular variation but does not eliminate it. The unmatched-word percentages therefore add information to the raw counts. Readers should avoid using the 79-word median as a budget for a college supplement with a completely different length and purpose.
What would a student do differently with these results? A reader might be tempted to say that two versions differing by only a few words need no further review. Comparing the full essay with the changed passages gives a better basis for deciding. If every changed word concerns timing, responsibility or the scope of an accomplishment, a short difference can warrant careful factual review. If the changes mostly smooth transitions without changing the account, the student may decide that a quick read for continuity is enough. The study describes how misleading length alone can be. It cannot prescribe the right number of review rounds.
Interpreting the empty alignment column
The blank result for prompt alignment is a substantive limit on the analysis. A reader could often propose a plausible Common App question after reading an essay. Doing so and then grading how well the essay fits that question would introduce a circular judgment. The reviewer would have selected a question partly because the response already suited it. A high alignment score produced in this way would say little about whether the applicant answered the question actually chosen.
Nor would checking a list of current prompts fully solve the problem. A current list tells us what a student in a particular cycle could select. It cannot identify which question was selected for an earlier version. A student might begin under one prompt, choose a different prompt as the essay develops, and finally use a broad open option. In that situation the essay could become a better fit without much textual change. The missing variable is the writer’s selected assignment at the relevant point in time.
A useful design for future research would keep two alignment questions distinct. First, did the student address each part of the original prompt? Second, did the student end up choosing a prompt better suited to the essay? Combining those decisions into one score would obscure a major difference in strategy. One student might reshape the writing to answer a question. Another might recognize that the writing already answers a different available question. Both decisions can be intentional. This retrospective series cannot determine how often either occurred.
What a family can ask at three different readings
A first reading serves a different purpose from a last reading. The following procedure is a suggestion for families adapting the findings to their own work. It has not been tested as an intervention. It requires no particular number of drafts, paid review service or standardized timetable.
The first reading: what reached the reader?
Let one reader see the essay without the student’s spoken introduction. Ask the reader to write two sentences. The first should describe what the student actually did. The second should state what the reader believes the experience meant to the writer. Ask the reader to mark a passage supporting each sentence. This gives the student a small, inspectable account of what their writing communicated.
Suppose the reader says the student organized a project and learned confidence. The writer had hoped the reader would notice a decision to turn a project over to someone else. The difference is valuable. The student can check how much of the essay describes the launch and how much describes the transfer of responsibility. A small change in emphasis might solve the problem. Adding a new opening about leadership could move the account further from what the student wanted to say.
Another reader may understand the intended lesson perfectly but be unable to recall an action supporting it. In that case adding a single decision, observed interaction, or consequence might help more than adding another reflective sentence. A student can also decide that the reader inferred something they prefer to leave private. Clearer writing does not require increasing disclosure. They can change the subject or reduce the identifying detail while preserving the point they are comfortable making.
The middle reading: what changed in the account?
After a substantial revision, ask the same reader to identify what they now understand that they could not infer before. Ask a second question in the opposite direction: what useful piece of understanding disappeared? This second question is easy to miss. A revision that adds a vivid scene may also remove the one sentence explaining why the student was there. An essay that becomes more fluent may blur the distinction between a collective accomplishment and the student’s own contribution.
The student can arrange proposed changes under three decisions: keep, revise again, or restore. Restoring earlier prose is an ordinary part of writing. A student who tries a more dramatic opening may discover that the quiet first sentence better fits the rest of the piece. The later version then supplies evidence for selecting an earlier choice. A scoring system that rewards every change simply for being later would misdescribe that work.
At this point, compare the middle of the two essays as carefully as the endings. An ending can announce a transformation with considerable confidence while the middle shows almost no change in behavior. Alternatively, the middle can already contain a nuanced account and the final paragraph can flatten it into a general slogan. It may be enough to let a concrete decision conclude the essay. The test is whether the reader can follow the writer’s reasoning and whether the writer agrees that it represents them.
Some students find it useful to read the two versions aloud to themselves. The listener can identify a sentence that is hard to say or a phrase that sounds unlike the rest of the essay. That observation calls for a question, not an authorship judgment. A writer may use a more formal register for a particular scene on purpose. The student can decide whether the shift is effective or whether it arrived through a sequence of individually reasonable edits that, together, made the paragraph feel unfamiliar.
