We have sixteen backlinks. Not one of them is a recommendation.

Build Notes — what shipped, what broke, and what we got wrong.

In the last edition I told you we had eight referring domains and that one of them was our own second property. That was wrong. It is two, and between them they account for ten of our sixteen links. I found out because I was fixing an unrelated typo.

Our largest single source of backlinks is buildvaillant.com, with five links. BuildVaillant is a web and product development studio. It is also our own venture — same company, same founder. I had been counting it as an external site linking to us. It is us linking to us.

Add the five from yieldaiglobal.com, our product site, and ten of our sixteen links come from properties we control. The genuine third-party picture is six domains and six links, not the eight and sixteen I reported.

I did not discover this by auditing anything. I was correcting a legal-name spelling — our registered entity is AgriVisionAI Inc., and two files still had it as “Agrivisionai Inc” from before the rename.

One of those files was llms.txt, which exists so that language models read our own description of the company instead of inferring one. Reviewing the change, I read a line I wrote myself: “BuildVaillant: a venture of AgriVisionAI Inc.”

I had written that. I had also, separately, been tracking buildvaillant.com every morning as though it were a stranger who had chosen to link to us.

Having got that wrong, I stopped trusting the rest of the number and actually examined it.

Two of the six are github.com and linkedin.com. Both are our own profiles, and both are nofollow by default, which means they explicitly pass no ranking signal.

One is wikidata.org. That is an entity record — it helps machines know we exist and who we are, which is genuinely valuable, but it is not a ranking signal and nobody chose to cite us.

The remaining three are bebee.com, tracxn.com and dealroom.co. These are auto-generated startup registry entries. No human decided to include us. A crawler found a company and made a row.

So: two properties we own. Two profiles we created. Three registry rows generated automatically. One entity record.

Sixteen links, and not one of them is an instance of a person reading our work and deciding to point at it. We have zero editorial links. I have been watching a daily number that was, almost entirely, measuring myself.

I want to be accurate in both directions, because the temptation after finding something like this is to overcorrect into drama.

This is the normal profile of a four-month-old, unfunded, founder-led company. Nobody has cited us because we have not yet given anyone a reason to. The entity records — Wikidata, Dealroom, Tracxn — are the correct foundation, and having them formed properly is genuinely worth something, because those are the records that downstream aggregators and AI answer engines read first.

We also have the two records I was about to tell you were missing. A Google Business Profile and a Crunchbase entry both exist. I had assumed they did not, which is its own small lesson about writing before checking.

Neither appears anywhere in our referring domains. For the Google Business Profile that is expected — it is a map listing, not a crawlable link. Crunchbase is less expected.

So I went looking for us on Crunchbase. I did not find us. What I found was AgriVision Solutions, AgriVision Equipment Group, AgVision, AgroVision, AGRVISION, Agriview and Agrovisio. There is also an agrivisionai.live, which is a different company entirely.

Our own README carries brand disambiguation as a standing copy rule, because several unrelated organisations share this name. I had been treating that as a writing problem — state the legal entity, state Detroit, state the founder. It is not a writing problem. It is a findability problem. Creating a record is not the same as being findable in it, and in a field of ten near-identical names, a record nobody can surface does close to nothing.

Correcting the tracker took one line — adding buildvaillant.com to the list of properties it should exclude. It now reports six third-party domains instead of eight, every morning, automatically.

But the configuration was never really the problem. I built a system to watch a number without first asking what the number contained. I had checked that the measurement was accurate. I had not checked that it was measuring the right thing.

The last edition was about not confusing inputs with outputs — about how easy it is, staring at a flat line, to report your activity instead of your results. This one is worse, and it is the more common failure. I did look at the output. I just never audited what was inside it. A number can be measured correctly, tracked diligently, reported honestly, and still be describing something other than what you think it describes.

The flat line I wrote about is still flat. It is just flatter than I said, and it starts lower.