The future of link building

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

What this module is

The first eight modules of this course are descriptive. They tell you how link building works right now, based on what I have watched happen across twenty-five years of client programs and on a dataset of 1,301,839 links from 22,260 referring domains that I can show you. Where the industry disagrees, I say so. Where a tactic is a waste of money, I say that too.

This module is different, and you should read it differently. It makes an argument about where link building is heading, and an argument is not a procedure. Some of what follows is settled — entity resolution and structured data, in Lesson 9.3, are how these systems work today. Some of it is a reasonable inference from the published architecture of retrieval-augmented systems. Some of it is speculation, and I have labeled it as speculation everywhere it appears.

Nobody outside a small number of companies knows how generative search systems weight their inputs. The people who do know are not publishing it. Any writing on this subject that sounds certain is either quoting something I have not seen or making it up, and I would rather be usefully honest than confidently wrong.

The argument, in short

A link is a machine-readable vote. That is what made it valuable: an href is unambiguous, directed, discoverable by crawling and countable, which is what allowed reputation on the web to be computed at all. The industry then spent twenty years optimizing the machine-readable part — manufacturing the artifact of the vote rather than earning the endorsement behind it.

A generative retrieval system does not need the href. To use you, it needs to establish that your brand is associated with a topic, that you said something specific enough to be worth quoting, and that sources it retrieves from named you. An unlinked mention in a publication a system actually retrieves can do work that a followed link on an unread page cannot.

Which means the reflex of checking whether coverage is dofollow before deciding whether it counted is increasingly the wrong question. Not because links stopped mattering — they have not — but because the link stops being the only currency, and most reporting still cannot represent the second one.

What is settled, what is inferred, what is unknown

Because this module is contested territory, it is worth stating up front which parts I would defend to what degree.

• Settled and actionable today. Entity resolution, consistent naming, sameAs relationships and structured data (Lesson 9.3). Nofollow has been treated as a hint rather than a directive for ranking purposes for some years (Lesson 9.5). Real, earned link profiles are mixed — 16.4% of links on this site's evidence-base profile carry a nofollow attribute.
• Well documented in general, opaque in the specific. The architecture of retrieval-augmented answering — query interpretation, retrieval, passage ranking, grounding, citation (Lesson 9.4). The general shape is public. No product's weighting is.
• Inference. That repeated, independent co-occurrence of your brand with a topic makes you more likely to be retrieved and named (Lesson 9.2). Plausible from the mechanics; I have no controlled study and will not invent one.
• Genuinely unknown. Whether unlinked mentions are a direct ranking input, and at what weight. How any specific system chooses which sources to display. How much of a model's behavior toward your brand comes from training rather than retrieval (Lesson 9.6).

Why it sits at the end

This module is deliberately last, and it is deliberately separate.

Threading a contested argument through eight modules of practical instruction would have muddied the instruction. You cannot properly evaluate a claim about where link building is going until you understand how it currently works — what a link actually does, why most of them do nothing, how the graph is read, how earned media is produced and measured. Putting the argument at the end lets it be clearly labeled as an argument, made to a reader equipped to push back on it.

The seven lessons build in sequence: the conceptual shift, why mentions accumulate into something links do not, the concrete entity work, how retrieval and citation actually operate, the honest position on link attributes, what can and cannot be measured, and finally what to change on Monday. If you read only one, read Lesson 9.3 — it is the most actionable and the least speculative thing here.

The lessons

1. 9.1From votes to evidenceWhy the link stops being the only currency — argued, not asserted.
2. 9.2Why unlinked mentions now carry weightCo-occurrence and association — what repeated naming builds that a link does not.
3. 9.3Entities and the knowledge graphStrings versus entities: naming, sameAs, Wikidata and structured data, done properly.
4. 9.4How AI search selects and cites sourcesRetrieval, grounding and citation — general mechanics and the limits of knowledge.
5. 9.5Why dofollow matters less than it didNofollow as a hint, and why the link attribute is the wrong first question.
6. 9.6Measuring brand presence in AI answersPrompt panels, share of voice and mention tracking — with honest limits stated.
7. 9.7What to do differently starting nowThe practical close — what to change, what to keep, and how to weigh the argument.