Most agencies guard their content process like a recipe. We’re publishing ours in full, because the process is the proof of expertise, and because a reader who follows every step still ends up where our clients end up: with a system that requires the discipline to run it.
Some background on why ours looks the way it does. We published the 36 AI search visibility factors, our framework for how ChatGPT, Claude, Gemini, and Perplexity decide which websites deserve citations, and rebuilt our own site against all 36. We built our own citation-verification tooling because no tool on the market checked claims against sources the way our standard required. Then we rebuilt how we write, because a site built for AI citations is wasted if the articles on it can’t survive fact-checking. What follows is the pipeline every article on this blog goes through, including this one.
.
What does an AI-citable article actually require?
Three things: a topic AI engines are already answering; claims an engine can verify, and structure an engine can extract. Miss anyone, and the other two don’t save you.
That’s the whole pipeline in one sentence. Topic selection makes sure you’re writing for questions where citations are actually being handed out. Verification makes sure every claim survives being checked against its source, because the first person to open that source may be a buyer, a competitor, or an auditor. A claim that contradicts its own citation is the cheapest credibility to lose. Structure makes sure the answer can be lifted out cleanly. The research points the same direction as the verification thesis: the Princeton-led GEO study (Aggarwal et al., KDD 2024) found that adding citations, quotations from credible sources, and statistics boosted a source’s visibility in generative engine responses by up to 40%. Engines reward pages whose claims are backed by checkable evidence.
Here’s each step as we actually run it.
Step 1: Topics come from citation gaps, not keyword lists
We don’t start with keyword tools. We start with a map of buyer questions and who gets cited for each one.
Our Citation Audit asks 50 real buyer questions to the AI engines, repeats each measurement, and records which domains get cited per question. The rows that matter are the ones where an engine cites a competitor and not you. Those questions are proven citation-worthy; the engine just points elsewhere. Each row is an article assignment. Questions where nobody gets cited go to the bottom of the list, because there’s no evidence the engines want to cite anyone there yet.
This is the difference between content strategy and content hope. Without the map, topic selection is guesswork dressed in a keyword report.
Step 2: The brief comes before the draft
No article gets written from a blank page. Every one starts from a brief that fixes the structure and, more importantly, the evidence.
The brief lists the section structure, the specific data sources for every claim we intend to make (report name, publisher, date), the schema requirements, and a must-avoid list we treat as hard rules: no invented statistics, no fictional case studies, no claims about our own results we can’t back with our own data. A claim that can’t be sourced dies at the brief stage, cheaply, instead of surviving into a published article where an AI engine or a competitor catches it.
Step 3: Write to a standard, with answer blocks on every heading
The draft follows a written standard, and its core is the answer block. Every H2 is phrased as a question. Directly under it comes a one-or-two-sentence answer. Then supporting detail: a dated statistic with a named source, a concrete example, or a comparison. The main answer to the article’s central question appears in the first hundred words, because AI crawlers don’t reward suspense.
The standard also requires that every number carries a named source and date, that each article contains at least one claim specific enough to be proven wrong, and that sources do real work rather than decorating the text. Each writing session outputs three files, not one: the article, an FAQPage schema file, and an Article schema file, written in the same pass so the structured data never drifts from the content. The correlation is documented: AccuraCast’s 2025 study of 9,000 AI citation sources found 81% of cited pages carry schema markup. AccuraCast itself frames that as correlation, not a citation requirement, and controlled tests suggest adding schema alone doesn’t move citations. Schema consistency is factor 6 in our framework for a narrower reason: structured data that contradicts the visible content confuses parsing, and confusion is what keeps a parseable page from being understood.
Step 4: The rule we never break: write first, verify second, never rewrite after verifying
This step is why clients hire us, so it gets the most detail.
