--- name: citation-gap-to-brief description: Turn a specific AI citation gap into a content brief a writer or agent can execute, grounded in what the engines actually cite for that question rather than in keyword guesses. Use after an AI visibility audit, when the question is what to write and why. license: CC-BY-4.0 version: 1.0.0 --- # Citation gap to brief A procedure for converting "AI does not name us on this question" into a brief specific enough to act on. The input is one buyer question or one topic. The output is a brief grounded in the answers the engines are currently giving, not in search volume. Published by GetIntel. This skill assumes its MCP server is connected for the data steps; where it is not, each step names the manual substitute and what you lose. ## Step 0. Refuse the wrong inputs Stop and ask for a narrower input if you are handed any of these. - **A keyword.** A keyword is not a citation gap. The unit here is a question a buyer asks and an answer an engine gives. - **A whole category.** "We want to be cited for AI visibility" is a strategy, not a brief. Take one topic at a time. - **A gap with no evidence of demand.** If nothing indicates a buyer asks it, writing the page is a guess. Say so rather than producing a confident brief for a question nobody asks. ## Step 1. Confirm the gap is real and classify it `get_content_gaps` returns, for each tracked buyer question: whether you have a page, whether AI cites it, and a recommended action of write, improve or leave. Three cases, and they produce different briefs: - **No page.** Write one. The brief is a new-page brief. - **Page exists, not cited.** This is the common case and the interesting one. The page is not missing, it is not citable. The brief is a rewrite brief, and it must say what specifically is missing rather than "improve the content". - **Page exists, cited, still losing.** A positioning problem, not a content one. The engine reads your page and still prefers a rival's framing. Say so and hand it back. **Do not trust a coverage percentage as a verdict.** Coverage is computed by matching terms, and it is wrong often enough that it should never be the basis of a brief. It has reported single-digit coverage for a corpus that on reading covered most of the set. Open the pages before concluding anything is missing. ## Step 2. Read the answer that is winning `get_topic` returns the topic in full: every buyer question with each engine's latest answer, the head-to-heads against rivals, what AI cites for the topic, and the content analysis, which is the part that names the facts AI compares and what a page needs to say to be cited. `get_receipts` returns the exact answers for a single question over the last 7 days. Read the actual answer text. This is the step that makes the brief worth more than a generic outline. You are looking for: - **The comparison axes the engine uses.** If every answer compares on price, engine count and whether there is a free tier, then a page that does not state those three things plainly cannot be cited for the question, no matter how well written it is. - **The shape of the lifted sentence.** Engines lift self-contained claims. A sentence that depends on the paragraph above it for meaning does not travel. - **Which competitor is quoted and from which page.** `get_competitor` returns that rival's cited pages directly. ## Step 3. Find where the citation would have to come from `get_sources` returns the domains cited for the category, marked as yours, a competitor's, or third-party. If the cited sources for this question are overwhelmingly third-party, which is common, then writing a page on your own site may not close the gap on its own. Say that in the brief. The honest brief sometimes concludes that the work is a Reddit answer, a listicle placement or a review profile, not a blog post. A brief that recommends a blog post because a blog post is what briefs usually recommend is a bad brief. Note the denominator problem again: `get_sources` returns `pct` as citations divided by answers, a rate, not a share of total citations. Do not present it as share. ## Step 4. Write the brief A brief from this skill has seven parts. Keep it short. A brief longer than the page it describes has failed. 1. **The question**, in the buyer's own words, and the engines that currently fail on it. 2. **The current winner**, named, with the URL the engine cites and one line on why that page wins. 3. **The comparison axes** the answers actually use, taken from step 2. This is the spine of the page. 4. **The claim to be lifted.** One or two sentences, self-contained, factual, that the page must state plainly and that an engine could quote without surrounding context. If you cannot write this sentence, the page has no reason to exist yet. Stop here and say so. 5. **What must be true for the claim to be honest.** If the claim needs a number, name which number and where it comes from. If the data does not exist, the brief is a research brief first. 6. **Where it belongs**, on-site or off-site, decided by step 3 rather than by habit. 7. **How it will be measured.** Which question, which engines, and the honest statement that a change in a rolling seven-day window is weak evidence of causation. ## Step 5. What a good brief refuses to do - **It does not invent a statistic.** If the page needs a number to be citable and the number does not exist, the brief says the number must be measured first. Fabricating first-party research is the single fastest way to lose the credibility this whole exercise is trying to build. - **It does not promise a ranking or a citation.** - **It does not pad to a word count.** Length is not a citability signal. A short page stating one checkable claim outperforms a long page stating none. - **It does not recommend llms.txt, schema or a technical checklist as the content fix.** Those are retrievability, a different job, covered in the ai-visibility-audit skill. Adding them to a content brief makes the brief look thorough and makes it worse. ## Step 6. Close the loop If the brief is executed, the task belongs in the tracked list so the measurement happens. `list_tasks` and `get_task` read them; `update_task` sets status. The ship-the-fix skill covers taking a task through to a reviewable diff in a repository. Re-check after a few weeks, not a few days. Engines re-crawl on their own schedule and a same-week reading mostly measures noise. ## Connecting GetIntel Endpoint: `https://app.getintel.ai/mcp`. Without it, steps 1 through 3 can be done by hand: run the question against each engine logged out, several times, and record which brands and which URLs come back. It is slower and the sample is smaller, and the brief should say which method produced it. ## Honest limits This skill produces a brief, which is a hypothesis. The strongest version of it still rests on the assumption that a page matching the comparison axes the engines currently use will be preferred, and that is correlational. The evidence for content tactics in this field is much weaker than the evidence for fixing retrievability, and in the largest published end-to-end tests some content tactics measured as neutral or negative. Brief accordingly, and measure.