How every figure published on this site is checked, labelled and, when it fails, retired.
Copy the procedure. That is the point of publishing it. Every claim below links to the document it came from, because a page about following citation chains that does not link its own sources would not be worth much.
Why this page exists
Most market statistics you meet in a deck have been copied at least three times before they reach you. Each copy is a chance for a range to become a point, a cohort to shift, a percentage to transpose, or a source to be attached that never published the number.
We have three times caught this happening to figures we were about to publish, or had already published, ourselves. Once with two seed-graduation rates attributed to Carta that Carta never published. Once with a correction to an exit-value series where the correction was the error. Once with a comparison of our own that broke rule two, which we found only by auditing a live page.
So what follows is not a statement of intent. It is the procedure that catches our own mistakes, and the register at the bottom is the evidence that it does.
Repetition is not verification. The more often a figure is copied, the more chances it has had to drift.
Take the single statistic your current decision depends on most. Not all of them, one. Run it through the seven steps near the bottom of this page before you read anything else here. If it survives, you have lost ten minutes. If it does not, you have avoided pricing a decision off a number that does not exist.
The five rules
1. One metric, one methodology, including over time
A single time series comes from a single provider with a single set of inclusion rules, and those rules must not change inside the period you are charting.
Two providers first. We do not average KPMG and CB Insights. We do not take one provider’s funding numbers and another’s deal counts and present them as one dataset. When two credible providers disagree, we publish both and name the denominator each uses. The disagreement is usually not an error: private-market databases have different inclusion logic, and neither converts into the other.
But check whether they are actually two providers. KPMG’s Venture Pulse is built on PitchBook data, sourced on every chart as “KPMG Private Enterprise analysis of PitchBook data” and stated outright on its methodology page. The PitchBook-NVCA Venture Monitor is PitchBook. J.P. Morgan’s venture commentary is published inside the Venture Monitor. Three widely cited names, one database.
Now the harder half, which we added after an audit found we were breaking it. A provider can change its own methodology mid-series. PitchBook changed how it computes aggregate exit values from Q1 2020, added SPACs and reverse mergers to the IPO exit type from Q1 2021, and applied a new M&A extrapolation in January 2025. A 2020 to 2026 exit-value chart therefore crosses three breaks. Drawing it as one continuous series is not automatically wrong, but doing it silently is.
Find any chart in your own deck built from more than one data provider. Label each series with its provider or rebuild from one source. If two providers disagree by more than a rounding margin, show both rather than picking the one that helps. Never average two providers: the mean of two incompatible methodologies describes nothing.
Then read the provider’s methodology page for changes inside your date range. If there are any, either start the chart after the last break or mark the breaks on the chart. Search the report for the phrase “analysis of” to find out whose data you are actually looking at.
2. One vintage, never stitched
Private-market data is revised upwards for months as late-reported deals are confirmed. The revisions are large enough to invert a trend.
KPMG’s Q1 2026 press release reported 8,464 global deals. Its Q2 2026 materials reference 10,277 for the same quarter: a 21% revision to a quarter already published. Q2 2025 moved from 7,356 to 8,860 the same way. Q1 2026 exit value was revised from $413.5 billion to $434.7 billion.
So a historical series must come from one current export, not from stitching together the figure each quarterly press release carried on the day it was issued. And a first-print quarter must not be compared with a settled one.
Check whether any multi-year chart you rely on was assembled from successive press releases. If it was, it is measuring database revisions as much as the market. Rebuild it from a single current export, or add the vintage date to every row and say plainly that the series is not chart-ready.
For any quarter-on-quarter comparison, ask which quarter is first print. A difference of a few percent between a mature quarter and a first-print one is too small to establish direction in a series that has recently been revised by double digits. Do not trade the direction of a private-market count until you know both periods come from the same vintage. We do not publish a numeric noise threshold, because two or three observed revisions are not enough to derive one.
3. Read the source, not the article about the source
A figure is verified only when we have opened the document that produced it and read the value there. Not a summary, not a news write-up, not a screenshot of a chart. Where a value sits in a chart rather than in text, we say which chart and which page. PitchBook’s definition of IPO exit value, for instance, is on page 99 of the Q2 2026 report, and it computes that value from the company’s valuation at its IPO price rather than from cash raised.
For every figure that will influence a decision, follow the citation chain to its end. If the article cites another article, keep going. Stop only when you reach the organisation that produced the data.
If the number is behind a form or paywall and you cannot see it yourself, it is not verified. Label it and say so when you use it. That sentence costs nothing and protects you completely.
4. Leave the gap open
If a public source does not expose a number, we do not estimate it and we do not digitise a bar chart and call the output official. KPMG publishes two stacked charts of deal share by series without printing the percentages. We describe the structure and leave the field unresolved.
A table with a hole in it and a note explaining the hole is a better document than a full table with one invented cell.
Go through your deck for cells that were filled because the row looked wrong empty. Replace each with “not published by [source]” and a line on what would be needed to fill it.
The instinct to complete a table is the most common way invented numbers enter a serious document. It never looks like fabrication at the time. It looks like tidiness.
5. Publish the corrections
Every figure we retire goes into the numbered register at the bottom of this page, with the wrong value, the right value and the reason. Including our own. Two of the nine currently listed are ours.
Keep a running corrections list for your own published numbers and make it visible rather than quietly editing. Anyone who already acted on the old figure needs to know it moved.
This is also the cheapest credibility available. A visible register tells a reader you check after publishing, which is the only real evidence that anything was checked before.
