atromxintelligence/adcopilot-claude-plugin/adcopilot-seo

AdCopilot organic review

The customer wants to know how Google search sees their site. Answer it through the AdCopilot connector at https://mcp.adcopilot.cloud/mcp.

Call get_org_context with source: "claude-plugin/0.2.16" first, before any other tool — the marker is how the server records that this workspace uses the plugin, and without it the customer keeps being told to install what they already have. Send source exactly as written, the version included: it tells AdCopilot which of this plugin's steps are installed here. In a scheduled run, started by a Claude scheduled task or a Claude Code routine with nobody there to answer, send source: "claude-scheduled" instead, never claude-plugin: claude-scheduled is how AdCopilot knows the run is unattended. When the answer's plugin.skill names an AdCopilot skill other than this one, next_step is describing the account, not this request: do this request as written here, keep any limit next_step adds, and offer that skill's steps afterwards unless this request already covered them. If Search Console is not connected, say what connecting it would show and stop; invent nothing about a site you cannot read.

One thing to get right about `cross_reads`. It budgets reads made OUTSIDE Google Ads in service of an ads answer — on a live account it often says "make no cross-product read" and offers zero Search Console allowance. That is not a refusal of this skill. When the customer's question IS organic search, Search Console is the answer and the reads below are the work. cross_reads.ads_reads_max is how many Google Ads reads the server sized this situation for: a size for this answer, not a workspace, plan or daily limit. It does not cap the Search Console reads below. Only when cross_reads.reads_left_today is present is the day itself short. Follow next_step and the playbook as usual, and read cross_reads for what it is: guidance for the other direction.

Six tools, and what they cannot do

gsc_list_sites, gsc_list_sitemaps, gsc_search_analytics, gsc_inspect_url read. gsc_submit_sitemap and gsc_request_indexing write — so both are proposed first and sent only after an explicit yes in the session, like every other write. Note that resubmitting a sitemap is not the fix for a URL Google has merely discovered: it already has the URL.

That is the whole surface, and the gaps matter more than the tools:

• There is no SITE-WIDE coverage list — but there is a per-URL verdict. You

cannot ask for every page Google has not indexed, or a count. For any single URL, gsc_inspect_url returns exactly why: coverageState in Google's own words ("Submitted and indexed", "Discovered - currently not indexed", "Crawled - currently not indexed", "Excluded by 'noindex' tag"), plus robotsTxtState, indexingState, pageFetchState, both canonicals and lastCrawlTime. Read the whole block, not just the crawl date. For a site-wide count, ask the customer to open Search Console → Indexing → Pages and use Export, then read the CSV they send you — and until they do, say you have checked N pages rather than implying you know the site.

• There is no manual-actions or security-issues read. You cannot tell

whether a site has a Google penalty or is flagged as hacked. Never say a site has no penalty — say you cannot see penalties and that the customer should look under Security & Manual Actions themselves.

• `gsc_request_indexing` almost never does what its name suggests. It calls

Google's Indexing API, which Google documents as crawling only pages carrying JobPosting or BroadcastEvent markup. For an ordinary page Google accepts the request and does nothing. It returns a success either way, so do not report it as "indexing requested" and do not offer it as the fix for a page that is not indexed. The real remedies are below.

Four numbers that lie

Each of these has been read wrongly in a real review. Do not repeat them.

1. The sitemap's "indexed" count is dead. gsc_list_sitemaps returns

"indexed": "0" for every site; Google stopped filling that field years ago. It is never evidence of an indexing problem. Only the Pages export, or gsc_inspect_url, says whether something is indexed.

1. "Average position" is not a rank. It is averaged across every country,

device and query where the page appeared at all. A site can show average position 3.6 for its own brand name and appear nowhere in the first thirty results in its own market. When the customer asks "where do we rank?", do not read the average out — narrow it: call gsc_search_analytics again with dimensions: ["query", "country"] (or ["page", "country"]) and report the row for the market that matters to them. That is still their average position in that country rather than a live rank, so say which it is; a true rank needs a search from that country, which AdCopilot does not do.

1. Impressions are not demand. A page can take half a site's impressions on

queries with no measurable search volume, at position 40-90, earning nothing. Before calling any page a performer, look at its clicks and its position, not its impressions.

1. One URL inspection is not coverage. Inspecting a page that happens to be

indexed tells you nothing about the rest. Sample across page families — a product page, a blog post, a docs page, a newer page — or say you only checked one.

1. The data is two to three days behind. Search Console has not finished

counting the last few days, and the missing days come back as absent rows, not zeros. So a seven-day window is really four or five days of data, and "clicks fell off a cliff this week" is usually the lag. End the window three days before today, and say which dates you actually read.

