
limelit
Connect Limelit Cloud: sign in from Claude Code with no API key, then see how ChatGPT, Claude, Perplexity, Gemini and Google AI answers rank and cite your brand, the prompts you lose, the sources to earn, and what to write next. Asks before anything that spends credits.
- version
- 0.1.0
- GitHub stars
- 447
Install
/plugin install limelit@limelitRun it inside Claude Code after adding the marketplace with /plugin marketplace add limelit-co/open.
Contents
Discovered from sourceSkills 8
Checking AI visibility
First move, every time
- Call
get_active_property. It names the brand, the engines tracked, and a setup block. - If
has_answersis false, nothing has run yet. Tell the user thenext_stepit returns, with its dashboard link, and report no numbers. - Otherwise call
analyze_my_visibility(trailing 30 days by default,days1 to 90). It returns brand visibility, citation share, the top competitors by mentions and the top cited sources in one call.
Going deeper
| The user asks | Call |
|---|---|
| Numbers for a set window | get_overview_kpis with days |
| How it moved over time | show_visibility_trend, or get_kpi_history |
| Share of voice against competitors | show_share_of_voice |
| Which engine is the weak one | get_matrix (one cell per prompt and engine) |
| Are the numbers current | get_run_activity |
Reporting rules
- Quote the n behind every figure: the number of answers it rests on.
- A figure resting on fewer than 20 answers is still settling. Say so.
- Never report a metric a tool did not return. Do not compute your own visibility from raw chats.
- Headline visibility leaves out prompts tagged as branded (prompts that name the brand).
get_matrixincludes them and flags them. - Name the engines a figure covers. Some engines are read through their API and some through scraped answers, and a window can mix both.
Next steps to offer
- Which prompts we lose: the
finding-lost-promptsskill. - What gets cited instead of us: the
finding-sources-to-earnskill. - What to write next: the
planning-what-to-writeskill.
Drafting and publishing
Write
generate_draft with source_url (the article to write against). Useful options:
brief_id: write from a reviewed brief (create_brief,list_briefs).target_query: the buyer query the post must capture, used in the H1, meta and first paragraph.style:how_to,opinion,data_breakdown,q_and_aorstory. Vary it across drafts.page_typewithpage_subject: ause_case,comparisonoralternativeslanding page.brand_strength:light,balanced(default) orstrong.
Drafting calls a model and can take minutes. It spends credits, so the plugin asks first.
Score and revise
score_content_qualitywithdraft_idscores 0 to 100 on seven checks and returns a rewrite instruction for each failing one. It calls no model and costs nothing, so run it after every revision.get_draft_structureandget_draft_keyword_coverageshow what the draft covers.update_draftsaves changes.- Add quotations and statistics only from sources you can link. Never invent one.
Publish
publish_draft with id makes the post public at a permanent URL, lists it in the sitemap
and exposes it to AI crawlers. It cannot be undone, so the plugin always asks first. Before
calling it:
- Show the user the title, the URL slug it will get, and the latest quality score.
- Get an explicit yes.
After publishing, offer the pushing-to-indexers skill so search indexes pick the page up.
For a site that publishes through its own CMS, use list_cms_destinations and
create_cms_draft instead.
Finding lost prompts
Start with get_active_property. If has_answers is false, give its next_step and stop.
1. Find the gaps
Call get_matrix (metric: visibility, sov or position). It is one cell per
prompt and engine, so it shows both the prompt you never win and the engine that is
weak across the board. Branded prompts appear here and are flagged; set them aside
when the question is about non-branded demand.
2. Read the answers that missed
list_chatswithshow=missedlists the stored answers that left the brand out, newest first. Narrow withprompt_id,targetordays.get_chatwith anidreturns one answer in full: the text, every brand found with its rank, every source cited, and the searches the engine ran to get there.compare_prompt_chatswith aprompt_idlines up every engine's answer to one prompt: who each mentioned, in what order, and what each cited.
3. Explain the gap
Look for the pattern across several missed answers, not one:
- Which competitors appear in your place, and at what rank.
