The One Page-Level Signal That Survives When You Control for Domain Authority

Most of the advice on improving AI citation rates points to the same levers: domain authority, backlinks, schema markup, directory presence. Get these right, the argument goes, and the citations follow.

That is partly correct. Domain authority does predict AI citation rates. Higher-authority domains get cited more often across every platform we audit. But it is an incomplete explanation -- and treating it as the primary lever leads businesses to the wrong fix.

A 2026 controlled study from Discovered Labs tested what predicts AI citations specifically at the page level, after accounting for domain authority. Their approach: domain fixed-effects controls, which hold domain authority constant across all pages in the analysis and isolate the contribution of individual page-level factors. The finding was striking. When domain effects are removed from the equation, almost every conventional page-level signal -- schema presence, content length, backlink count, heading structure -- loses its predictive power. One factor survives: vocabulary alignment.

The statistical finding, documented in our June 2026 content-format-citation-signals investigation (Scout session 51, 2026-06-18): β=+0.37, q≈10⁻⁷³. That q-value is not a typo. It reflects extreme statistical confidence -- not that vocabulary alignment has a huge effect, but that the relationship between vocabulary alignment and AI citation rate is highly unlikely to be noise.

What "vocabulary alignment" means, and where businesses miss it

Vocabulary alignment means that the language on your page matches the language buyers actually use when they type queries into AI tools.

That gap is almost universally larger than businesses expect. A residential plumbing company writes: "licensed residential emergency plumbing services available 24/7 throughout the Greater Toronto Area." A homeowner types: "burst pipe plumber near me" or "water everywhere call plumber" or "emergency plumber coming tonight." These are different vocabularies. The business wrote in its own professional register. The buyer typed in an incident register -- specific, urgent, situation-first.

AI platforms are retrieving pages to answer the buyer's query. A page written in professional category language is not the same source as a page written in buyer incident language, even when both describe the same service.

This is not a Perplexity-only issue. But Perplexity makes the mechanism most visible. In our investigation of cross-platform retrieval architecture (`knowledge/platform-divergence-data.md`, Scout session 42, 2026-06-09), we documented that Perplexity retains approximately 88% of the original query words when it runs its retrieval search. When a buyer types "burst pipe emergency plumber near me," Perplexity searches for sources that contain language close to that phrasing. A page written entirely in professional language -- "residential plumbing services," "licensed contractor," "comprehensive HVAC maintenance" -- does not clear the vocabulary threshold for those queries. It is passed over, not because the content is wrong, but because the words do not match.

ChatGPT is more forgiving on vocabulary because it generates 91% unique sub-queries from the original prompt before searching -- it decomposes "burst pipe emergency plumber near me" into many angles before retrieving. A page doesn't need to match the buyer's exact phrasing to be retrieved by ChatGPT. But that architectural difference does not make vocabulary irrelevant on ChatGPT -- it means vocabulary precision solves a different part of the problem there, while distribution breadth (directory presence, third-party listings) matters more.

How Claude reads your page

Our platform-citation-behaviors investigation (Scout session 14, 2026-05-12) documented a consistent pattern across eight Sourcepull audits: Claude's extractor operates at the sentence level in a way other platforms do not. A page with strong overall quality but poorly structured individual paragraphs may be read by Claude without being cited.

The specific pattern: Claude wants the key answer surfaced in sentence one of each paragraph or section. If sentence one is context-setting or acknowledgment ("We understand that emergency plumbing situations can be stressful..."), the answer is buried. RAG retrieval systems frequently return only the first sentence or first paragraph of a chunk. If that first sentence contains no useful answer, the source is not cited -- even if the answer appears in sentences two and three.

This makes Claude the most demanding platform for writing quality. A page can satisfy Perplexity's vocabulary requirements and still miss Claude's citation threshold if its sentences are ordered wrong. The answer must lead, not follow.

The fix is tractable. Take any service description and invert its structure. Instead of: "At ABC Plumbing, we understand that burst pipes are stressful. Our team of licensed plumbers is available around the clock to help you with any emergency." Write: "Emergency plumbers typically arrive within 60-90 minutes for burst pipe calls in Hamilton. ABC Plumbing dispatches from [neighborhood] and covers Hamilton and surrounding areas." Same information. The answer is now in sentence one.

The FAQ content specifications that follow from this

The Discovered Labs vocabulary finding and the Claude sentence-level observation converge on a practical content specification for FAQ sections, documented in our June 2026 content-format investigation.

**Answer length:** 40-80 words per answer is the extractable chunk. Below 40 words is typically too thin -- insufficient specificity for AI systems to cite as a source. Above 80 words and RAG retrieval systems may split the answer or return only the first chunk, which may not contain the complete answer. The 40-80 word range fits inside a single retrieval chunk and keeps the answer self-contained.

**Question language:** Use exact buyer phrasing, not professional category language. "How long does it take for an emergency plumber to arrive?" rather than "What is your service response time?" The first is what a homeowner under urgency types. The second is what the business wants to answer. These are not interchangeable; the vocabulary in the question sets the vocabulary in the answer, which determines whether it is retrieved.

**Answer structure:** Direct answer in sentence one, supporting specifics in sentences two and three. "Emergency plumbers in Hamilton typically arrive within 60-90 minutes, depending on call volume and location. ABC Plumbing maintains a 75-minute average response time across the Hamilton service area." The answer lands in sentence one. The specificity follows.

**Number of questions:** 4-8 per page. Fewer than 4 is insufficient surface area. Above 8 and answer quality tends to drop as the business runs out of specific answers and starts filling space with thin content that doesn't get cited.

Which platform to check first

The Discovered Labs finding tells you where to start diagnosing. If your Perplexity score is low while your ChatGPT score is moderate or higher, vocabulary mismatch is the likely explanation. Perplexity's literal retrieval makes it the most sensitive platform for vocabulary gaps -- a mismatch that ChatGPT's fan-out architecture partially compensates for will show directly in Perplexity citation rates.

If Claude is weak relative to Perplexity, the issue is more likely sentence structure than vocabulary. The content may be using buyer language already but burying the answer after context-setting sentences.

A Signal Check at sourcepull.ca returns per-platform scores across ChatGPT, Perplexity, Gemini, and Claude in about three minutes. The platform-by-platform breakdown shows which gap is driving your overall score -- and that determines whether vocabulary precision or structural rewriting is the right first investment.