Content: the first pillar of Answerability.

Before an AI engine can retrieve your page or trust your source, there has to be a page that answers the question — in a form the engine can lift. That is the Content pillar, and it fails more quietly than the other two.

§1Why Content comes first

The three pillars of Answerability fail independently, but they do not fail with equal visibility. Retrieval failures are loud — a crawler is blocked, a page renders only after JavaScript, a sitemap 404s — and they are findable with a checklist. Trust failures are slow but legible once you look — no Wikidata item, no external corroboration, no resolvable author. Content failures are the quietest of the three, because the symptom is not an error. The symptom is silence: the engine answers the buyer's question fluently, at length, citing several sources, and none of them is you. Nothing broke. There was simply no answer-shaped content of yours for the engine to reach for.

This is why we put Content first in the reading order. The three pillars fail independently — a page that exists can still fail Retrieval or Trust on its own terms, and fixing one does not predictably move the others — and a citation requires all three at once, not in sequence. But there has to be answer-bearing content for the other two pillars to have anything to evaluate: an engine cannot retrieve, or trust, a page that was never written. That makes Content the natural place to begin reading, and usually the natural place to begin work — not because it causally gates the others, but because it is the substrate they act on. A company that opens with the engineering checklist often spends weeks making an unanswerable site perfectly crawlable.

Content asks two things in sequence, and they are different problems with different fixes. First: coverage — does content exist that addresses the question at all? Second: shape — is that content written so an engine can extract it? A coverage gap is solved by writing something new. A shape gap is solved by rewriting something that already exists. Conflating them is the most common planning error we see: a company commissions a content sprint to "answer more questions" when its real problem is that the answers it already publishes are buried in narrative prose the engine cannot lift.

§2The buyer-question universe

Content is scored against questions, not keywords. The unit is the buyer-intent prompt — the actual sentence a buyer types into an engine — organized by who is asking and where they are in the decision. We build the universe from buyer archetypes — typically four to six, though the count tracks how differentiated the buying audience is: a niche B2B service may need four, a broad consumer brand many more. The archetypes together produce the standing audit set of roughly sixty prompts, spread across the stages of the journey: awareness, comparison, risk, pricing, fit, and post-purchase. Prompts-per-archetype flex inversely with the archetype count; the total is held near sixty for the reason the primer gives — enough surface to characterize a category without drowning the re-audit signal in noise. The archetypes are the load-bearing input. Generic prompts produce generic answers; archetype-specific prompts produce the answers that distinguish a sponsor from its competitive set.

The distinction matters because the engines answer the specific question, not the category. "What is cost segregation" and "best cost segregation firm for a $4M multifamily acquisition with a 2026 placed-in-service date" are different retrieval events with different winners. The first surfaces educational explainers; the second surfaces firms and the listicles that rank them. A company that has written the explainer and not the decision-stage answer is invisible exactly where the commercial intent concentrates.

Where do the prompts come from? From the sponsor's stated ideal customer profile, from sales-call language where transcripts are available, and from the vocabulary buyers actually use in adjacent communities — vertical forums, Reddit threads, trade press. No keyword tools are involved, because keyword tools report what people type into a search box to get a list, not what they ask a model to get an answer. The two phrasings diverge: search queries are terse and noun-heavy; engine prompts are full questions, frequently with situational detail the buyer would never type into Google. Content scored against search keywords will systematically miss the conversational, situated questions that the engines are actually answering.

A company can publish constantly and still answer none of the questions its buyers are asking the engines.

§3The content-coverage map

The coverage map is the central Content artifact. It is a grid: buyer questions down the rows, the company's existing content across the columns, and in each cell a judgment — does this content answer this question strongly, partially, or not at all. The map makes two things legible at once: which questions have no answer-bearing content, and which existing pages are doing the answering. It is the artifact clients act on first, because it converts an abstract anxiety ("are we visible in AI?") into a list of specific, writable gaps.

A simplified coverage map for a single archetype looks like this:

The pattern in this illustrative map is the one we see most often: strong awareness coverage, missing comparison and pricing coverage, partial risk and fit coverage. Companies write the content that introduces them and the content that closes a deal in a sales conversation, and they skip the middle — the comparison and pricing questions buyers now ask an engine before they ever make contact. The engine answers those questions from whatever third-party content exists, which is how a category's comparison conversation ends up owned by review sites and listicles rather than by any of the firms being compared.

§4The three content gaps

Content failures resolve into exactly three shapes. Naming them precisely matters, because each has a different remedy and a different cost.

Gap one — the answer does not exist. No page on the site addresses the buyer's question. This is a pure coverage gap, solved by writing. It is the cheapest to diagnose and frequently the cheapest to fix, because the question is known and the company usually has the expertise to answer it — it simply never published the answer in a form aimed at this question.

Gap two — the answer exists but is not answer-shaped. The content is on the site. A human who reads the whole page would find the answer. But it is written as prose to be read top to bottom, with the load-bearing claim in the third paragraph, no question-format heading routing the engine to it, and no self-contained passage the engine can lift without dragging in surrounding context. This is a shape gap, solved by rewriting. It is the most common Content failure and the least visible, because the company can point to the page and say "but we cover that" — and they do, for a human. The engine cannot extract it.

