> ## Documentation Index > Fetch the complete documentation index at: https://docs.synscribe.com/llms.txt > Use this file to discover all available pages before exploring further. # How to get cited by AI answer engines (turn a research trace into keywords) > Use the fanout-content-mining skill to turn an LLM's research trace — the fan-out queries it fired and the pages it read — into keywords and content angles optimized to be cited by AI answer engines, read through a similarity lens and a gap lens. Give Pi an **LLM research trace** — the fan-out search queries an LLM fired and/or the pages it read while answering a buyer's question — and ask it to run the **`fanout-content-mining`** skill. Pi works a funnel (raw fan-out query to keywords to content ideas) and reads the trace through two lenses: **similarity** (ride the topics and formats LLMs already retrieve, so your content lands in the set they cite) and **gap** (find what the LLM kept re-asking or read but couldn't get an answer from, and fill it so the LLM goes "huh, that's new"). It's the operational move behind [Part 1 §1.2 — two search surfaces](/theory/two-search-surfaces): you're no longer ranking #1 for one phrase, you're being citable across the machine's whole fan-out. The output is a report of keywords and single-angle content ideas you can hand to [Creating blogs](/platform/creating-blogs). > 🎬 **Video planned:** supply a research trace, mine the fan-out into keywords, pick angles > optimized for LLM citation. See the shot-list. ## Where the trace comes from The whole play depends on watching the machine's actual queries instead of guessing them. Synscribe ships two ways to capture a trace (per [§1.2](/theory/two-search-surfaces)): a Chrome extension that reveals ChatGPT's web-search queries, and **Birdseye** (a macOS app) that shows Claude Code's queries and the pages it read. Either gives you the two inputs the skill wants: * **Fan-out query log** — the searches the LLM fired, in order. This is buyer *intent*, and where a query keeps recurring, *unmet demand*. * **Pages-read log** — the title, extract, and URL of each page it opened, plus the search term that surfaced it. This is what *answered well* versus what *failed* (a 403/404, or a "cannot answer from the content provided" / marketing-only page — both count as the answer being absent). One log is enough; both is better. Paste in the seed buyer question too if you have it. ## What Pi produces The skill keeps the funnel — it doesn't flatten it: 1. **Reads the fan-out** and keeps the queries carrying real buyer intent in *verbatim* form (the exact string is the unit it decomposes), tagging each with intent, entities/jargon, a gap signal (did it hit a wall?), and a format signal (comparison / directory framing = the LLM wants side-by-sides). 2. **Fan-out to keywords**, working each query in both directions: *abstract up* to the de-branded, category-level concept the buyer was circling (the most ownable keywords), and *decompose down* into the targets hiding inside a fat, entity-stuffed query. It flags a competitor's proprietary API/webhook name as a low-ownership keyword and abstracts it up to the concept instead. Every keyword is tagged similarity or gap, with the source query (or read page title) as evidence. 3. **Keywords to content ideas**, one keyword at a time, going wide: several single-angle pieces per keyword covering direct patterns (informational, comparison, decision, BOFU, regional cut) and the sharper *lateral* ones (consequence/economics, jobs-to-be-done, persona cut). Each idea names its type, title, the exact question it answers, the play, one line on *why an LLM would cite it*, and the must-include entities. **The highest-leverage move is originating the answer nobody has published.** For the strongest gaps, the skill deliberately proposes net-new pieces — original data/benchmarks, a coined framework or taxonomy, a point-of-view thesis, or a brand announcement — the kind that makes an LLM cite *you* as a primary source rather than one more echo. Every net-new idea carries a visible `NEEDS VALIDATION` flag naming exactly what to verify before publishing (never fabricate a launch or a metric). ## Read the two lenses right > **Demand ≠ saturation.** "The LLM kept asking" = demand. "The LLM found a good answer" = saturated. > Demand *plus no good answer* is the best opportunity — weight those highest. Ride the retrieved > formats for similarity; deliberately produce the missing shape for the gap. The skill also filters for **ownership**: prefer category, decision, comparison, and taxonomy pieces where competitors appear as examples *inside* the piece, not as its subject. It won't propose content that just explains a rival's product — that educates buyers about a competitor and reads oddly coming from you. ## Real prompt Paste (or point Pi at) the captured trace and let it run the funnel: ```text theme={null} here's the fan-out query log and the pages-read log from a ChatGPT session where a buyer was researching [the buyer question]. run fanout-content-mining: give me the keywords and blog/LP ideas optimized to get me cited by AI, and rank the gaps worth owning. flag anything that needs validation before I publish. ``` Pi returns a skimmable report: the seed question, the verbatim queries worth working, the keywords (each tied to its source query), the content ideas grouped under their keyword, and a closing ranked "gaps worth owning" table. Take the ACCEPT-worthy keywords through [keyword evaluation](/platform/evaluate-serp-cliff) before building, and hand the winning ideas to [the blog writer](/platform/creating-blogs). *** **Why (theory):** [Part 1 §1.2 Two search surfaces](/theory/two-search-surfaces) · **Feeds:** [2.2.2 Evaluate keywords](/platform/evaluate-serp-cliff) · [Creating blogs](/platform/creating-blogs) · **Sibling how-tos:** [2.8.1 Fix a page that won't rank](/platform/fix-page-not-ranking) · [2.8.3 Bring your own DataForSEO key](/platform/byok-dataforseo)