Crunchbase held 41 of the 45 real deals whose company it lists, every one of USD 100 million or more among them, logged a median of a day before the article I had read. Dealroom’s free pages held 30 of 32. I had expected at least a quarter to be missing.

So, is Crunchbase up to date on funding rounds? On this test, yes. Complete, and fast. I went in to show a gap and came out with the opposite, and that changed what I build.

The idea I wanted to be true

Every day ins7ghts reads what practitioners publish about data and AI, and among the things it picks out are deals: who raised, who bought whom. The tempting story writes itself. An AI that reads the press every morning must catch deals Crunchbase and Dealroom haven’t logged yet.

I have taken back enough numbers this year to not put that one in a post untested. So I wrote the test down first, with the bar on paper before any result: if at least a quarter of the real deals the press reported were missing from Crunchbase and Dealroom, the claim could go out. Below that, it couldn’t.

Before Crunchbase and Dealroom: half the deals weren’t deals

The first surprise came from our own side, and by we I mean me and my AI team. I drew 120 deal mentions from July to September 2026 articles, 90 funding rounds and 30 acquisitions, and had AI models check each against its article: is this a 2026 deal that the article reports?

Only 60 of the 90 funding mentions were. Of the 30 acquisitions, 7.

The rest were older deals named as background (a list of famous rounds in an investor’s blog, last year’s acquisition in a profile) or rumours. A machine reading “SpaceX raised” in a sentence about the past still reads “SpaceX raised”. So before comparing anything with Crunchbase or Dealroom, the honest number of real deals was 67, not 120.

Then Dealroom

Dealroom’s free company pages, for the companies whose page I could confirm by name: 30 of 32 real deals were there, and all 9 of USD 100 million or more.

Even the misses mostly weren’t. Where Dealroom seemed to miss a deal, our article had come months after it. K1X’s round, for one: on Dealroom in April, in our article in August. I was the late one.

Then Crunchbase

Crunchbase took a second try. The first run stopped after three companies on the rate limit, and I let it stop, because the plan said so. The run that counted went through Crunchbase’s own AI connector on my Pro account.

41 of 45 deals were there, all 18 of USD 100 million or more. And the timing: Crunchbase logged them a median of 1 day before our article.

The same week, Crunchbase put a new line on top of its home page: “Before It’s News, It’s In Crunchbase.” On deals, my test agrees with them.

One thing I’ll take from this for our own work. Several companies I couldn’t place at all, because the name pointed somewhere else. On Crunchbase, Clay matched a UX agency, and Enigma matched four companies of that name, none of them the one in the article. A bare name is a weak key, for them and for us.

What changed because of it

I said it changed what I build. Here is how.

Funding is context in ins7ghts now, never the lead. Who raised, who bought whom: that is Crunchbase’s and Dealroom’s ground, they cover it well, and I’m not going to pretend otherwise in a sales conversation. Like trying to beat the land registry at knowing who owns which house.

What a company does beyond the money is a different story: the launches, the partnerships, the people moving between companies. That is where I went next, and there Crunchbase’s picture looks quite different. That test has its own write-up.

And this test goes on the list of ones that failed. I keep that list on purpose, next to the ones that passed, so I know which sentences I am allowed to say.

The next test shows up first in Ins7ghts Daily, my daily newsletter on data and AI, before it becomes an article like this one.

Reading a pitch deck that leans on its funding history and its named customers? With the Deck Screen, every checkable claim in a pitch deck gets one verdict, Confirmed, Contradicted or Unverifiable from the public record, with the source attached.

The small print

Articles of July to September 2026; 120 sampled deal mentions, one per company; 67 real 2026 deals by their article. AI models did the reading and the first grading and checked each other’s work; every company the two sources did not find was checked by eye. Dealroom’s free company view, companies matched by name with an identity check by eye on every miss. Crunchbase through my Pro account on 31 October 2026, so a deal logged up to that day counts. “Held” means the deal is on the company’s profile; “logged” is Crunchbase’s own date for the entry. Crunchbase and Dealroom cited here as counts and a few names. More of these tests are on the 7wData research pages.