- Which sources the engine cited. A page that keeps getting cited without you is a
source to earn (see
finding-sources-to-earn). - Which searches the engine ran. Those are the phrasings your pages need to answer.
Rules
- Quote n. A prompt with only a few answers per engine has not settled.
- Quote at most a short excerpt from an answer, and link the chat rather than pasting it.
- Answers vary from run to run. One missed answer is noise; a run of them is a gap.
- Do not start new runs to "check again" unless the user asks. Runs spend credits
(see
running-prompts).
Finding sources to earn
Start with get_active_property. If has_answers is false, give its next_step and stop.
The worklist
list_top_sources: the most cited sites in your category.list_source_gaps: the same sources ranked by gap, meaning how often a source was retrieved or cited versus how often your brand appeared alongside it. Pick the grain withdimension(domains,hosts,urls) and filter withsource_type(owned,competitor,informational,other).list_source_urlswithvalueset to a bare host (for exampleexample.com, not a URL) shows exactly which pages on that site get cited.compare_source_to_brandwithvalueset to a tracked competitor's domain puts its source presence next to yours over the same prompts.
Ready-made shortlists
find_easy_wins: pages where at least one tracked competitor is cited and you are not. These already have proven reach in AI answers.list_opportunities: the ranked Plan queue (win, defend, winning), each with a score, a proposed action and the citation evidence.
Choosing the action
- A competitor's own page cannot take an inclusion pitch. Write a stronger page instead
(see
planning-what-to-write). - A publisher, directory or review page can. Limelit drafts the pitch; the user sends it from their own mailbox. Limelit never sends outreach itself.
- An owned page that is cited less than before is a refresh candidate:
list_citation_drops.
Rules
- Quote n with every count.
- Never report a metric a tool did not return.
Planning what to write
Start with get_active_property. If has_answers is false, give its next_step and stop.
Pick the target
| Situation | Call |
|---|---|
| One best next post, with a reason | suggest_my_next_blog_post |
| The full ranked action list | list_opportunities (filter bucket: win, defend, winning) |
| Pages competitors win and you do not | find_easy_wins |
| Your posts AI engines cite less than before | list_citation_drops |
| Search Console demand you are missing | gsc_findings (see reading-search-console) |
Shape the piece
- The searches engines ran behind lost answers (from
get_chat) are the questions the page must answer.get_fanoutsandget_fanout_phrasescollect them. create_briefturns a source URL into a reviewable angle, outline and question list. Review it with the user before drafting.- A competitor's page wins? Consider a comparison or alternatives landing page instead of
a blog post (
generate_draftwithpage_type).
Then
Hand off to the drafting-and-publishing skill to write, score and publish.
Rules
- Recommend one thing first, with the evidence and its n, then the runners-up.
- Do not invent statistics or quotations for a plan. Every claim needs a source you can link.
Pushing pages to the indexes
Check first
get_indexing_status returns the IndexNow setup, each host's key file state, the Google
sitemap state, every pushable page with its last push, and each page's Google standing
(indexed, not_indexed, unknown_to_google, not_checked).
- A page with
eligible=falsesays why inreason. - A host whose key file is not served is skipped, never guessed.
- Google only takes per-page indexing requests by hand: give the user the page's
google_deep_linkand they click Request indexing in Search Console. - Brave takes manual submissions only; the status includes its form URL.
Push
Only push when the user asks for a push. It spends no credits, but it asks outside engines to recrawl the site, and a push cannot be unsent.
- Call
push_to_indexerswith nodry_run(preview is the default). Choose pages withurls(exact URLs from the status) orscope(posts,owned,all). It returns the exact batches it would send. - Show the user the batches and get a clear yes.
- Call again with
dry_run=false. The plugin asks for permission on this call.
Owner only.
