Cross-link discovery: finding the answer that already exists
When self-service fails, the most common reason is not that the answer was missing. It is that the reader could not find it. That is a link problem, and links are fixable.
Start with the failure, because it is more precise than it first looks. Gartner’s 2024 survey found that customers could not find content relevant to their issue in 43% of cases — the single most common reason self-service fails (Gartner, August 2024). The content existed. The path to it did not.
Nielsen Norman Group separates that failure into two useful halves. Findability is when “users can easily find content or functionality that they assume is present”; discoverability is when “users encounter new content or functionality that they were not aware of previously” (NN/g, Low Findability and Discoverability). Search, at best, serves the first. A cross-link serves the second — and across a split estate the second is the harder problem, because the reader cannot search for a resource they do not know exists.
Across a fragmented estate this gets worse, for a structural reason: the reader is not searching one site. They are searching whichever silo they happened to land in. A docs search will not surface the forum thread. A help-centre search will not surface the blog post. The content is there; the path is not.
Why “related” is not good enough
The obvious approach is embedding similarity: encode every page, find the near neighbours, link them. It produces plausible-looking output and it is not very useful, for two reasons.
The first is that similarity is symmetric and reading is not. A tutorial and its API reference are similar to each other, but the useful link runs one way at one moment and the other way at another. The second is that “related content” tells the reader nothing about why to click. A link that says this thread troubleshoots the error on this page earns a click. A link that says related earns a shrug.
So Kenaptic types its edges from a deliberately small vocabulary — explains, troubleshoots,
is-design-reference-for, discusses-real-world-issue-with, is-prerequisite-for. A small
controlled set beats an open one here: it keeps the types comparable across an estate, and it
forces the model to commit to a claim we can show the reviewer and defend.
Three signals, blended — and a number that is not a probability
A candidate is scored on topic overlap between the two pages, embedding similarity between the relevant passages, and a directional judgement about whether the proposed relationship actually holds. Topic overlap acts as a soft pre-filter and a booster rather than a hard gate, because a hard gate on topics throws away exactly the cross-silo pairs the product exists to find — a forum thread and a reference page often share very little surface vocabulary while being precisely about the same failure.
The score we show a reviewer is decision support, not a calibrated probability. We say so in the product and we will keep saying it. A number that looks like a confidence invites people to threshold on it and stop reading the evidence, and the evidence is the point.
Evidence, then a human
Every proposal arrives with the passages from both pages that justify it, the relationship type, and the exact change that would be written. The reviewer approves, edits the type, or rejects. Nothing publishes without that step.
This is also how the product gets measured. We keep two numbers apart on purpose: recall, whether the engine rediscovers links we already know should exist, and precision, the share of proposals a human actually approves. Collapsing them into one score would let a system look excellent by proposing very little, or by proposing everything.
Where the link goes matters as much as whether it is right
A correct link in the wrong place is close to worthless. A reference block at the foot of a long troubleshooting page is read by almost nobody who needed it. So an approved link can land as a foot-of-page reference, as an additive sentence in the prose at the relevant point, or both — and the sentence is written to sit in the flow of the paragraph rather than announce itself as machine output.
It is additive by design: Kenaptic adds a sentence, it does not rewrite yours. The words on your page stay your words.
Kenaptic proposes; a person decides. Every link is reviewed before it publishes and retractable in one click — see Security.