After the draft is finished, it goes through our citation checker, a tool we built ourselves. It extracts every factual claim, opens the live source behind each one, and returns a verdict per claim: supported, not supported, source inaccessible, or conflict between sources. Claims whose sources can’t be confirmed get removed, not kept with a hopeful footnote. When a removed claim was load-bearing, the conclusion resting on it gets softened or cut too. The final gate is human: flagged claims are read and decided by a person.
The order is the rule. Write to the standard first, then verify, and never restructure the article after verification, because any sentence written after the check is a sentence the check never saw. A polish pass that “just tightens the wording” after fact-checking quietly reintroduces the exact risk the fact-check removed. If a verified article needs restructuring, it goes back through verification. Slower and non-negotiable.
We earned this discipline the honest way: watching our own drafts fail our own checks. Sources that looked solid turned out not to contain the figures attributed to them. Numbers that were verified belonged to a different company than the one named. None of it was visible from inside the draft. It surfaced only when a tool opened every source and compared. That experience is baked into the Verified Content Retainer: every client article passes the same gate before publishing.
Step 5: Publish onto a site built for the 36 factors
An article inherits the site it lives on. A perfectly structured, fully verified article published on a slow, schema-broken, poorly crawlable domain is handicapped before any engine reads a word of it.
So the publishing environment is part of the pipeline. Our site is maintained against all 36 factors (crawler access, load performance, heading hierarchy, structured data consistency, internal linking, and freshness), and every new article slots into that environment with correct schema and contextual internal links. Freshness is a factor for a reason: Ahrefs’ analysis of 17 million AI citations found AI assistants cite content roughly 26% fresher than Google’s organic results, with ChatGPT showing the strongest preference for newer pages. The average AI-cited page is still about 2.9 years old, so freshness helps at the margin rather than expiring content after a year.
If you want to know where your own site stands, that’s exactly what the free 60-second audit measures: your score across all 36 factors, with a prioritized fix list. It checks readiness, not citations, which is the right first question: if the foundation fails, nothing you write on top of it gets seen.
Step 6: Re-measure, and only then claim anything worked
The pipeline ends where it started: measurement. After publishing against a set of target questions, we wait for the engines to re-crawl, then re-run the same questions and compare. Cited where you weren’t before? The gap-targeting worked for that question. Still absent? The article goes back for another pass, usually on specificity, because vague content is invisible to engines choosing between sources.
The self-serve version of this loop is our AI Citation Tracker: five buyer questions, tested across ChatGPT, Claude, Gemini, and Perplexity with 3x repeated runs, results in about 90 seconds and $5 to start. The full version is the $1495 Citation Audit: 50 questions built for your category, every citation verified, every finding evidence-graded, credited in full toward any fix engagement. The rules behind both are public in our methodology.
We hold ourselves to the evidence rule in our own reports: an observation is not proof, and no change gets credited until a before-and-after shows it. We ran this whole system on ourselves and published the baseline. The re-measurement follows, whatever way it goes.
Frequently asked questions
What is a write-then-check workflow? A content workflow where the article is fully written to a structural standard first, then every factual claim is verified against its live source, and no rewriting happens after verification. The order exists because sentences added after the check are sentences the check never examined.
Why phrase headings as questions? AI engines are question-answering systems. A question heading followed by a direct answer gives the engine an extractable unit that maps onto queries buyers actually type. A vague statement heading forces the engine to reconstruct the answer, and it usually cites a source that didn’t make it work that hard.
Do you need a fact-checking tool, or can this be done manually? Manual works at low volume: list every factual claim, open every source, and confirm each one. The tooling exists because at one or more articles a week, the manual version is the step that quietly gets skipped. And it’s the step protecting everything else.
How long does the full pipeline take per article? The brief takes one to two hours, the draft several more; verification runs in minutes, and human review of flagged claims can add an hour. It’s slower than writing without a pipeline. It’s much faster than recovering trust after publishing something wrong.
The shortest path is to run the free audit to score your site against the 36 factors, test five buyer questions in the tracker for $5, and write one article for that question through the six steps above. One question, one article, one re-measurement. If you’d rather it were done for you and proven, that’s the platform.