Stop Writing Marketing Copy. Start Writing Evidence.

Marketing copy is written to persuade a person. Evidence is written so a machine can lift one paragraph and stand behind it. AI answer engines do not choose pages, they choose passages: Google composes its generative results by retrieving specific information from pages in its index, and Microsoft says how content breaks into retrievable chunks now matters in ways ranking never measured. Every paragraph you want an AI to cite has to survive being pulled out of context and read alone, which means it has to pass five tests: completeness (does it answer the question by itself), freshness (can a machine tell it is current), authority (is it clear who is qualified to say it), attribution (is the claim traceable to a source), and entity clarity (does it name which business it means). Ask those five questions of every important paragraph and you turn copy that only reads well into copy a machine can quote.

Your best-written sentence and your most citable sentence are usually not the same sentence. One is built to move a reader toward a call. The other is built to answer a question completely enough that a machine will repeat it and attach your name. The flagship in this series showed that rank and citation are separate outcomes: we track the same buyer questions across Google, Google's AI answers, and ChatGPT, and the site that ranks is often not the site that gets cited. This piece is about the writing change that closes that gap.

AI systems cite passages, not pages

The unit of writing that matters has shrunk from the page to the paragraph. Ranking asked a page-level question: of all the pages about this topic, which one should a human open? Citation asks a passage-level question: of all the sentences on all the pages, which specific one can support this answer? Those are not the same question, and the copy that wins the first can lose the second.

Microsoft's search team describes the same mechanism from the other side: in grounded AI systems, factual fidelity and how content breaks into retrievable chunks matter in ways that never appeared in a ranking signal. A ranking system reads your whole page in context. A grounding system tears one chunk loose and reads it alone, with none of the surrounding copy to lean on. Classic SEO never had to survive that. Citation does.

So the practical test changed. You are no longer only asking "is this page good enough to rank?" You are asking, of each important paragraph, "if a machine lifted this and nothing else, would it still be true, current, credible, sourced, and clearly about my business?" Those are the five questions below. Run them like a checklist.

Completeness: a paragraph that leans on the rest of the page cannot be cited

Question one: does this paragraph answer the buyer's question on its own? A citable passage names the who, the what, the where, and at least one checkable proof point inside the same block of text, because the machine will not read the rest of the page to fill in the gaps.

Watch what completeness does to a single sentence:

"We're passionate about helping our clients get the results they deserve."

"Redwater Physiotherapy treats post-surgical knee and shoulder rehabilitation in Sherwood Park, with direct billing to most major insurers and evening appointments."

A human skimming the page takes roughly the same brand impression from both. An extractor can only use the second, because it is the only one that contains information: a name, a service, a location, and two concrete facts a buyer would act on. The first sentence is not wrong, it is empty, and an answer cannot be built out of empty. If your key paragraphs depend on the headline above them or the paragraph before them to make sense, they will not make sense once they are pulled loose.

Freshness: a machine cannot cite what it cannot date

Question two: can a machine tell this passage is current? Grounding systems prefer information they can trust is fresh, and the only recency they can read is the recency you write down: a date, a reference to the current year, an updated figure. "Latest" and "up to date" are invisible, because they were equally true and equally unprovable five years ago.

Compare:

"We use the latest roofing materials and techniques."

"In 2026 we install Class 4 impact-rated asphalt shingles across Edmonton, which may qualify homeowners for a hail-damage insurance discount."

The second sentence carries a signal the first does not: a year a machine can compare against today, plus a specific, checkable detail. That is the difference between a passage a grounding system treats as current and one it has to guess about, then often skips. Put the date in the sentence, not just in a "last updated" line in the footer that the extracted chunk leaves behind.

Authority: name who is qualified to make the claim

Question three: is it clear who is qualified to make this claim? Authority is not a dial you read on a dashboard, it is a property you write into the passage. A credential, a licence, a first-party basis, or a named person tells a citing system the claim comes from someone entitled to make it.

That is the honest nuance. The report launched February 10, 2026 and it scores citations, not authority, so no tool hands you an authority number to optimize. You build authority into the sentence and let the citation follow. Compare:

"Trusted by many happy patients."

"Dr. Amrit Sahota, a licensed Alberta dentist since 2009, places same-day crowns using in-house CEREC milling."

The second names a person, a credential, a licence, and a method. It reads as a claim someone stands behind. The same move works with first-party specificity: "trusted by businesses across Alberta" says nothing a machine can weigh, while "Choice OMG monitors more than 60,000 data points a day across 33 client accounts" is a first-party basis a citing system can attach to a source. Vague trust is not authority. Named, checkable qualification is.

Attribution: an unsourced number is one an AI will not repeat

Question four: can this claim be traced to a source? ChatGPT search responses can include inline citations and a Sources panel, which means a citing system is actively looking for claims it can forward with a source attached. A bare, unattributed statistic gives it nothing to stand behind, so it tends to reach for a claim that names its origin instead.

Compare:

"Most Edmonton businesses are invisible to AI answer engines."

"In our own July 2026 cross-engine tracking, a question we rank #4 for on Google was cited by neither Google's AI answer nor ChatGPT."

The first is a floating assertion with no owner. The second names who measured it and when, so a machine forwarding it knows exactly what it is standing on. You do not have to publish someone else's study to do this. Your own dated, first-party observation is a source. The rule is simple: if a paragraph makes a number-shaped claim, the same paragraph should say where the number came from.

Entity clarity: the passage has to say which business it means

Question five: does the passage name which business it means? Once a paragraph is lifted out of context, "we," "our team," and "the company" have no referent, because the page header that told the reader who "we" was is gone. The sentence itself has to carry the name.

Compare:

"We've been the area's go-to team for over a decade."

"Northgate Auto Repair has served north Edmonton drivers since 2011."

A reader on the page resolves "we" from the logo at the top. An extractor holding only the sentence cannot, so the first version becomes an anonymous quote about nobody. The second survives extraction because the subject is named inside it. This is strictly about the words in the paragraph: make the business the grammatical subject of your important claims instead of a pronoun. Keeping your name, address, and listings consistent across the wider web is a real and separate job, and it is not this one; this question is only about whether the sentence on the page names itself.

Run the five questions before you publish

Treat the five questions as a QC pass, the same way you would run a health check before a deploy. Marketing that runs like infrastructure does not ship a page and hope; it checks the page against a standard first. Take your most important service page, read each key paragraph as if it were the only thing a machine could see, and put it through the list:

A paragraph that passes all five can be lifted out of your page and still stand as evidence. A paragraph that fails one is copy that reads well and cites badly, which is exactly the gap the flagship measured: rank and citation are separate outcomes, and the writing is where you close the distance between them. Most service-page copy fails at least two of these questions today, not because it is bad writing, but because it was written to persuade a reader who could see the whole page, not to feed a machine one chunk at a time.

This is the work our SEO practice now runs on client pages: not just chasing positions, but rewriting the passages that carry the claims so a machine can quote them. If you want a second set of eyes on whether your most important pages read as evidence or just as marketing, ask us and we will run the five questions on the pages that matter most to your buyers.