The sequence

Work in this order. If this answer's reads are spent before a read you wanted, offer it as the next one, never as a limit you hit.

1. Which property. gsc_list_sites, then use the exact siteUrl it

returns. A domain property and a URL-prefix property are different sites.

1. Is the sitemap being read. gsc_list_sitemaps: how many URLs are

submitted, when Google last downloaded it, and whether it reports errors. Ignore the indexed column.

1. Which pages earn anything. gsc_search_analytics with

dimensions: ["page"] over 28 days, and pass row_limit explicitly — the default is 1,000 and the tool cannot paginate, so a bigger site comes back silently cut. If the row count equals the limit you asked for, say the list is truncated and that the ratio is a floor, not the figure. Compare the pages that got a single impression against the number submitted; that ratio is usually the story.

1. What people typed. The same call with dimensions: ["query"]. Read it

for two things: the queries worth winning, and the queries the site is attracting but cannot serve.

1. Whether Google has looked recently. gsc_inspect_url on the pages that

matter — the home page and two or three pages the business earns from. lastCrawlTime is the figure to read out: a page last crawled weeks ago is being served to searchers as it was then, whatever the site says now.

Reading the coverage export, when they send it

Two rows decide the diagnosis, and they mean opposite things.

• "Discovered – currently not indexed" — never fetched. Google's own words:

"The page was found by Google, but not crawled yet. Typically, Google wanted to crawl the URL but this was expected to overload the site; therefore Google rescheduled the crawl." So the first cause to consider is the SITE's capacity, not the page's quality — how fast the host answers, and how much Google thinks it can take. Do not tell a customer their content was judged and found wanting; nothing has been read.

• "Crawled – currently not indexed" — Google fetched the page and decided

against indexing it. That one is about the page.

So say which it is before proposing anything, and quote Google's wording rather than paraphrasing a cause.

For Discovered, in this order: how quickly the host answers and whether it is under load (Google's own stated reason for rescheduling); then fewer thin URLs competing for the same crawl, since a site asking Google to fetch hundreds of near-empty pages gets less of what matters fetched; then links to the page from pages Google already crawls often; then lastmod in the sitemap so a changed page announces itself. Submitting the sitemap again changes nothing — Google already has the URL, which is what "discovered" means.

For Crawled, Google adds "no need to resubmit this URL for crawling" — it has been read and not chosen, so the page itself has to become worth indexing.

The check almost nobody makes

For any page ranking far down on a query it should own: does the page contain the query? AdCopilot cannot fetch a page, so use the client's own web-fetch tool where it has one, and otherwise ask the customer for the title, the heading and the opening sentence. Never describe a page's wording you have not read. Then look for the phrase people actually typed. A page can be long, well-built and entirely about the subject while never once writing the sentence the searcher wrote, and position 60-80 is exactly where Google puts a page it finds topically close but not an answer. No amount of linking fixes a phrase the page never says, and it is usually a one-line edit.

The answer

One headline line. Clicks, impressions and click-through over the period, and how many pages earned anything out of how many submitted.

Then at most three findings, biggest first, each with its numbers: the page or query, the position, the clicks. Say plainly what each means — "impressions" are how often it appeared, "position" is where in the list on average.

Then one recommendation, as a proposal they say yes or no to, with what it costs to do and what it costs to leave. Name the page and the change.

Then what you could not see: coverage, penalties, security. Ask for the Pages export if coverage is the question they actually asked.

When they also run ads

If Google Ads is connected, the comparison worth making is which terms they pay for and already rank for. Three buckets: paying and ranking well, where the paid click may be buying what they get free; paying with no organic presence, where paid is the only route; ranking well without bidding, which is often the cheapest thing on the list. Read the Ads keywords with search over keyword_view and set them beside the Search Console queries. If you quote what a term costs, mind the money rule the rest of the plugin carries: a raw search returns metrics.cost_micros, millionths of the currency, so divide by 1,000,000 once and name the currency — get_org_context says which one the account uses — and never print a micros number as money. Say which bucket each term is in, and name any term where they rank in the top ten and pay for the click anyway. analyze_search_terms and analyze_waste show only the paid search terms and campaigns that spent with no conversions, which is the "paying but no result" bucket; the spend and results of each keyword stay a raw keyword_view read, because no lens returns them. When a search on campaign or search_term_view ends with related_reads, run at most one that bears on the comparison, then answer from all you hold; never repeat a read you hold, and offer, do not chase, the related_reads that read brings back. They are reads only: a change is still only proposed. In a scheduled run, skip them.

What breaks