The fastwrite Writing Report 2026: what 78,251 sources reveal about how students actually write
Almost everything we believe about how students write academic papers comes from surveys, anecdotes, or memory. People report what sounds reasonable, not what they did. We took a different route: we measured 78,251 actual literature additions across 14,311 writing projects.
Some widely held beliefs survive the data. Several don't.
The ten headline findings
| # | Finding | Number | Base |
|---|---|---|---|
| 1 | Literature work happens in the afternoon, not at night | 4.7% of additions between 00:00 and 05:59 | 78,251 additions |
| 2 | The most productive hour of the day | 3 pm (8.6% of all additions) | 78,251 additions |
| 3 | Night work is rare at the project level too | only 7.3% of projects have a single night-time source | 10,781 projects |
| 4 | The typical reading list is short | median of 4 sources | 10,777 projects with literature |
| 5 | Literature arrives in one burst | 68.3% of projects add everything on a single calendar day | 10,781 projects |
| 6 | Half of it takes under an hour | 53.5% of projects with ≥2 sources: span under 60 minutes | 8,668 projects |
| 7 | Students cite recent work almost exclusively | median publication year 2021; 54.4% of sources are five years old or newer | 77,848 sources |
| 8 | The classic text has all but disappeared | only 0.8% of sources were published before 1980 | 77,848 sources |
| 9 | Saturday is the quietest day of the week | 11.1% of activity, against 16.5% on Tuesday | 78,251 additions |
| 10 | Enabling a feature is not using it | 68.8% allow AI literature search; 38.2% actually use it | 10,511 projects |
Where the data comes from
fastwrite is a writing assistant for academic work, used primarily by students in German-speaking Europe. Users upload their own PDFs and can optionally allow fastwrite to add external scholarly papers. On 29 July 2026 we took an anonymised snapshot of the production database.
Observation window:
4 June 2025 to 29 July 2026
| Metric | Value |
|---|---|
| Writing projects | 14,311 |
| Users with at least one project | 12,409 |
| Literature sources included | 78,251 |
| — self-uploaded PDFs | 64,435 |
| — external papers found via fastwrite | 13,816 |
| Distinct titles | 63,182 |
| Total pages recorded | 2,553,222 |
92.7% of users created exactly one project. This dataset describes people writing a single piece of work, not prolific repeat users.
What we did not analyse:
no chat texts, no document contents, no names, no institutions, no grades. The analysis runs on metadata and on the bibliographic titles of the literature only.
Finding 1: The night-owl student is a myth
The image of a student hunting for sources at three in the morning is culturally universal. In the data it is a rounding error.
| Time of day | Share of literature additions |
|---|---|
| Night (00:00–05:59) | 4.7% |
| Morning (06:00–11:59) | 19.1% |
| Afternoon (12:00–17:59) | 47.7% |
| Evening (18:00–23:59) | 28.5% |
The single most active hour is 3 pm (8.6%), followed by 2 pm (8.5%) and 4 pm (8.1%). The least active is 5 am, at 0.23% — one thirty-seventh of the afternoon peak.
At the project level it is starker still. Of 10,781 projects containing literature, only 789 — 7.3% — have even one source added at night. For just 3.9% of projects does the majority of literature arrive between midnight and 6 am.
And the obvious objection — that deadline months must look different — doesn't hold. Across all 13 fully observed months, the night share stays between 3.5% and 5.8%, with no visible spike in the busiest months.
Finding 2: Saturday is the quietest day of the week
Weekends make up 28.6% of the calendar but only 24.7% of literature work. Academic writing, in this dataset, is a weekday activity.
| Weekday | Share |
|---|---|
| Monday | 16.4% |
| Tuesday | 16.5% |
| Wednesday | 15.8% |
| Thursday | 14.3% |
| Friday | 12.3% |
| Saturday | 11.1% |
| Sunday | 13.6% |
Activity falls steadily from Monday to a Saturday trough, then recovers on Sunday. Sunday is more productive than Friday.
Finding 3: The typical project has four sources
Across all 14,311 projects the median is 3 sources. Excluding the 24.7% of projects that contain no literature at all, it is 4.
| Cohort | n | Median | Mean | 90th pct |
|---|---|---|---|---|
| All projects | 14,311 | 3 | 5.5 | 13 |
| With ≥1 source | 10,777 | 4 | 7.3 | 17 |
| With ≥5 sources | 4,924 | 7 | 13.2 | 28 |
The distribution is heavily skewed. Among projects with literature, 45.7% have five or more sources, but only 18.3% reach ten and only 1.5% reach fifty. The largest project in the dataset has 232 sources.
An important caveat: a fastwrite project is not necessarily a complete thesis. It may be a chapter, a seminar paper, or an abandoned attempt. The number describes projects, not grades.
Finding 4: Literature arrives in a burst, not a stream
This is the finding that surprised us most.
68.3%
of all projects with literature add their entire reading list on a single calendar day. Even among projects with three or more sources, it is 56.1%.
Looking only at projects with at least two sources, the median span between first and last source is 0.02 days — roughly 29 minutes.
| Span between first and last source | Share of projects |
|---|---|
| under 1 hour | 53.5% |
| under 24 hours | 65.5% |
| under 7 days | 81.0% |
| under 30 days | 93.2% |
From the other direction: 65.2% of all 78,251 sources arrive on a day when the same project receives at least five. 38.7% arrive on a day with ten or more.
