How I Design a Source List

I design a source list as five retrieval tests, and I would not add publishing volume while the entity test and the citation-quality test are empty. Answer engines retrieve sources. They do not hand back a ranked list of our pages because the pages were relevant to a keyword. Research on these systems frames the job as an evidence-selection problem: does the source contain a claim the system can use, is the entity identified, and can a crisp answer be pulled from the URL? A no on those questions keeps the page out of the output, however much we have published.

The evidence-layer write-up of the same five tests is Source Architecture for AI Search Visibility: Evidence Layer. That page is the deeper resource. This essay is how I would use it as a founder: the order I would fund the layers, the two I would not skip, and the measurement that would tell me volume had become the right next spend.

The five tests, in the order I would brief them

Crawlability is the admission test. If the page is not in the indexes these engines draw from, it is absent. In 2026 that means more than a sitemap Google has seen. Engines ingest from different pipelines, including Bing, direct deals, crawled collections, and APIs. Perplexity's search API was built to depend less on Google and Bing. Semantic HTML, a clean canonical, and a fast response are the ticket into pools we cannot enter any other way. I would fix a blocked or client-rendered page before I wrote another one. I would not call that the strategy. It is the door.

Entity recognition is the test most publishing plans skip. Engines retrieve entities. Gemini maps a query onto its Knowledge Graph before it surfaces a source. A retrieval layer checks whether the brand or the person is a named entity, not only a string. First-person pages are not enough. Third-party pages that name the company or the person in a factual sentence are what fill the entity record. I would rather have ten independent sentences that use the same name than a hundred pages that say "I."

Citation quality is the compounding test. A study of 134 URLs across answer engines found that sources cited by more than one engine scored 71% higher on quality than sources cited by only one. The engines are converging on sources the ecosystem already treats as authoritative. An early citation raises the chance of the next retrieval. A source in Perplexity becomes a more plausible source for ChatGPT. A source in neither stays in neither. I would build a distributed record several pipelines can validate on their own, which is a different assignment from a single logo.

Answer extractability is the difference between a citation and a contribution. The paper From Citation Selection to Citation Absorption separates being cited from actually supplying the answer. Many sources are mentioned and unused. Absorbed sources put a direct answer near the top, use headings, define terms in the language of the query, and state claims that survive without the surrounding essay. Question blocks, numbered steps, explicit definitions, and comparison tables raise the chance of absorption. A long narrative can be cited and still contribute nothing. I would structure the pages we want used, and I would keep a column for "cited" separate from a column for "used."

Cross-engine distribution is the same architecture showing up in more than one pool. The citation-behavior research says sources that appear across engines are not a coincidence of taste. The features that make a source retrievable for Perplexity are largely the features that make it retrievable for ChatGPT or Gemini, because the pipelines are variations on retrieve-then-select. Building for one engine is a bet on one door. Building the five tests is an entry into several pools at once.

The two layers I would fund before more pages

Most source lists I am handed are layer one and a partial layer four. The site is crawlable. Some pages have headings. There is no distributed entity record, and there is no citation base across domains the engines have already qualified. Without those two, crawlability, formatting, and cross-engine hope do not compound. The machine has no authoritative entity to retrieve.

The fix I would fund is earned placement that names the entity, links to the owned source, and lives on a domain already in those pools. Each placement is an entry. Each consistent attribution is a vote for the record. More posts on our domain, while layers two and three are empty, add candidates the retrieval pass has no reason to trust.

I would track the percentage of category answers in which we appear, and I would call a rising number evidence that the architecture is working. I would not call a rising on-site word count evidence of anything in this system.

A high-quality placement on a trusted domain can show up in Perplexity in days. A distributed entity record across many sources takes months. I would tell the team both clocks so a quiet month on the entity work is not mistaken for failure, and a single fast citation is not mistaken for a finished record. Owned pages can improve extractability, layer four, without any press. They cannot build layer two or layer three. Those require sources that are not us. The model is the same for a person and for a company. The entity is the person's name or the brand. The strongest lists build both, so a company citation names the founder and a founder citation points back to the company.

The order of those layers is the order on Machine Relations. The name I would use for layer two, so a naming cleanup does not get filed under "SEO hygiene," is entity clarity.

I would keep a one-page score for the list so the five tests do not become a metaphor. Crawlability is pass or fail on the URLs we want retrieved, including the pipelines that are not Google's index. Entity recognition is a count of independent pages that use one name in a factual sentence. Citation quality is how many of those pages are the kind of source the 134-URL study found scoring 71% higher when more than one engine cites them. Extractability is whether the owned page puts a liftable claim near the top, which is the absorption half of the 2026 selection paper. Cross-engine distribution is whether the same source shows up in more than one engine's answer, rather than a hope that a Perplexity hit will spread by itself. A source that fails two of those does not get more volume. It gets the earliest failing test repaired. Adding posts while the entity count is flat is how a calendar imitates a retrieval strategy.

The decision

I would score every proposed source against the five tests and I would sort the quarter's spend toward the first failing test in the list, with a bias I will state: if crawlability is merely adequate, I spend on entity recognition and citation quality before I spend on more pages. I would require each new placement to name the entity in a factual sentence and to link to an owned page that can be extracted. I would review absorption on the tracked questions, not only the clip.

Scope: a source list whose job is retrieval by ChatGPT, Perplexity, Gemini, and the engines in the studies above. The 71% quality gap is from the 134-URL study. The selection-versus-absorption split is from the 2026 paper. It does not govern a list whose only job is to reach a human reporter, and it does not claim I have any engine's ranking code.

What would change my mind: a later measurement in which single-engine citations matched the quality of cross-engine citations, so the 71% gap closed, and in which owned pages with no third-party mention were absorbed as often as pages with an outside attribution. I would then add volume earlier. I would not add it because a content calendar had empty weeks. Empty weeks are a scheduling fact. They are not a retrieval result.

FAQ

How is a source list different from an SEO target list?

An SEO list chases keyword relevance and links in a web index. This list chases entity recognition, citation quality, and whether an answer can be lifted, across retrieval systems. Indexation still matters, and it is test one. The priority after the door is open is the entity and the outside citation, which a keyword list does not require.

How soon should I expect a citation?

A strong placement on a domain these engines already use can appear in Perplexity within days. The distributed entity record, the part that takes many independent sources, takes months. I would report those as different clocks. I would not average them into one "SEO takes six months" sentence.

Can the list be built without press?

Partly. Extractability can be improved on pages we own. Entity recognition and citation quality require someone else to name us in a way a machine can attribute. A source list with no third-party lines is a formatting project. I would keep it, and I would not tell the board it is a retrieval strategy.

Does the list change for a founder rather than the company?

The tests are the same. The entity string changes. For me, the string is my name. For the company, it is the brand. I would build both, and I would make each citation point at the other, so the graph does not contain a person with no company and a company with no person.