The last reading: is this the version the student approves?
An essay can reach an acceptable length and still contain an inaccurate attribution. It can answer every part of the question and still contain a detail about another person the student no longer wants to share. The final reading should leave scope for those decisions. One practical approach is to give the student a quiet pass through the complete response after everyone else has stopped suggesting wording. Ask them to identify any sentence they would hesitate to explain if asked about it later.
Next, compare the approved text with the text actually entered. A passage may appear twice because two versions were combined. A paragraph that looked separate in the writing environment may sit flush against the previous one. A choice made after shortening may have removed the reference that makes a pronoun clear. Looking at the destination version reveals these problems more efficiently than polishing a separate copy again. Georgia Tech’s applicant guidance specifically recommends rereading the transferred essay. Georgia Tech Undergraduate Admission
For a family keeping its own record, a useful endpoint is “the student approved what was entered for this application.” The student can privately retain the submission confirmation if they wish. A counselor’s last review, the student’s final decision and the college’s receipt can occur at different times. This distinction also explains why our study could say more about revision than about submission. The question “who saw which version?” requires evidence different from the question “how did the writing change?”
When a student changes direction entirely
The study concentrated on essays that remained recognizable across two versions. Consider the case this design would miss. After drafting a full statement about a team, a student realizes that the paragraph they most want to write concerns the job they took after the season. They start again. The new essay may grow out of the writing process even though the words, incident and structure barely overlap. Treating it as a revised version of the first statement would imply continuity that the writer might reject.
A family can preserve both versions and ask what the first one helped the student discover. Perhaps describing the season revealed which responsibilities continued after it ended. Perhaps a reader misunderstood the point, prompting the student to choose a different event. Or perhaps the second idea simply interested the student more. Each explanation requires the writer’s own account. There is no basis for choosing among them from two texts alone.
For research, the clean solution is to let a topic switch count as its own type of decision and follow the writer prospectively. The first question becomes whether the student continued, combined, replaced or paused an essay. Textual comparison can then be used where a common history exists. Counting every replacement as a 100% rewrite would falsely equate starting a different response with editing the same one. Excluding replacements from a narrow comparison is defensible, provided the report explains what its design leaves out.
Comparison with published guidance and research
The sources here describe writing, revision and what particular admissions offices tell applicants. They help explain what a family can look for when reading. Their authors did not review the seven essays in our comparison.
In her 1980 paper on revision strategies, Nancy Sommers questioned models of composition that put revision in a distinct stage after the writer has already settled what to say. That question is relevant here because a change in a student’s way of expressing an idea can also change the idea the reader takes away. Our two observations per student cannot reconstruct the intervening decisions or identify the writer’s thought process. Sommers studied writing and revision in a different setting. The connection concerns what counts as revision, and does not supply a comparison sample of admissions essays. Sommers, 1980
Yale’s admissions office encourages applicants to choose meaningful topics, retain control over their essays and ask people who know them for feedback. The distinction between topic and perspective matters for interpreting the Ivy material. An essay might retain its topic through many versions while the writer changes the perspective substantially. Conversely, a more distinctive-sounding topic might still communicate very little about the applicant. Yale’s guidance expresses one institution’s advice. It does not tell us how Yale would evaluate any essay in this study. Yale Undergraduate Admissions, Essays
Georgia Tech asks applicants to consider authenticity, what a reader learns about the writer, and whether the selected prompt has been answered. Its guidance also recommends checking the writing after it has been transferred into the application. These suggestions explain the value of two comparisons this study could not consistently make. Matching drafts to an identified prompt would support a direct alignment assessment. Comparing the agreed final version with the actual submitted text would establish whether the writing under review reached the college in that form. Georgia Tech, Personal Essays
The University of California advises applicants to develop concrete examples in personal insight responses and to use feedback while maintaining their own voice. The genre differs: a short, direct response has different demands from the extended personal statements reviewed here. The shared issue is how much evidence a reader needs to understand a claim. Percentages or length changes from this study should not be transferred to UC responses, and the guidance itself does not validate any Ivy statistic. University of California, Personal Insight Questions
A revision exercise for students and readers
The following examples were written for this report. They are fictional teaching examples. They do not paraphrase student writing or reconstruct any private experience.
