agentleFS
XPOZpublicXPOZpublic/geo-seo-agent/.claude/skills/content-gap-analysis/SKILL.md

content-gap-analysis

Skill0 starsChanged 3 months agoMITView on GitHub

Analyze GEO content gaps. Find prompts and topics where the company is absent from AI answers (and search results, where measured), diagnose why, and propose prioritized content briefs.

6 other projects use a version of this.

Adopted by

Most recently taken up by

Used with

Projects with it also use technical-seo-checker, serp-analysis and 1 more.

Other files people use alongside this one.

Files found in the same public projects as a version of this one, more often than they turn up anywhere else. Each project counts once, large collections count for less, and this file's own repository is left out.
  • opencueSkill

    technical-seo-checker

    Use when the user asks to "check technical SEO"; audits crawl, indexing, CWV, mobile, security, redirects, and evidence gaps. 技术SEO/网站速度

    Also in 3 projects that use this

  • opencueSkill

    serp-analysis

    Use when the user asks to "analyze SERPs"; reviews ranking factors, features, intent, AI Overviews, and snippets. SERP分析/搜索结果

    Also in 3 projects that use this

  • indranilbanerjeeSkill

    keyword-research

    Research keywords via the connected data source: intent, gaps. Grouping → keyword-cluster. \"what keywords should we target\"

    Also in 5 projects that use this

The file

We split this skill into the rules it gives an agent

1 rule. 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • 0

    Common

    Most of the 6 edited versions we read keep these.

  • 1

    Distinctive

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Every brief also carries its evidence (the snapshot findings that motivated it) and its success criteria (which prompts should move, …

    line 62

“Common” is measured: kept by at least 6 in 10 of the 6 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Content Gap Analysis

The judgment stage of the loop. /visibility-measure records facts; this skill reads them, diagnoses why the losing prompts are being lost and what makes the winning ones work, and turns the diagnoses into prioritized, evidence-backed briefs for the human to approve. It sits between the sensor and the production skills, which all refuse to invent topics, so nothing gets written that didn't pass through here.

Everything below is guidance with its reasoning, not a checklist: deviate when the evidence argues for it and say why. Two hard constraints: briefs become actionable only through human approval (flipping proposed → approved is the human's edit, never this skill's), and every proposal must be traceable to snapshot evidence; a suggestion the data doesn't support doesn't get proposed.

Process

1. Orient

  • The latest snapshot in memory/visibility-snapshots/ and its deltas summary. If no snapshot exists for the current week, invoke /visibility-measure first; diagnosing from stale readings produces stale briefs.
  • The prompt panel (strategy/target-prompts/target-prompts.md): Value anchors priority.
  • The content plan (strategy/content-plan/content-plan.md): the routing default and which channels are in scope.
  • The content index (content/index.md): what exists and what's already proposed/approved/in-production; never propose a duplicate of work already in flight.

2. Diagnose each lost or weak prompt

Classify each active prompt from the snapshot (won / partial / lost, per engine), then find the root cause for the ones being lost. The cause decides the fix; proposing content before diagnosing is how wasted briefs happen. Common causes, as guidance rather than an exhaustive taxonomy:

  • No asset serves the intent: nothing on the site answers this question → new content, routed per the content plan's routing default (intent shape suggests the format, but the winning citations for the prompt decide; override with evidence and say why in the brief).
  • Asset exists but isn't cited: audit the live page before concluding anything: answer-first opening? extractable paragraphs? fresh modified date? FAQ? Pull its Search Console standing where connected, and check the latest memory/site-readiness/ report for the page's mechanical readings (structured data, extractability share, crawler access); if none exists or it predates a site change, request a /site-readiness run. The rework brief must name specific findings; /content-rework hard-stops without a diagnosis, by design.
  • Surface split: cited by one engine but not another, or ranking on Google while absent from AI answers (and vice versa). The split itself is diagnostic: a page Google ranks but engines don't cite usually has an extraction problem, not an authority problem.
  • The answer is assembled from surfaces you're not on: listicles, review sites, community threads, directories. No amount of own-domain content fixes this. For the standard channels (review sites, marketplaces, GitHub, integrations, all active by default), propose the presence work directly: a proposed row in the listings/ roster (or content/github/ for owned-repo work) with the citation evidence, consumed by /channel-presence after approval. When the surface is one the plan doesn't work (community threads most commonly, or a channel marked n/a), record it as scope-change evidence for the content plan instead; never propose rows the plan can't execute.
  • Unwinnable as phrased: dominant established players own it on terms the company doesn't compete on. Note it for /target-prompts-discovery to consider retiring; the panel has one writer, and it isn't this skill.

The snapshot names that a prompt is lost; the raw run it links to shows how: the full answers, every page each search retrieved (not just yours), and the exact URLs cited. That is the evidence for judging what winning citations look like and whether an answer is assembled from surfaces you're not on. visibility.json is far too large to read whole; slice it per prompt using the recipes in ../visibility-measure/references/visibility-json.md.