The status system
Every row in every dataset carries one of these. Nothing is published without one.
| Status | Meaning | Can it go in a chart? |
|---|---|---|
| A — read in source | We opened the primary document and read this value in it | Yes |
| B — derived | Calculated from A-status values. Not printed in the source | Yes, labelled as calculated |
| C — secondary | Appears in a press release or secondary report, not in the primary document | Yes, with the source named |
| N — not published | The source does not publish this figure | No |
| X — corrected | A circulating value the primary source contradicts | Only as a correction |
A dataset is not verified because it has sources at the bottom. It is verified row by row, and the share of A-status rows is the honest measure of how much checking happened.
Add a status column to your next data table, then count the A rows. If the share is low, you now know something about your own document that you did not know before you counted.
When someone hands you a “fully sourced” dataset, ask what proportion of rows was read in the primary source. If there is no answer, it was collected, not verified.
What we do not do
- We do not fill gaps for completeness. An explained hole beats an invented cell.
- We do not assert an unverified negative. “No provider publishes this” is itself a claim, and we cannot check every vintage of every provider. We write “does not reproduce from the current vintage” instead, because that is what we actually tested.
- We do not treat a number as more solid because it appears in many places. Repetition is usually the opposite of verification, especially when the sources share a database.
- We do not describe exit value as money returned to investors. IPO exit value is computed from the company’s valuation at its IPO price, not from cash raised, and counts only the first majority liquidity event. The distinction changes the meaning entirely.
- We do not quote our own numbers without the date they were checked.
Adopt the fourth one today, because it is doing the most damage in this market right now. Whenever you meet “exit value”, establish whether it means a valuation mark or cash distributed. In 2026 those readings differ by orders of magnitude, and almost every headline uses the first while readers assume the second.
The check, step by step
Ten minutes per figure, and it works on anything.
- Find the producer. Not the article quoting it. The database, the central bank, the working paper. Check whether apparently separate sources share one.
- Open the original. If it is behind a form and you cannot see the number, the status is C or N, never A.
- Compare the exact wording. Ranges get rounded, cohorts merged, percentages transposed. Most errors enter here.
- Check the vintage, not just the date. A 2026 piece often quotes a 2024 dataset, and a 2026 dataset revises its own 2026 numbers.
- Check who benefits. An industry reporting its own returns is a different class of evidence than a census or a peer-reviewed paper.
- Check the definition behind the label. “Exit value”, “deal”, “mega-round” and “early stage” mean different things to different providers, and sometimes to the same provider in different years.
- When two credible sources disagree, publish both and name the denominators. Choosing the convenient one is how bad numbers spread.
Run steps 1 to 3 on every figure that reaches a decision document. Add 4 to 7 for anything quoted publicly or that a counterparty may check.
In our own work roughly one figure in three does not survive in the form it was circulating. Budget for that failure rate rather than being surprised by it, and build the schedule so a broken number can be dropped instead of patched at the last minute.
Corrections register
Every figure retired on this site, with the reason. Nine so far. Rows 3 and 9 are ours.
| # | Circulating figure | Verified | Status | Why it failed |
|---|---|---|---|---|
| 1 | 2020 exit value $522.4B | $622.4B | X | The Q3 2020 chart label reads $236.5B, not $136.5B. The widely repeated correction is itself the error |
| 2 | H1 2026 total $560.4B or $560.3B | Both official | A | The same KPMG release uses both. Rounding, not error |
| 3 | 2021 annual $750.8B or $750.9B | Neither is printed | B | Ours. Our sum of chart labels gives 750.8; KPMG’s press release says 750.9 from unrounded data. We had presented one without noting the other |
| 4 | Q1 2026 funding $330.9B | $332.9B | X | Revised between vintages |
| 5 | Q1 2026 deal count 8,464 | 10,277 latest vintage | X | Late reporting, a 21% revision. Vintages must not be mixed |
| 6 | Q2 2026 deal count 8,440 or 8,467 | 8,440 in the report | A | Both appear in KPMG’s own materials. The report also notes counts are partly estimated |
| 7 | +136% YoY, +128% HoH | +118.1%, +123.0% | X | Does not reproduce from the current vintage. It does reproduce from a mid-2025 vintage, so it is a vintage artefact rather than an invention |
| 8 | Annual totals presented as KPMG data | Derived sums | B | The report prints no annual totals |
| 9 | Exit count fell 4.4% in Q2 2026 | Unresolved | N | Ours, published and then withdrawn. It compared a revised quarter with a first-print one inside a series that revises about 20%. Found by auditing our own live page |
Update cadence
Datasets on this site carry the date they were last checked at the top, not in a footnote. Quarterly series are re-checked when the provider publishes a new vintage, and the whole series is replaced rather than extended, because extending is stitching.
If a figure has not been re-checked since its stated date, that date is the honest warning.
Put a “last checked” date on your own recurring numbers and treat anything older than one provider cycle as stale until re-read. For quarterly venture data that is roughly ninety days.
Corrections and credit
If a figure on any page here is wrong, out of date, or if you have a primary source for something marked unresolved, tell us. Corrections are credited on the page and added to the register above.
Applied
- Venture Capital’s $1.9 Trillion Exit Record Was Mostly a Valuation, Not Cash — why a $1.9 trillion exit quarter is mostly a valuation
- Venture Data 2026: which numbers check out and which do not
- Record Venture Funding 2026, Record Trouble Raising