Read for an action
An invented early sentence reads, “I learned how to lead when our rehearsal fell apart.” A possible revision reads, “When three players came in on different beats, I put down my instrument and asked each section to clap its entrance.” A reader now knows what the student noticed and what they decided to do. The new sentence remains incomplete as a piece of reflection. It says little about what the student believed beforehand, whether the intervention worked or why they still think about it.
A useful question is “which part happened differently because you were there?” The student may answer that somebody else actually proposed clapping the entrances. If so, the revision has uncovered a factual issue before it hardens into a leadership claim. An accurate alternative might focus on the student being the first person willing to admit they were lost. That change would shift the essay’s center toward listening and vulnerability. The writer, rather than the reader supplying feedback, has the information needed to make that choice.
Read for a consequence
Imagine a student writing about a community noticeboard. One conclusion says, “This taught me to value communication.” Another says, “The next Saturday, I stopped rewriting the notice and sat beside it long enough to see where people got confused.” The second sentence gives the reader a way to see what changed. It does not have to name a value. Whether it belongs depends on whether the action happened, whether the student wants to discuss it, and whether the preceding paragraphs establish what needed explaining.
A reviewer can place the two conclusions beside the same opening and ask which question each conclusion answers. The first may be adequate in an exceptionally short response where the evidence is already vivid. The second may open up a more interesting question than it resolves. What did people misunderstand? Did anyone ask for a change? The writer can decide whether to supply one more concrete observation, narrow the claim, or let the reader draw a modest conclusion. The exercise turns a general request for “more reflection” into a specific decision about evidence.
Read for the function of an image
Consider a student who begins an invented essay with a box of misprinted labels at a school repair table. The labels return in the last paragraph. In one version, the return merely reminds the reader that the labels were funny. In another, the student now understands that people stopped asking for repairs because they could not tell where to leave their things. The physical object stays the same. Its meaning changes because the writer has learned to see the process from a visitor’s perspective.
To assess repetition, cover the last reference and reread the paragraph. If the point is as clear and the ending feels just as complete, the reference might be expendable. If covering it removes the evidence for the change in perspective, the repetition is doing work. Counting occurrences of “label” would give an identical answer in the two fictional versions. Reading in sequence distinguishes their effects. A trusted reader can help the student decide whether the effect they intended is visible to someone seeing the essay for the first time.
Read for order
Suppose a writer describes waiting outside a locked practice room and later explains why they had arrived an hour early. As written, the reader may initially interpret the wait as a problem with a schedule. Moving one sentence about the writer’s purpose earlier can make the same wait read as preparation for a difficult conversation. There is no need to add a new achievement or a more dramatic incident. The revision changes what the reader understands while reading the scene.
List the actual jobs of the paragraphs using short phrases: introduces a question, shows the first attempt, explains an assumption, describes the consequence, reconsiders the question. If two paragraphs have the same job, the writer has a choice. Combining them may free space. Keeping them may be essential if the contrast between attempts reveals change. A reverse outline is an instrument for seeing those choices. It does not prescribe a particular number of paragraphs or a fixed shape for an essay.
Test the relationship between cut words and lost information
For an invented revision exercise, ask a student to remove a sentence that feels redundant, then ask a reader who has seen only the revised passage to explain what happened. If the reader can no longer tell who made the crucial decision, the sentence may have carried more information than its author realized. Restore the necessary piece in the student’s own words and look for a lower-cost cut elsewhere.
A family can use the same test before making a large cut. Ask what would be lost if an entire example disappeared. Perhaps it supplies the only evidence that the writer acted on an earlier lesson. Perhaps it demonstrates exactly what the previous example already established. The word savings could be identical in those cases while the cost to the essay differs. This is one reason why judging a revision by the number of words removed can lead a reader away from the student’s purpose.
Check alignment with the actual assignment
Consider two invented prompts. One asks how an experience changed the writer. Another asks what the writer contributed to a group. The same event could help answer both. In the first, the reader needs an earlier and later understanding. In the second, the reader needs an identifiable action and its effect on others. A beautiful scene that establishes neither distinction leaves the main question unresolved.
