How much do we miss?

Every tracker claims to be comprehensive. This one publishes the test. We fixed a list of 169 real events that happened between 2026-07-01 and 2026-07-28, assembled from public sources without ever looking at our own database. Then we checked how many of them we actually hold. We held 36 of 169. Measured 2026-08-10.

The direction

Coverage of the independent gold set has gone from 9% to 21.3% across 4 weekly measurements between 2026-07-28 and 2026-08-10, a gain of 12.3 points. Higher is better: this is the share of independently listed events this tracker holds.

event captured captured with every detail correct

Where we fall short

These are our worst categories in this measurement, computed from the table below rather than chosen.

• Leadership changes outside the US: we hold 3 of 43 (7%).
• Funding rounds outside the US: we hold 20 of 92 (21.7%).
• US funding rounds: we hold 8 of 22 (36.4%).

Part of this is structural and will never fully close. In the United States a listed company changing an officer must file an 8-K, and a company selling securities must file a Form D. Those are mandatory, machine readable and free, which is why our best cell is US leadership changes. There is no equivalent anywhere in the world for a private funding round. A seed round in Nairobi, Jakarta or Bogota exists as a press release and whatever local outlet chose to write it up, in whatever language. Anyone telling you their private funding coverage outside the United States is complete has not measured it.

The rest of it is simply us, and we are not going to dress that up as a structural problem. Filing regimes do not explain a gap this wide. It is a young tracker reading a small number of collectors, at a volume far below what the world actually announces in a month. The events it does catch skew to the ones that arrive as machine readable filings. Even in the category with a mandatory filing we are well short of complete. This number is bad. We publish it anyway. A bad measured number is worth more than a good unmeasured claim. This is also the only version of this page that can improve for a reason you can check.

Every table below scores twice. The second score is stricter. An event we hold with one wrong detail (the amount, the date, the company or the place) passes the first score and fails the second.

Recall by category

The four cells that matter most. "Event captured" means we hold an entry for it. "Captured with every detail correct" also requires the country, the amount, the date and a working source link to be right.

By kind of update

By where the event happened

By what kind of document announced it

A mandatory filing is a different collection problem from a press release in a local outlet. This row is the honest measure of that difference.

By size of the event

Large means a funding round of $50M or more, or a change at a large listed employer. Small events are deliberately over represented in the test set, because measuring only the large ones would flatter the result.

By country

Most countries carry only a handful of events, so treat a single country cell as an indication and not a rate. Under each score is why it is what it is: the sources we read there, the publishers we probed and could not wire, and the queue.

Against the whole market, by country

A different question from the tables above, so it gets its own table and its own units. The gold set asks how much of a fixed list of landmark events we caught. This table asks what share we hold of everything a market announced, using external market-size counts. External counts use their own definitions and publication dates, so these shares are indicative and not a parity claim. We recorded both sides of each row together on 2026-08-05. Read the last column first: a ratio between two differently defined counts is not a coverage score, and the one row where the units differ says so.

United States, measured separately

A separate reference set, assembled the same way and measured the same way, for one country. Its cells are hiring markets rather than countries. We held 21 of 51 events announced between 2026-06-01 and 2026-07-31, measured 2026-08-11 against set 87aa6b62e0a542ef.

By hiring market

Each market carries between eight and sixteen events, so treat these as indications and not as rates. This table is a work list.

By what kind of document announced it

By size of the employer

What is wrong with this test set

Every sampling method has biases. These are ours, written down before the result was known, because a benchmark that hides its own caveats is a brochure.

• This set covers FUNDING ONLY. Leadership changes, pay actions and ways-of-working signals are not tested here and their US coverage is unmeasured. That is a real gap and not a rounding of scope: US leadership events at privately held employers could not be enumerated from original sources at all. Open web search returns executive-moves databases, which may never be cited; the only free chronological index that remains is SEC EDGAR, which is our own feed. Every leadership pass attempted came back either over 90% exchange-listed filings or empty.
• Discovery ran through public web search, so the set is biased toward employers whose round was written up in English by an indexed publisher. A round announced only on a company blog with no pickup is under-represented, and so is any round covered exclusively by a deal database.
• The single largest cause of a verified event being dropped was that no fetchable page stated the company's headquarters. Modern startup coverage has largely stopped writing 'San Francisco-based', and small rounds get the shortest write-ups, so this bias runs against small employers specifically. The set is 35% small-band and the true population is far more than that.
• Several publishers refuse an automated fetch. Business Wire, GlobeNewswire, sec.gov, CNBC, Reuters, Fierce Biotech, Endpoints News and a number of metro dailies were unreachable during assembly, and verified events behind them were dropped. Coverage of this set therefore reflects which publishers allow a fetcher as much as where events happened.
• Three of the four passes stopped on an exhausted search budget rather than on an exhausted window. San Francisco reached 13 of a 16 target and rest-of-US 14 of 16, and eight named metros in the rest-of-US cell were never enumerated at all. The set is a valid sample of what the method reaches; it is not the complete list of US rounds in the window.
• Metro cells carry 8 to 16 events each, so a metro cell is an indication and not a rate. At eight events the 95% interval is about fifty points wide. The metro breakdown is a work list, and only the overall figure and its interval should be read as a measurement.
• 51 events resolve the headline to a 95% interval about 26 points wide at worst. That is roughly the resolution of the sibling layoff tracker's 57-filing benchmark and it is not enough to separate two cells that differ by ten points.
• The window ends 2026-07-31 and the corpus is still backfilling into that period, so a result measured soon after sealing understates what the tracker will eventually hold for this window.

