Agency Autopilot

Create a GA4 property FOR one of this agency's clients — with its web data stream (which yields the G- measurement ID for the site tag) — and link it to that client. Day-one set-up for a new client. Previews unless confirm is true. Follow with ga4_create_key_event, wp_install_tracking (ga4_id = measurement_id) and ga4_link_google_ads.

ga4_create_property

Link a client's Google Ads account to their GA4 property so key events and audiences flow into Ads. Both must be linked to the same client here. Previews unless confirm is true. Never unlinks.

ga4_link_google_ads

Google Ads accounts linked to a GA4 property. A missing link is why Ads shows no Analytics conversions or audiences.

ga4_list_google_ads_links

Key events (conversions) defined on a GA4 property: event name, counting method, and whether it is custom. Use before creating one, and when a report shows zero conversions.

ga4_list_key_events

Run a GA4 report. Dimensions and metrics use GA4 API names (e.g. dimension 'sessionDefaultChannelGroup', metric 'sessions'). Dates accept YYYY-MM-DD or relative forms like '28daysAgo' and 'today'.

ga4_report

Apply the Google Ads fixes from audit_client in one step: pause every keyword the audit flags as spending with no conversions, and add the search terms the user chose as negative keywords (phrase match). Only items the audit flags right now are accepted. Previews unless confirm is true; each keyword can be re-enabled with ads_set_keyword_status. Does not change budgets, extensions or Meta campaigns.

apply_audit_fixes

Audit one client and return a scorecard: conversion tracking, keyword spend with no conversions, search terms, impression share and ad extensions (Google Ads, last 30 days); leads recorded, campaigns spending without leads, ad fatigue and click-through (Meta ads, last 30 days); rankings within reach and low click-through titles (Search Console, last 28 days); whether the website's leads are being measured (Analytics); and the website itself: on-page SEO, broken links, text cut off mid-sentence, and from the weekly site health, speed with its causes on each money page, whether Google has indexed the pages that matter, and the Business Profile. Each area is ok, warning or critical, with the top actions and the tool that makes each fix. Use for a new client, or when asked what to fix first. Reads only.

audit_client

Audit a business the agency does NOT manage yet, from public data: Google Business Profile completeness, reviews against nearby competitors, the website (speed, titles, broken links), tracking tags, and where it appears in Google Maps-style searches for its core terms. The pitch tool: a scorecard with the three fixes most worth making, in a few minutes, with nothing connected. A section it could not check is marked unverified and claims nothing. Free plan: 5 audits a month; paid plans unlimited. Nothing is sent to the prospect. A website alone is enough; name and city are optional.

audit_prospect

Decline a change Autopilot proposed (the ids are in autopilot_status), when the person says no: Autopilot will not make it. Pass why if the person said; it is how Autopilot learns what this agency wants.

autopilot_decline

Dismiss one open Autopilot alert (the dashboard's Dismiss): it stops showing on the dashboard and in status, and stays on the client's Autopilot page until the data shows it is fixed. Alert ids come from autopilot_status. Call without confirm to preview.

autopilot_dismiss_alert

End the Autopilot run that autopilot_start began in this conversation: stores the short report on the run page (with any proposals the run recorded) and unlocks nothing else: further changes need a new autopilot_start. Emails the agency only when it is needed, a client report waits for review or went out automatically, or a client stopped its reports or one bounced, unless notify is false (a routine that sends its own summary).

autopilot_finish

The log of every change Autopilot made for a client: when, what, why (its evidence), the result it expected, and the measured result once there has been time (leads before against after). Includes proposals the agency approved. Use for "what has Autopilot done for this client", "what has Autopilot changed", "why did it do that", "did it work". For what it has learned, use autopilot_status. Read-only.

autopilot_log

One change Autopilot proposed and is waiting for the agency's yes (the ids are in autopilot_status): what it is, why, and the exact tool call that makes it. When the person says yes, make that call with that tool yourself (for a review reply you may use the person's edited text as comment), then record it with autopilot_proposal_done. When they say no, use autopilot_decline. Changes nothing.

autopilot_proposal_get

Preview Autopilot's Google Ads clean-up for one client: reads the account and returns the keywords it would pause and the search terms that spent with no conversions, plus the proposals that need a person. Changes nothing. Takes up to a couple of minutes. Premium plan.

autopilot_preview

Records that the agency approved a proposal and the change was made: call it right after the tool call from autopilot_proposal_get succeeded. It checks that call went through (the same tool, the same ids, in the last 2 hours), so a proposal is never marked done without the change. Makes no change itself.

autopilot_proposal_done

Record how this agency wants its Autopilot to work, in the person's words ("never touch the brand campaign", "write posts in German", "always ask before publishing pages", "don't do that again"), for every client (leave client out) or one. Preferences come first in every Autopilot run and win over data from other clients. Call it whenever the person says how Autopilot should or should not behave. Adds only; autopilot_preference_remove drops one.

autopilot_preference

Remove one Autopilot preference by its key (from autopilot_preferences), so Autopilot stops following it. Only when the person asks to drop that preference; say which one it was.

autopilot_preference_remove

Run Autopilot on one client in THIS conversation, on the agency's own assistant plan (no separate AI cost): returns Autopilot's briefing (leads, forms, demand, Ads health, notes, last run) and its instructions, and locks this conversation to that client and the changes the agency allowed with autopilot_set, checked on every call. Anything outside the limits is recorded as a proposal instead of running. In shadow (or with Autopilot off) nothing is changed. Work the client as the instructions say, then call autopilot_finish. Use for "run Autopilot on X" and in scheduled routines. Premium plan.

