They sell both ebooks and ereaders. Their ereaders and ereading software has the ability to synchronize data. They are in a far better position to judge how much of what is being read by looking at their data than we are without their data.

Are there issues with their data? Probably. Are these uncertainties quantifiable? In some cases, yes and in other cases no. They're going to have a hard time quantifying the uncertainties introduced by people who never connect to their services, but that won't matter to Kobo very much anyhow since such data is collected for their own marketing purposes. Figuring out how many people completed a book should be trivial, even ignoring the content, since they can examine the distribution for the last position read across all readers.

That is based upon data that we know Kobo collects. The reality is that they could collect much more data if they really wanted to. While there are likely limits to what they do collect, simply to protect the reputation of their company, they could tell how much attention a given reader gives to a given page. Consider readers collectively, on devices with displays and display settings, and they can probably isolate the paragraphs garnering people's attention or causing them to set down the book.

Is that proof that Kobo knows what they're talking about? Of course not. Yet it does point out that Kobo has the means of figuring out how many people finished a book and that they can probably do so with reasonable accuracy, at least for Kobo customers.