From votes to evidence

Lesson 44 of 50Module 9 · The future of link building5 min read

By Bill Hartzer · Hartzer Consulting · 25+ years in search, domains and expert witness work

Read this module differently from the other eight

Modules 1 through 8 teach link building as it works today. They are descriptive. Where I tell you that a followed link from a relevant page on a site with real traffic moves rankings, that is not a prediction — it is an observation, repeated across twenty-five years and a great many client programs, and you can go and verify it yourself this week.

This module is different. It makes an argument about where link building goes next, and an argument is not a procedure. Some of what follows is well established. Some of it is a reasonable inference from how the underlying systems are built. Some of it is speculation, and I will label it as speculation every time. Nobody outside a handful of companies knows how generative search systems weight their inputs, and the people who do know are not publishing it.

So weigh this module accordingly. If you read one sentence here and go and cancel your outreach program on the strength of it, you have misread me badly. The argument is additive: something new has started to matter, not something old has stopped.

The vote was useful because it was countable

The insight behind PageRank was not that a link is an endorsement. Librarians and citation analysts had understood citation as endorsement for decades before anyone indexed the web. The insight was that on the web, endorsement had become machine-readable. An href is unambiguous. It is directed — it points from one page to another and not the other way. It is discoverable by crawling, which means you can find every one of them without asking anybody. And it is countable, which means you can build a matrix out of it and run linear algebra on the result.

That combination is rare and it was enormously valuable. Text on a page tells you what a publisher says about themselves. The link graph tells you what everybody else says about them, in a form a machine can total up. For an engine trying to rank a web that had no other reliable quality signal, this was the closest thing available to a measurement of reputation.

Everything the industry built after that followed from those four properties. The metrics you learned in Module 2 — Trust Flow, Citation Flow, the various domain-strength scores — exist because the graph is countable. The audit habits in Module 3 exist because the graph is discoverable. The whole discipline is downstream of the fact that reputation had been encoded in a tag attribute.

Twenty years of optimizing the machine-readable part

Here is the problem with a countable signal: it invites counting. Once the industry understood that links were the currency, an enormous amount of effort went into producing the artifact of the vote rather than the endorsement the artifact was supposed to represent. Directories, article farms, reciprocal exchanges, blog networks, paid inserts — every tactic graded high risk in the reference section of this site is an attempt to manufacture the machine-readable part without earning the human part.

Google spent the same twenty years trying to tell the difference. Penguin was the loud version of that effort. Most of it was quieter: discounting rather than penalizing, ignoring rather than punishing. The practical result, which Module 2 covers in detail, is that a large share of links now do effectively nothing. My own profile is a fair illustration — 69.3% of the 22,260 domains linking to billhartzer.com sit at Trust Flow 0, on a profile where I have never bought a link. Links accumulate. Most of them are noise.

The lesson I draw from that history is not that links stopped working. It is that a signal loses value in proportion to how cheaply it can be produced, and the industry got very good at producing this one cheaply.

What a generative retrieval system needs instead

Now consider what a system that composes an answer actually needs from a source, as distinct from what a ranked list of ten blue links needs.

A ranking system needs to order documents. Reputation signals help enormously with ordering, and a countable graph is a good source of reputation signals. A generative system that retrieves passages and writes an answer over them has a different job. Before it can decide whether to use you, it needs to establish three things:

• That you are associated with the topic. Not that a page of yours contains a keyword — that your brand, product or person is one of the things routinely discussed in connection with this subject.
• That you said something worth quoting. A specific, attributable, self-contained claim beats a page that is merely relevant.
• That sources it retrieves from named you. Corroboration from documents the system actually pulls, in the words of those documents.

Notice that none of the three requires an href. A sentence in a trade publication reading “According to Hartzer, the median referring domain stops linking after about three years” supplies association, quotability and corroboration at once, whether or not my name is wrapped in a link. The link would be nice. It is not what makes the sentence useful to the system.

This is the whole argument compressed: the link was valuable because it was a machine-readable form of something else. When machines can read the something else directly, the wrapper matters less.

Evidence does not replace votes

I want to be exact about the claim, because the loose version of it is being sold hard right now and the loose version is wrong.

Links are not going away. Every one of the following is still true, and I expect all of them to stay true for a long time:

• Crawlers still find pages by following links. Discovery is a link problem before it is anything else. A document nothing points to is a document that is hard to retrieve, by any system.
• Classical ranking still runs on link signals. Traditional search results have not disappeared and continue to send the large majority of search traffic to most sites.
• Links still send people. A link in a piece somebody reads produces readers, and readers produce the branded searches and direct visits that no algorithm needs to be persuaded of.
• Link data is still the best available map of who is discussed alongside whom. Even for entity work, the graph is evidence.

The claim is narrower and, I think, harder to argue with: the link stops being the only currency. It becomes one form of evidence among several, and there are now situations — covered in Lesson 9.5 — where a piece of coverage with no followed link outperforms a followed link on a page nobody reads. If your reporting cannot represent that outcome, your reporting is measuring the wrapper.