Every category, over time

The same breakdowns across every measurement, as raw counts. This is where an improvement shows up as an improvement. A new set of feeds in a region moves that region's row and leaves the others alone. A movement that appears everywhere at once is usually the test changing rather than the tracker.

By category, over time

By kind of document, over time

By country, over time

Countries appear here only in the measurements whose test set contained them, so a blank is "not tested" and never "found nothing".

What is wrong with the ones we do hold

23 of the 36 events we hold have at least one field wrong or missing. These need a better extractor rather than a new source, and we count them separately for that reason. A record with no country is invisible to every geographic filter on this site even though we have it.

How this was measured

We chose one bounded period first: 2026-07-01 to 2026-07-28. Within it we assembled 169 events that genuinely happened at named employers, from public sources only: company announcements, regulatory filings, national and trade press. We never built the reference set by querying our own database, because a list drawn from what we already hold measures nothing at all.

We then sealed the set. We fixed its contents on 2026-07-30, before any matching ran, and the file carries a digest (aacab2b2ed7a207c) that we record alongside every published figure. We added and removed nothing after seeing the result. Four of the events were known misses before the exercise began, and we included them deliberately, because a test set that drops its known failures is not a test.

We then looked up each event through the same public API anyone else can use. We matched on employer name, kind of event, and a date window that allows for a late write up. An event counts as held when we have a record of it. It counts as fully correct only when the country, the amount, the date and the source link are all right as well. Where a rule could go either way, we wrote it to favour counting an event as held, so the misses reported here are conservative.

How this stays honest as it repeats

The measurement runs on a schedule and commits its result, so the series above appears without anybody deciding to produce it. That solves one problem and creates another: re-running the same fixed list of events forever would stop measuring reach and start measuring memory. Once those particular events are collected the figure walks towards 100% and means nothing at all, which is the most flattering way a benchmark can fail.

So we retire a test set rather than reuse it indefinitely. We replace it when the window it covers has aged out, or earlier when the number against it stops moving, and the run itself says which. Every past set stays on disk with its digest so any figure ever published here can be re-derived from the exact list it was measured against.

Assembling a replacement is not automated, and we are not going to pretend otherwise. Building a set means open ended research across national press in many languages, and judgement about what counts as an event. A generator that took the easy path would drift towards finding exactly the things we already collect, and the number would climb for the worst possible reason. So the schedule does the measuring, the arithmetic and the alerting, and a person still assembles and seals each new set. That is the honest sliver, it is a few times a year, and naming it is cheaper than a benchmark nobody should believe.

What is wrong with this test itself

A benchmark that hides its own weaknesses is a brochure. These are the ones we know about.

• The window ends on the day of the first measurement, so events announced in the last two days had almost no time to be collected. That makes the result slightly pessimistic at the tail rather than flattering.
• Some countries carry only one or two events, so a single country cell is an indication and not a rate.
• The set counts announced events. A funding round disclosed only in a later filing, and a leadership change announced internally first, are both out of scope here.
• Three rest-of-world items are debt facilities rather than equity rounds, which some readers would not count as a funding round at all.
• This measures two of the four kinds of signal the tracker collects, funding rounds and leadership changes. Pay and ways-of-working signals are not tested here and their coverage is unmeasured.
• Eighty nine events is enough to separate the big categories and not enough to say anything reliable about a single country.
• The widening is a SUPERSET of 2026-07-v1, so 89 of the 169 events have been measured once already. Their result against v2 is partly a re-measurement, and only the 80 new events are being asked for the first time.
• Fifty of the 80 new events come from countries that were not in v1 at all, so most of the new geography carries one or two events and is an indication rather than a rate. The regional totals are what can be read; the country cells mostly cannot.
• Coverage of the new regions reflects which publishers allow an automated fetch as much as where events happened. Every research pass reported the same obstruction: national dailies in Nigeria, Ghana, Uganda, Rwanda and Greece, and the disclosure portals of the Indonesian, Malaysian, Sri Lankan, Kuwaiti, Athens and Johannesburg exchanges, refuse a fetcher outright. Real events were found there and dropped unverified, so those countries are under-represented in the set and their recall will look better than it is.
• Announcement date and effective date diverge sharply outside the US and western Europe, most visibly in African leadership changes, which are often announced weeks before they take effect. This set uses the announcement date throughout, which excluded several otherwise valid events.
• Ten funding events carry amount_disclosed=false because the publisher stated no figure. They are held to the same standard on every other field, and their amount cannot contribute a defect either way.

The measurement is a script in the public repository, and anyone can re run it against the live API. That is what makes this number contestable rather than a claim. Coverage changes, so a recall figure without a date is worthless: this one was measured on 2026-08-10 and will be measured again.

The test set, and every result

All 169 events, with the source that proves each one happened, so the reference set itself can be audited rather than taken on trust. Sorted worst first.

Measured 2026-08-10 against test set 2026-07-v2 (aacab2b2ed7a207c). The test set and the script that runs this are public. Found a real event from this window that we missed? Tell us and it goes into the next test set. Every source we read · Back to the tracker