What Document Analytics Can Never Tell You

Every article in this series ends with a short list of conclusions you must not draw. This one is that list, gathered up and argued properly — because the limits are not footnotes on the feature. They are the feature. A signal you know the boundaries of is something you can act on; a signal you do not is something that will eventually make you confidently wrong in front of a client.

Here is what no amount of tracking will ever give you, and what to do instead in each case.

It cannot tell you what they thought

This is the whole of it in one line, and everything below is a special case. Behaviour is observable; judgement is not. A person can read every word and remain unconvinced, or skim and be sold, and no timing, count or pattern separates those two.

The chain that gets people into trouble is short and each link looks reasonable: opened is a fact, time on screen is a reasonable proxy for reading, reading is a reasonable proxy for interest, interest is a reasonable proxy for intent. Four reasonable steps, and at the end you have a claim about somebody’s mind derived from a click. The three levels — opened, read, understood takes that chain apart link by link.

It cannot tell you which part they read

There is no scroll depth here and no section-by-section event. Nothing records where in a document a reader got to.

That is a deliberate absence rather than a missing feature. Scroll position measures where a window sat, not where a person looked — a reader who scrolls to the end and back while thinking about the opening section produces a “read to the end” that means nothing. Publishing a section heatmap built from that would give you a picture precise enough to plan around and wrong in ways you could not detect.

So: no conclusion about the pricing section, ever. If you need to know what somebody thought of the price, the mechanism for finding out is a question, and it always was.

It cannot tell you who a person is unless they told you

Identity comes in exactly three grades, and they are not interchangeable:

• Identified. They were signed in when they opened it. This is the only grade that is proof of who read the document.

• Self-reported. They typed an address into the email gate. Nobody checked it. It is evidence that somebody wanted the document enough to hand over an address — genuinely useful, and not evidence of identity. It carries a badge saying so wherever it appears, including in the alert email.

• Anonymous. Nothing is established. If you mailed the link to a named address it is shown as “someone with the link sent to …”, because a link forwards and that address is a fact about your mailing list.

Collapsing the middle grade into the first is the single most consequential error available in this category, because a name is what you act on. A typed-in address dressed as a confirmed one is how a proposal ends up addressed to somebody who never opened it.

It cannot tell you exactly how many people read it

A reader count is a floor, which is why it is printed with an “at least” sign. A person is counted by identity where one is known and by browser otherwise, so the same human on a phone and a laptop is two, and a reader who blocks cookies is new on every single visit.

The sign is there so that the number cannot be repeated as though it were exact. What a unique-reader count actually counts has the full account. The rule for you: never repeat one of these numbers to a client as a total.

It cannot tell you where they were

Country, mostly. City, unreliably. DB-IP, the source of these lookups, puts country accuracy at 95–99% and city at 55–80%, and the city error is systematic rather than random — an address commonly resolves to an ISP hub, so a whole region’s readers can appear in one capital.

That is why every city on the screen carries an approximation mark and a line stating the error next to it, rather than a footnote at the bottom of the card. A hedge that only survives if the whole card is read is not a hedge.

The failure this prevents is specific: reading “opened from Warsaw” when your client is in Kraków, and concluding the document was forwarded to somebody else.

It cannot tell you that a document was forwarded

Nothing labels a forward. What arrives is a visit that does not look like the visits you have already seen, and a forward is often the best explanation for that — alongside a second device, a cleared cookie jar, or the same person reading at home.

The honest sentence is “a reader appeared who does not resemble my original one”. Everything past that is inference, and it is worth holding at inference strength even when it is almost certainly right.

It cannot tell you why it was not opened

Zero opens is the least ambiguous number in the dataset and the most misread. It diagnoses the envelope — the subject line, the timing, the recipient — and says nothing whatever about the document, which nobody has seen.

Rewriting your pricing in response to an unopened proposal is the classic unforced error in this whole area.

What it can tell you, which is not nothing

Read back, the list looks bleak. It is not, and the reason is that the thing tracking is actually good for was never a conclusion about a person.

• That a document arrived and was opened, with a timestamp. A fact, and one you previously did not have.

• That somebody came back. Deliberate, and the strongest signal in the set.

• That the audience widened. Actionable, because it changes who you are writing for.

• That nothing has happened at all. Which is a diagnosis of your covering message, and a cheap thing to fix.

All four change your timing and your framing. None of them changes what you know about the person. That is the correct scope of the whole category, and a tool that promises more than it is promising you something it cannot deliver.

And one thing that is not a limit but a rule

Never quote any of it back. “I saw you opened it twice on Thursday” is true and it converts a useful signal into an uncomfortable conversation. What you bought is the right day to call and the right thing to lead with; nothing you say should reveal how you knew.

The other half of that bargain is that the reader is told. Every document shared through Quixli carries a visible notice saying the owner can see view statistics. Tracking nobody was told about is surveillance rather than analytics, and the notice is what keeps that distinction real rather than rhetorical.

The complete inventory — what is recorded, what is filtered, what is graded and what deliberately does not exist — is on the features page. If you are working through a specific number right now, start with what a short read actually means or what silence after a read means at each interval.