Reading Search Console
These tools need a Search Console connection on the Limelit property. If one returns empty or not connected, tell the user to connect Search Console in the Limelit dashboard.
| The user asks | Call |
|---|---|
| Overall property picture | gsc_property_summary |
| Traffic for some or all pages | gsc_search_analytics (days up to 90, url_list empty for all) |
| What to work on | gsc_findings: content decay, cannibalisation, CTR gaps, quick wins, content gaps, traffic drops |
| Is a page indexed | gsc_url_status with the absolute url |
| The week in one view | gsc_weekly_digest |
A quick win or content gap whose query a tracked prompt already asks carries that prompt.
Cross-check it with finding-lost-prompts to see whether AI answers miss you there too.
Search Console is classic search, not AI answers. Keep the two kinds of numbers apart when you report them.
Running prompts
Runs spend credits. Never start one unless the user asked for a run.
Preview, confirm, run
- Call
reevaluate_all_prompts(orreevaluate_promptwith aprompt_idfromlist_prompts) with nodry_run, ordry_run=true. That is a preview: it returns the answers planned and the estimated credits, and runs nothing. - Show the user the plan and the cost. Wait for a clear yes.
- Call again with
dry_run=false. The plugin asks for permission on this call even if the Limelit server is allow-listed. - Poll
get_run_activityuntil its status isdone,failedorcancelled. Report answers completed and failed.
Fresh answers take time to show up in the headline numbers. Re-read with
analyze_my_visibility after the run is done, not during it.
Spend
get_spend_summary: today's and month-to-date cost, the daily and monthly budgets, and a day-by-day series. Use it when runs stopped: a budget cap is the usual reason.set_spending_capschanges the caps. Owner only, and it changes billing behavior, so confirm the new values with the user first.
Adding prompts
create_prompt adds a tracked prompt and starts spending on its runs. Confirm the wording and
that the user wants it tracked before calling it.
Commands 4
Use the finding-sources-to-earn skill. List the top easy wins and source gaps, group them by
the action each needs (stronger page, inclusion pitch, refresh), and recommend the first three
to work on with the evidence and n for each.
Use the finding-lost-prompts skill. Find the prompts and engines where the brand is missing
most often, read the missed answers, and explain the pattern: who appears instead,
which sources the engines cite, and which searches they ran. Do not start any runs.
If an engine or topic was given, focus on it: $ARGUMENTS
Use the planning-what-to-write skill. Pick the single highest-leverage post to write next,
explain why with the evidence and n, name the buyer query it must capture, and offer to create a
brief. Do not generate or publish a draft until the user says to.
If a topic was given, stay within it: $ARGUMENTS
Use the checking-ai-visibility skill. Report brand visibility, share of voice, citation share,
the competitors ahead, and the top cited sources over the trailing $ARGUMENTS days (30 if no
number was given). Quote the n behind each figure and flag any figure under 20 answers as still
settling. End with the one gap most worth closing and offer the skill that closes it.
Agents 1
You are an analyst reviewing how AI answer engines mention and cite one brand, using the Limelit tools. Work read-only: never call a tool that runs prompts, drafts, publishes, pushes, tracks or changes anything. If the review suggests one of those, recommend it instead.
Steps:
get_active_property. Ifhas_answersis false, return itsnext_stepand stop.analyze_my_visibilityfor the trailing 30 days, andget_overview_kpisfor 7 days to see the recent direction.get_matrixto find the weakest engine and the prompts never won.- For the two or three worst prompts,
list_chatswithshow=missedandget_chaton a few answers: who appears instead, what is cited, what the engine searched for. list_source_gapsandfind_easy_winsfor the sources to earn.list_opportunitiesandlist_citation_dropsfor the action list.
Return, in this order and in under 300 words:
- Headline: visibility, share of voice and citation share, each with its n.
- Weakest engine and the prompts most often lost, with n.
- Who wins those answers instead, and the sources behind them.
- The three actions most worth doing next, each tied to the evidence above.
Rules: quote n with every figure, call any figure under 20 answers still settling, never report a metric a tool did not return, and never invent a number.
Hooks 1
MCP servers 1
Sponsors
Details
- Marketplace
- limelit-co/open
- Source
- limelit-co/limelit-plugin
- Category
- Marketing
- Keywords
- geo, aeo, ai-visibility, share-of-voice, mcp