The engines can also be asked anything live: the ask scripts run any prompt with web search enabled and save the full transcript under recordings/chats/<engine>/. A live probe is a single sample (noisy, per the panel's own caveat), and proposals stay traceable to snapshot evidence.

3. Learn from the won prompts

Step 2's mirror, on the prompts where the company is cited. Won prompts are the program's only ground truth for what earns citations: the strategy's success factors (answer-first, freshness, metadata) are borrowed assumptions until your own winners confirm, refine, or contradict them.

Per won prompt, read the winning samples in the raw run: which page was cited, on which fan-out query, and what the page does: format, opening, freshness, extractable structure, where and how the brand is named. The sharpest evidence is contrast with step 2: a trait separating your cited pages from your retrieved-but-passed-over ones is a stronger signal than any external best practice.

The patterns found become inputs, not a report: a rework brief names the in-house winner whose shape the page should adopt, a new-content brief borrows the winning playbook, and the run record keeps the patterns so they compound across runs instead of being rediscovered. A winner that contradicts a strategy principle is a routed finding (§7), not a silent observation. Small-sample caution applies here as everywhere: two samples per engine make one week's winner a hypothesis; a pattern earns trust by holding across prompts and weeks.

4. Judge the declining pages (when Search Console is connected)

The snapshot flags leading pages losing traction; deciding whether a decline is real is this skill's call, not the sensor's. Weigh trend length, seasonality, and cannibalization by newer pages before concluding. A real decline is a rework candidate even if the page targets no panel prompt: existing equity is cheaper to defend than new ground is to win.

5. Mine product-usage gaps (when usage signals exist)

What paying users actually do with the product (analytics, support questions, sales-call themes) reveals workflows without a page: gaps no external research would surface. Ask the operator what signals exist and weigh them alongside the panel-driven gaps.

6. Prioritize

Judgment anchored on three questions per candidate: how valuable is the prompt (its Value on the panel: closeness to purchase), how winnable is it (who's cited today, from the snapshot's competitor data), and how severe is the gap (fully absent vs. mentioned-but-not-cited). Propose briefs sized by the actual need the diagnosis found; every diagnosed gap the evidence supports can become a brief. Don't pad a thin week or trim a rich one. The prioritization still matters: order the batch so the human can approve the top and cut the tail.

7. Write the proposals and route the rest

Per accepted candidate, in the channel the diagnosis routed it to:

  • Produced content (action: new-content / action: rework, routed to self-published or docs per the content plan's routing default): an item folder (content/self-published/items/<slug>/ or content/docs/items/<slug>/) with brief.md, and a proposed row in that channel's index. A new-content brief carries target prompt IDs, content type (per the content plan), angle, key points, primary + secondary keywords, meta description. A rework brief carries the live URL, the named diagnosis (what specifically is failing), plus target prompt IDs.
  • Presence channels: a proposed row in the listings/ roster (gatekeeper surfaces: review sites, marketplaces, directories, integration listings, community lists) or in content/github/ (owned-repo work), with the rationale and evidence in the listing's folder where the conventions define one. Consumed by /channel-presence after approval.

Every brief also carries its evidence (the snapshot findings that motivated it) and its success criteria (which prompts should move, roughly when; expect the lag the sensor documents). Evidence is what makes review possible; success criteria are the thread retro will pull.

End the run with the proposals summarized for review (gap, evidence, proposed fix, expected outcome) plus the non-brief findings routed to their owners: retirement notes for discovery, competitor and market entity learnings refreshed in memory/competitors/ per that index's conventions, anything needing eyes in the "awaiting human" list.

When the citation landscape argues for changing the plan's scope (opening a non-standard playbook surface, reversing an n/a, or refocusing a channel that consumes effort without producing citations; the snapshot's "Where the answers come from" reading is the input to watch), append an evidence-backed proposal to the Proposed changes section of strategy/content-plan/content-plan.md, naming the snapshot/run-record evidence behind it. Approval stays the human's edit; the approved change carries its evidence into that plan's change-log.

Write the run's diagnostic record to memory/content-gap-analysis-runs/<date>.md (conventions in that folder's index): the full diagnoses and their evidence, the prioritization reasoning, and the deferred candidates: diagnosed gaps that didn't make the batch are the next run's head start, and the record is where retro finds what each brief predicted and why. With the reasoning there, the journal entry stays brief. Update the journal and memory/state.md per ground rules.

Never

  • Writes or edits content: briefs only.
  • Edits the panel or approves its own briefs or plan proposals.
  • Proposes work the content plan can't execute (non-standard surfaces, channels marked n/a): it may propose the scope change (evidence-backed, §7), never rows into it before approval.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.