The honest caveat: this does not mean literature research takes half an hour. It means literature reaches the software in a single transfer. Many users evidently collect elsewhere first — in a browser, a library database, a downloads folder — and then upload everything at once. What we measure is the moment of handover, not the thinking before it.
That is still worth knowing. Reference tools are usually built for continuous curation. They are being used as a drop box for one bulk upload.
Finding 5: Students cite almost nothing old
A publication year is available for 99.6% of sources. The median year is 2021 — the typical source in use is about five years old.
| Age of source | Cumulative share |
|---|---|
| ≤ 1 year | 21.8% |
| ≤ 3 years | 39.5% |
| ≤ 5 years | 54.4% |
| ≤ 10 years | 78.4% |
| ≤ 20 years | 94.1% |
| older than 30 years | 2.3% |
Only 634 of 77,848 sources — 0.8% — were published before 1980. To put the shift in perspective: this dataset contains more sources from 2025 and 2026 combined (16,943) than from the entire period before 2010 (7,819).
Notably, self-uploaded PDFs and papers found automatically by fastwrite share the same median year of 2021. The recency bias does not come from the search tool; it matches what students select on their own.
Finding 6: In German-speaking Europe, scholarship is still German
This finding is specific to our market, but it is a useful data point for anyone studying non-English scholarly practice.
We determined the language of each bibliographic title using two independent methods — a statistical language model (fastText) and a transparent rule-based classifier. On the 39,902 titles where both methods reach a decision, they agree in 99.8% of cases.
| Method | Coverage | German | English |
|---|---|---|---|
| fastText only (strict) | 54.0% | 75.7% | 24.3% |
| Both methods combined | 76.2% | 72.3% | 27.4% |
Both approaches land in the corridor of 72–76% German-language literature, despite a 22-point difference in coverage.
At the project level the distribution is strongly bimodal. Of 5,946 projects with at least three classifiable titles, 48.8% use German-language sources exclusively and 11.5% use English exclusively. Only 39.7% mix the two.
There is also a counter-intuitive twist. The papers found automatically by fastwrite are more German (82.8%) than those students upload themselves (69.7%). A German-language research question returns German-language results. The common worry about AI over-representing English scholarship inverts here: the language of the query narrows the search space.
Finding 7: Allowing a feature is not the same as using it
fastwrite asks explicitly, at project setup, whether external scholarly literature may be included. That decision has been stored reliably since December 2025, so the following figures cover projects created from that point onward (n = 10,511; 99.8% with a stored decision).
- 68.8% allow external literature
- 31.2% deliberately restrict the project to their own uploads
But of the projects that allow it and contain any literature, only 38.2% actually added an external paper. Estimating adoption from settings alone would overstate it by a factor of 2.6.
Where the feature is used, it is used thoroughly: for 24.5% of those projects, external papers account for 100% of the literature. There is no average usage pattern here — there is non-use and there is heavy use.
What this data cannot show
We consider this the most important section of the report.
- Timestamps record when something was saved, not when work happened. A source stored at 3 pm may have been read at midnight. Every temporal claim here refers strictly to the moment of saving in fastwrite.
- An added source is not a read source. Nor a cited one. This data says nothing about reading or citation.
- A project is not necessarily a finished piece of work. It may be a chapter, an experiment, or a test.
- The dataset is not representative of students generally. It describes people who chose to use fastwrite — a self-selected group.
- Language is inferred from titles, not full texts. A German title can belong to an English text and vice versa. 23.8% of titles remain unclassified, mostly because they are filenames rather than titles.
- There is no data on grades, discipline, institution, or country. Claims about quality, success, or time saved cannot be derived from this dataset.
- Known inconsistencies were not silently corrected: 12 sources with no attributable project, 300 implausible publication years, 260 projects with a contradictory settings state.
Frequently asked questions
How many sources does an academic paper have on average?
In the fastwrite dataset the median is 4 sources per project containing literature (n = 10,777) and the mean is 7.3. The distribution is heavily right-skewed: only 18.3% of projects reach ten or more sources.
Do students work on their papers at night?
Mostly not. 4.7% of all literature additions occurred between 00:00 and 05:59, and only 7.3% of projects contain any night-time source at all.
How recent is the literature students use?
The median publication year is 2021. 54.4% of sources are five years old or newer, 78.4% ten years or newer. Just 0.8% were published before 1980.
How long does literature research take?
Measured as the span between the first and last source saved in fastwrite: 53.5% of projects with two or more sources fall under one hour, and 68.3% complete everything within a single calendar day. This measures the upload, not the searching that preceded it.
Is this dataset publicly available?
The aggregated figures in this report are. The raw dataset contains usage data and is not published.
How to cite this report
For individual figures, a reference to the report and the data date is enough. If you need additional breakdowns for an article, a paper, or your own research, get in touch — we're happy to share aggregated tables.
Methodological note: all times in Europe/Berlin, including daylight saving. Only fully processed self-uploads and external papers added via fastwrite were included; deleted, inactive, and failed entries are excluded. The 78,251 rows are literature additions, not 78,251 unique publications — the same publication can appear in multiple projects (63,182 distinct titles).