How AI search selects and cites sources

Lesson 47 of 50Module 9 · The future of link building5 min read

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

What I am describing and what I am not

I am going to describe the general architecture of retrieval-augmented generation — the family of systems that answer a question by fetching documents and writing over them. That architecture is publicly documented in the research literature and in the developer tooling that thousands of engineers build with every day. It is not a secret.

What I am not going to do is tell you how any specific product works internally. I do not know how Google's AI features rank candidate passages. I do not know what corpus any given assistant retrieves from at any given moment, how often it changes, or how the citation shown to you is chosen from the material the model saw. Neither does anyone selling you a course about it. When you read a confident claim about the internals of a named product, ask where the person got it.

Everything below is therefore either general mechanics — how systems of this kind are built — or inference from those mechanics, marked as such.

The general shape of a retrieval-augmented answer

Strip away the product differences and systems of this type do roughly the following:

1. Interpret the question. The user's phrasing is turned into something searchable. In many implementations one question becomes several sub-queries, because a compound question needs several kinds of evidence.
2. Retrieve candidates. The system queries an index — sometimes a conventional search backend, sometimes a vector store, often both — and gets back a candidate set of documents or passages.
3. Rank and select passages. The candidates are scored and cut down to what will fit in the context the model works with. This is the step that decides whether you are in the running at all, and it operates on passages, not whole sites.
4. Generate, grounded in the selected text. The model writes an answer conditioned on the retrieved passages rather than on memory alone. This is what “grounding” means and it is the whole point of the architecture.
5. Attach citations. Sources are attributed to statements in the output.

Two consequences follow directly, and neither is speculative.

The unit of competition is the passage. A page that answers a question in one clear, self-contained paragraph is more retrievable than a page that answers it across four sections and a table. That is not a stylistic preference; it is a property of how chunked retrieval works.

You have to be in the corpus. If the system retrieves from a web index, then classical crawlability, indexation and — yes — the link-driven discovery taught in Module 1 are prerequisites. Retrieval cannot select what it never fetched.

Two different ways a model can know about you

This distinction is worth being precise about, because conflating the two produces most of the bad advice in this field.

Parametric knowledge is what the model absorbed during training. It is fixed at training time, has a cutoff, cannot be updated by you, and cannot be audited. If a model asserts something about your company with no citation, it is drawing on this.

Retrieved knowledge is what the system fetched at the moment of the query. It is current, it is the thing citations point at, and it is the part you can plausibly influence — by being in the index, being retrievable at passage level, and having said something worth selecting.

Practically, aim at the retrieved half. It is faster-moving, it is where citations come from, and it is the only part where your work this quarter can show up. Whether being named frequently in public text makes a future model more likely to mention you unprompted is a reasonable inference — that is, after all, roughly what training on text does — but I cannot measure it, you cannot measure it, and I am labeling it speculation rather than dressing it up.

What tends to get selected

From the mechanics, some properties follow that are worth writing toward. I would put these as strong inferences rather than proven rules.

• Self-contained statements. A passage that makes sense lifted out of its page is easier to select and easier to quote than one depending on three paragraphs of setup.
• Specific, attributable claims. A number with a source and a method beats a hedge. This is the practical argument for original research: a system composing an answer needs something concrete to attribute, and it will attribute it to whoever produced it.
• Clear attribution of who is speaking. Named authors with stated expertise give a system a person to credit. Anonymous content gives it a URL.
• Direct answers near the question. Headings that state a question and paragraphs that answer it immediately are structurally easier to match than the same information buried mid-article.
• Corroboration elsewhere. A claim repeated by independent sources is safer for a system to use than one appearing in exactly one place.

Notice that this list is largely a description of good editorial practice. That is not a coincidence, and it is the main reason I am relaxed about the direction of travel. The work that makes you retrievable is mostly the work that makes you worth reading.

A citation is not proof of influence

One trap worth naming, because a lot of nascent measurement rests on it.

The sources shown under a generated answer are not necessarily the full set of material the system considered, and being cited is not the same as having shaped the answer. Attribution in these systems is a separate step from generation. A source may be cited for one clause while the substance came from elsewhere. Conversely, a document may have contributed to the retrieval that shaped an answer and never appear in the citation list.

So when you count citations — which Lesson 9.6 covers — count them as a visible proxy for something you cannot see directly, not as a complete accounting of influence. It is a legitimate metric. It is not the ground truth it looks like.

What this does not change

A short and deliberately unexciting list, because the excitable version of this lesson is everywhere and it is wrong.

Crawl and index still gate everything. A page that cannot be fetched cannot be retrieved. Site speed, rendering, canonicalization and internal linking remain prerequisites rather than legacy concerns. Authority still helps — a system choosing between two passages saying the same thing has to break the tie somehow, and source quality is the obvious tiebreaker, which is the same problem classical ranking solved with link signals.

And the material you produce for this purpose is the same material Module 6 tells you to produce for digital PR: original data, clear claims, a named expert willing to be quoted. There is no separate content strategy hiding behind this lesson. There is the usual one, executed with more attention to whether individual passages stand alone.