Have the writer identify the exact question selected for the destination application, underline each distinct request, and put a mark beside the passage that responds to it. If an answer depends on something the applicant has explained aloud but left out of the essay, the family reader has a useful observation to offer. If a broad personal-statement prompt genuinely allows the writer to choose any subject, avoid imposing a narrower prompt the student never selected. The family has an advantage the retrospective researcher lacked: it can ask the student directly which question they intend to answer.
Resolve conflicting reader advice
One reader might ask a student to start with dialogue. Another might ask for the moment after the conversation. Both are offering an opening, but they may have identified different problems. Ask what each reader found confusing or what each wanted to understand sooner. A reader who could not identify the setting has a different concern from a reader who could identify it but felt the second page contained the most interesting insight. The student can use both observations without adopting either proposed opening.
Keep the original version available for comparison. Read each candidate opening into the same next paragraph. Which one establishes an expectation the rest of the essay actually fulfills? Does either introduce a claim the student would have difficulty defending? Does either make a private fact about another person more identifiable? A student can decline an otherwise elegant suggestion on any of these grounds. A revision process that leaves this decision with the student is consistent with the emphasis Yale and Georgia Tech place on applicant voice. Yale Undergraduate Admissions · Georgia Tech Undergraduate Admission
Limitations and what a stronger study could measure
Seven students give us seven comparisons. We could identify changes in the writing, but we could not consistently identify the exact prompt, determine who suggested each change, or verify the submitted version. We did not have a second independent set of ratings.
The seven pairs are a small, selected sample of commercially advised students. The wider number screened is a count of candidate collections, and the targeted reading process was not a census of every essay within them. Reporting seven divided by 47 as the proportion of students who revise would answer a question the study never investigated. No denominator of all completed applications or all students served in either year was available for that purpose. A national estimate, subgroup comparison or admission-rate calculation would require different data.
The time between observations was not standardized. One pair could encompass a short polish and another several rounds of rethinking. We did not know the number of conversations, minutes of work, or changes that vanished before a later version became visible. A student might replace a paragraph and later restore it verbatim. The endpoint comparison would count no change there. Accordingly, these measurements describe differences between versions, not the effort or value of the writing process.
We did not verify a submitted copy for every applicant. An explicit final designation is stronger evidence than simply being the last version we found, but the student could still make a last-minute correction, select another essay, or submit different versions to different colleges. If the question concerns what admissions officers actually read, the appropriate endpoint would be the verified response for a specified application. A different unit of analysis would then be needed for a student who sends several variants.
Authorship was also outside the measurement. The materials show expressions on a page and, in some cases, suggestions made during review. They do not identify who composed or approved each line. A claim that a later draft sounds more adult would be an impression about style. It cannot establish who wrote it. No authorship detector, inferred personality trait or estimate of adviser contribution was used. No private text or identifiable case is necessary for the public to understand the results.
The qualitative coding was performed in one review. Independent readers might disagree, particularly on whether a repeated image had acquired a different function or whether a shorter reflection expressed a narrower belief. A follow-up could recruit two readers, give each the same rubric and separate copies, conceal the identities of students and advisers, and ask for a written reason for each decision. The readers could compare judgments after locking their original scores. Agreement and disagreement would both be worth reporting, with sparse student groups protected.
A prospective study could invite families to retain three approved observations. The first would be the earliest complete essay the student recognizes as a draft. The second could be a version designated ready for review. The third would be the exact essay sent to a specified institution, if the student authorizes that use. At each point, record the actual selected prompt and whether the student considers the essay a continuation of the same response. Offer a separate route for topic changes. Replacing an essay can be a meaningful writing decision even when a paragraph-level comparison no longer makes sense.
Such a design could also ask a short question before feedback: “What do you hope a reader will understand after this draft?” A reader could answer the corresponding question after reading. Changes in agreement across versions would capture an aspect of communication that word counts cannot. A participant should be able to decline, especially when the essay is personal. Any research use of submitted material would need appropriate consent and a retention plan. A richer design would answer a different question from whether a student was accepted, an outcome influenced by the full application and an institution’s decisions.