Gap three — a competitor owns the answer. The answer exists, it is reasonably well-shaped, but a competitor's version is more extractable, more corroborated, or simply got there first and became the canonical source the engines reach for. This gap is the most expensive, because closing it is not only a Content problem — it bleeds into Trust. You are not writing into empty space; you are trying to displace an incumbent the engine already treats as the answer.

The capture below shows the third gap operating in a live category. We asked five engines a decision-stage question; the four that returned web-grounded answers are shown. Read it for which kind of source each engine reaches for when it answers "best firm" — almost never a firm's own page, almost always a third-party ranking.

Two lessons sit in that capture. First, the question "best X" is answered by comparison-shaped content, and the firms that publish their own credible comparison content (KBKG) get cited for it — answer-shape is a content strategy, not just a copy-editing pass. Second, the source mix differs sharply by engine, which is why Content cannot be scored against a single engine: the same question routes to different content depending on who is answering. We return to that divergence in the engine notes.

§5What answer-shaped content looks like

Coverage gets you into the candidate set. Shape determines whether you are lifted from it. The properties below are the ones we score, and they are the same properties that, in the Princeton GEO experiments, moved visibility in generated answers — not because the model rewards the format directly, but because formatted-for-extraction content is what a model can actually quote.¹

• Self-contained passages of roughly 134–167 words. Long enough to carry a complete thought, short enough to be lifted whole. Content in this band is reliably more extractable than either 60-word fragments (which get concatenated with neighbors and lose attribution) or 300-word paragraphs (which get paraphrased or skipped).
• The answer in the first 40–60 words of a section. Openings appear to be weighted disproportionately in extraction; a passage that defers its answer to the third sentence is frequently passed over for one that leads with it.
• Definitions in canonical "X is…" form. "Cost segregation is an engineering-based tax study that…" is extractable; "We help clients with what's sometimes called…" is not. Engines surface definitional content heavily.
• Tables for comparison and steps. Engines extract tables reliably and re-render them as tables. A comparison written as prose is far less likely to be cited than the same comparison written as a table — which is part of why third-party comparison tables out-cite vendor prose, as the capture above shows.
• Question-format headings that mirror the prompt. "What does a cost segregation study cost?" as an H2 routes the engine to the answer beneath it. "Our approach to value" does not.
• Consistent terminology. One noun per concept, throughout. Synonym variety reads as stylistic polish to a human and as topic drift to a model deciding what the page is "about."

None of these are stylistic preferences. Each is a property that makes a passage liftable into a generated answer without truncation, padding, or loss of attribution. The goal of a Content rewrite is not more words; it is the same expertise, re-shaped into units a model can quote.

A second category, captured the same day, shows the same dynamic in software rather than services. The question is decision-stage; the winner is the product whose own content is answer-shaped for it.

§6Common failure modes

The library that answers the company's questions. The most common and least visible failure. A large content operation publishes steadily — but about the topics the company finds interesting, the launches it wants to promote, the thought-leadership angles its executives prefer. Mapped against the buyer-question universe, the coverage is lopsided: deep on the company's preoccupations, thin on the buyer's decision-stage questions. Volume is not coverage.

The buried answer. The page covers the question, but the answer is the payoff of a narrative rather than the opening of a section. Engines reaching for a quotable passage find none and move on. The fix is structural, not additive: surface the answer, add the question-format heading, break the prose into liftable passages.

The design-heavy page. A visually polished landing page where the substance lives in image captions, hover tooltips, scroll-triggered reveals, or a hero graphic — all invisible to extraction. The more designed the page, frequently the less extractable. What a human experiences as richness, the engine experiences as an empty document.

Terminology drift. The page calls the same thing three names across three sections, for variety. The model, trying to decide what the page is about, sees a page that is weakly about three things rather than strongly about one.

The unanswered question-heading. A heading phrased as a question — "How much does it cost?" — under which the body never actually answers, deferring instead to "contact us." The engine is routed to the heading and finds no answer to lift, which is worse than not raising the question, because it signals false coverage.

§7How Content interacts with the other pillars

Content is necessary and not sufficient, in both directions. A page can be perfectly answer-shaped and still uncited because the engine cannot reach it (a Retrieval failure) or because the engine does not trust the source enough to quote it over a corroborated competitor (a Trust failure). The capture in §4 shows the interaction directly: KBKG is cited by every engine not only because its comparison page is answer-shaped, but because KBKG is a sufficiently established entity that the engines treat its ranking as authoritative. A new entrant could publish an identically-shaped comparison page and be cited far less, because the Content is matched but the Trust is not yet built.

This is why Answerability is reported as a composite over three pillars rather than as one number. A company that scores ninety on Content and twenty on Trust has done real, visible work — and will still watch competitors get cited, because the bottleneck moved rather than closed. The value of scoring the pillars separately is that it tells you which work will actually move citations next, rather than flattering you with an average that hides the constraint. Content is where most engagements should start, because the other two pillars need answer-bearing content to act on — but a strong Content score does not, by itself, move Retrieval or Trust. Starting there is not the same as finishing there.

← Back to research

Note 2026-02 · Published 2026-05-24 · Independent research practice · [email protected]