Examples

Paste any of these into a connected assistant. Each one lists the tool sequence it should produce, so you can tell a working connection from a plausible-sounding guess, and each names the thing that most often gets read wrong.

Every recipe starts with list_projects, because every Website-scoped tool takes an id it returns. If an assistant skips it and invents an id, that is the tell.

1. Where am I invisible?

"List my websites. For the first one, show me AI visibility over the last 30 days broken down by assistant, and tell me how many checks each rate is based on."

Sequence. list_projects → get_prompt_visibility

Watch for. The denominator. A rate with no check count behind it is not a measurement, and a null rate means unknown rather than 0%. If the assistant reports a null as zero, correct it: it is telling you that you are invisible when nothing has been measured.

2. Who gets named instead of me?

"For my main website, show me the tracked prompts where my brand is not mentioned, and which of my saved competitors appear in those answers instead."

Sequence. list_projects → list_prompts → get_prompt_visibility → list_competitors

Watch for. Whether the competitor set is real. If it is full of directories and marketplaces, the answer is about your category rather than about your rivals, and the fix is to prune the set in the product.

3. Cut my recurring spend

"Show me my active prompts and their mention rates. Recommend five to pause, and tell me what pausing them saves per run. Do not change anything yet."

Sequence. list_projects → list_prompts → get_prompt_visibility → get_credit_balance, then set_prompt_status only after you agree

Watch for. The arithmetic. A prompt is charged per assistant per run, so five prompts on five assistants is twenty-five checks saved per run, not five. This is the single most effective change available on most accounts.

4. Run a site audit properly

"Start a site audit on my main website. Check my credit balance and tell me the cost first. Then poll until it finishes and give me the issues grouped by severity."

Sequence. get_credit_balance → list_projects → start_site_audit → get_site_audit_status (repeatedly) → get_site_audit_issues

Watch for. This is a dispatch-and-poll pair. start_site_audit returns a job id, not results. An assistant that reports "no issues found" straight after starting it has read the dispatch response as the answer.

Also watch for. Checks reported as not assessed. That is a real verdict and it is not a pass: it means no audit path that ran could produce a verdict for that check.

5. Explain a traffic drop

"My Search Console clicks fell around the 14th. Show me the trend, then ask the site timeline what happened around that date, and tell me what the timeline cannot see."

Sequence. list_projects → get_search_console_trend → get_site_timeline

Watch for. Two things. The most recent day or two are provisional and will rise, so they are never a decline. And the timeline is a record of what RankX AI did, not a history of the website: an edit made directly in WordPress, a plugin update or a hosting incident is invisible to it, so an empty result means "no record", never "nothing happened". The tool states its own blind spots on every response; a good answer passes them on.

6. Find striking-distance queries

"Show me the position distribution for my main website, then the queries in positions 4 to 10 with the most impressions and the worst click-through."

Sequence. list_projects → get_search_console_opportunities → get_search_console_performance

Watch for. Average position improves as the number falls. An assistant reporting a "20% increase in average position" has produced a phrase that means nothing.

7. Is an AI Overview costing me clicks?

"For my main website, compare click-through on queries where Google shows an AI Overview against those where it does not. Tell me if the sample is big enough to conclude anything."

Sequence. list_projects → get_search_console_opportunities

Watch for. The tool refuses to conclude below a minimum number of queries on each side, and says so. That refusal is the answer, not a failure, and an assistant that reports a difference the tool declined to draw has invented it.

8. Find pages Google is not indexing

"Which pages on my website are not indexed but still earn impressions? Then inspect the worst one and tell me Google's own reason."

Sequence. list_projects → get_index_coverage → inspect_url

Watch for. The denominator again: coverage counts describe pages that have been inspected, not every page on the site, and the response states its own. inspect_url is the only way to see a page that is not indexed at all, because a non-indexed page has no clicks or impressions to appear in any other tool.

9. Why is there no traffic data?

"There is no Analytics data for my website. Why?"

Sequence. list_projects → get_google_integration_status

Watch for. The answer should name one of four states and its remedy: not connected, needs reconnecting, connected but never synced, or ready. It should never report zero traffic. This is the tool that exists so an assistant can answer precisely instead of guessing, including the case where a connection is healthy but no property has been selected.

10. Research to brief

"Research keywords around 'ai visibility tracking'. Show me the ten best by volume against competition, do not save anything, and then create a content brief for whichever one I pick."

Sequence. get_credit_balance → list_projects → research_keywords → create_content_brief → get_content_brief

Watch for. research_keywords spends only for a fresh discovery; a repeat inside the cache window is free. create_content_brief starts background research, so the brief is not complete when it returns: poll get_content_brief until it is.

11. Generate a draft and read it back

"Generate an article from brief X. Confirm the cost with me first, then poll until it is ready and show me the opening three paragraphs and the GEO score."

Sequence. get_credit_balance → generate_article → get_content_item (repeatedly)

Watch for. Another dispatch-and-poll pair. A null body means the article has not been drafted yet, not that it came back empty. And nothing here publishes: the draft lands in Content Studio for a human.

12. Fix meta descriptions safely

"Find the pages on my connected WordPress site with missing or thin meta descriptions. Dry run a fix for the worst five and show me the exact before and after. Do not apply anything yet."

Sequence. wp_describe_site → wp_list_content → wp_get_content → wp_update_seo with the dry run on

Watch for. Three things. wp_describe_site first, always, because whether SEO fields are writable depends on the site's plugin. The dry run is the default and the assistant should not switch it off without being told to. And if the write is refused, the refusal carries the remedy: read it rather than retrying.

What a bad answer looks like

Worth knowing, because a confident wrong answer is the failure mode here:

• A number with no denominator. Every rate in RankX AI has one.
• A null reported as zero. The tools distinguish them and so should the answer.
• "No issues found" immediately after starting an audit. The dispatch response was read as the result.
• "Zero traffic" for a Google surface. The tools refuse rather than reporting zero, so a zero means the assistant invented one.
• A conclusion the tool declined to draw. The opportunities tool says when a sample is too small.
• A live write with no dry run shown first. Every WordPress write dry-runs by default, so a real change with no diff means the default was overridden.

Where to go next

• Agent Skills for the same jobs as reusable workflows.
• The tool reference for every tool.
• Troubleshooting when a recipe does not run.