How to read and share the numbers
When sharing a figure, include what it counts. “Seven earlier and later essay pairs” is accurate. The same evidence does not support a claim about every first and final draft or how much editing colleges prefer.
Read the detailed method and checks
The unit in the numerical results is a pair of essays from the same student. A sentence such as “the median personal statement shortened by 16 words across seven ordered comparisons” keeps the unit and the comparison intact. Describing the same number as a reduction for the average college applicant would introduce both a different statistic and a different population. An even more consequential change would be to describe the result as the effect of professional editing. We did not observe the same students completing the same task in the absence of that support.
Reporting just the earlier and later lengths can create another misunderstanding. Readers may see the later median of 645 and infer that the process moved students toward a particular target. A count close to a familiar word limit can arise for several reasons. Some students may already have been writing to a known length. Others may have expanded a short version or chosen a different emphasis. The archive does not preserve a verified target and rationale for each revision, and the count used in the research can differ slightly from the count displayed to the applicant. We cannot tell how many students were trying to arrive at any particular number.
The medians for the four reading dimensions are ordinal summaries. A score of two identifies a category of change. It does not mean twice as much specificity or twice as much reflection as a score of one. Adding the four scores would create a numerical ranking without a defensible unit. For example, a major addition of one scene could change the evidence, interpretation and structure together. Treating those as independent achievements would reward one writing decision several times. We therefore do not provide a total score or name a most-improved essay.
Sparse counts affect what can safely be checked in public. Publishing an individual earlier length, later length and several distinctive changes could make a student’s essay recognizable to someone who knew the writer. Removing the name would provide limited protection. This is why the public materials provide aggregate medians, while a restricted calculation ledger retains the internal observations needed for reproduction. A reader can check the method with the worked sentences and interpret the reported sample. Rebuilding the student-level calculation requires authorized access to the underlying evidence.
It is reasonable to ask whether the sample size is large enough to support any practical use. Seven pairs can establish that a type of change occurred in the reviewed writing and can supply a transparent description of those particular comparisons. They provide little leverage for deciding how frequently a type of change occurs across schools or cohorts. They can motivate a more precise question for a family or a better designed future study. They cannot resolve a disagreement over how much an applicant elsewhere should revise.
One defensible use is in a discussion about what information to collect during an application season. Knowing only that a student “revised the essay” leaves several later questions unanswered. A small amount of additional information, such as whether the revision changed a factual point, responded to a prompt or addressed a reader’s misunderstanding, may be more valuable than preserving every intermediate sentence. A student should have a say in what is retained and why. Documentation that primarily serves future research should never be mistaken for information the family must supply in order to receive ordinary advising.
Another use is pedagogical. An adviser could give a workshop group two entirely fictional versions of a paragraph, ask participants to predict what a reader would learn from each, and then compare those predictions with the words that changed. A group can do this without assigning a quality score or asking students to share personal material. The exercise gives people practice distinguishing an observation about text from an inference about the person who wrote it. That distinction is essential to both a respectful writing conversation and a defensible research claim.
Conclusion
Across seven students, a modest change in total length often accompanied a substantial change in wording. A family can make the read-through more useful by asking which actions are clearer, where the student explains their meaning, and whether the ending still sounds like the writer.
In seven ordered essay comparisons, the median signed length change was −16 words, the median absolute change was 79 words, and the median smaller unmatched word count was 221. Close reading found changes in the role of examples, the interpretation of events, the shape of the narrative and the function of repetition. The evidence for strictly verified first-to-final pairs was too limited to support a separate public estimate.
A student choosing between drafts can ask a question the measurements leave open: “Which version gives a new reader the most accurate understanding of what I did and why it matters to me?” Word counts establish whether the answer fits the available space. Comparing actions, interpretations and paragraph functions helps the student decide what belongs in that space. The Ivy Institute, internal materials, data, and information.
References
The Ivy Institute, internal materials, data, and information.
Sommers, N. (1980). Revision strategies of student writers and experienced adult writers. College Composition and Communication, 31(4), 378–388. Article
Yale University, Undergraduate Admissions. Essays. Applicant guidance
Georgia Institute of Technology, Undergraduate Admission. Personal essays. First-year essay guidance
University of California, Undergraduate Admissions. Personal insight questions. First-year applicant guidance