autopilot_start

Saves the set-up answers (the agency's, and a client's with client_answers) and applies them: the change types Autopilot may make itself, the client-report choice and standing rules for its runs. A client that is on moves to shadow if the answers call for it; saving never switches a client on (off stays off, shadow stays shadow): autopilot_set does that.

autopilot_setup_save

Switch Autopilot for one client: "shadow" investigates and reports what it would do, changing nothing; "on" makes the kinds of change allowed below itself (small, reversible, capped, never brand terms); "off" stops it. Shadow or on also starts a daily lead watch (needs Analytics, Search Console or Google Ads). The work runs in the agency's own Claude: "run Autopilot on X" (autopilot_start) or a weekly scheduled task (autopilot_routine gives its prompt). Budgets, bidding, new campaigns, taking pages down and messaging always come back as proposals for a person. Suggest shadow first. Premium plan.

autopilot_set

Tell Autopilot a fact about a client that its data cannot show, so it stops asking: e.g. "Test lead sent 10/10 from /contact/, reached HighLevel and the thank-you page", "Client is away until 11/1, keep spend flat", "Don't touch the ADU pages". Notes are dated and read at the start of every Autopilot run. Changes nothing in any account.

autopilot_note

Everything Autopilot follows for this agency: preferences the person stated and ones learned from what they approved, declined and undid, each with its key and which clients it applies to. Reads only.

autopilot_preferences

Autopilot's set-up questions, so it works the way this agency does: required before a client is first switched to shadow or on. The agency's part is asked once (what they do for most clients, how hands-on, how they want to hear from it, the client report); each client's part is short (what the agency does for it, what a win is, what's off-limits, low-star review replies, anything else). Each question has its options and default, with any answers already given. Ask them in the chat in one message, or give the person setup_page (a web page of answers to click). Call it as soon as set-up is needed, without asking first. Changes nothing.

autopilot_setup_questions

The weekly Autopilot prompt, and how to run it every week in the assistant this conversation is in (a scheduled task where the assistant supports one with connectors, otherwise a weekly ask): for every client in shadow or on, or for one client. Runs use the agency's own assistant plan (no separate AI cost). Changes nothing.

autopilot_routine

Autopilot's state. With no client: everything Autopilot needs from the agency across all clients (needs_you, each with what to do), then every client's mode and last run; use for "what does Autopilot need from me". For one client: its mode, what it has learned about the client (learned: the lessons its runs wrote down, which every run follows), what has worked across all Autopilot clients (across_autopilot), what needs a person, and its last four runs. Use for "what has Autopilot learned", "what is it doing", "does it need anything". For every change with its reason and measured result, use autopilot_log. Read-only.

autopilot_status

Record why the person declined an Autopilot proposal or turned Autopilot off for a client, in their words. Call it after autopilot_decline or autopilot_set returns ask_why and the person answers. It is how Autopilot gets better: the reasons are read every week.

autopilot_feedback

Publish a post (update) on a Business Profile location — text, optional call-to-action button and photo. Preview unless confirm is true; posts are public the moment they are created.

gbp_create_post

List the Google Business Profile locations this sign-in manages, with the gbp_location_id every other gbp_* tool needs. Call this first; never guess a location id.

gbp_list_locations

Business Profile performance for one location over a date range: impressions (Maps and Search, mobile and desktop), calls, website clicks, direction requests, conversations and bookings, with daily values. Use for client reporting beside Ads and Analytics.

gbp_performance

Recent posts (updates) on a Business Profile location, with their state and links.

gbp_posts

Post the owner's public reply to a review. Runs as a preview unless confirm is true; the reply text is published under the business name, so show it to the person first.

gbp_reply_review

Recent reviews for one Business Profile location: rating, text, reviewer, and whether the owner has replied. Use it to find reviews waiting on a reply.

gbp_reviews

Undo review_leave_unanswered: the review counts as waiting for a reply again (and Autopilot may answer it). Previews unless confirm is true.

review_leave_unanswered_clear

Mark a review the agency (or the owner) leaves unanswered on purpose, e.g. a low-star review the owner handles himself. From then on it is not counted as waiting for a reply, Autopilot won't reply to it, the reply rate leaves it out, and reports never call it unanswered. Previews unless confirm is true. review_leave_unanswered_clear undoes it.

review_leave_unanswered

Add a note to a HighLevel contact — useful for recording what was agreed or what you changed. Additive and visible to the agency's team.

highlevel_add_note

Appointments booked in a date range, with how many were canceled or no-showed. For a local service business an appointment on the calendar is usually the real booked job — use this, not just pipeline stages, when asked what the marketing actually produced. Each appointment carries `location` when one is set, and `meeting_link` when that location is a joinable video link.

highlevel_appointments

List the calendars in a HighLevel sub-account, so appointments can be filtered to one of them.

highlevel_calendars

List recent conversations in a HighLevel sub-account, newest first, with how many are unread. Use this to answer 'who is waiting on us' — an unanswered inquiry is the most expensive thing in a local business's pipeline.

highlevel_conversations

Create an opportunity (a potential job) in a pipeline, attached to an existing contact. Get the pipeline and stage IDs from highlevel_pipelines and the contact ID from highlevel_find_contact or highlevel_create_prospect. Call without confirm first to preview.

highlevel_create_opportunity

Create a new prospect (contact) in a HighLevel sub-account. Needs an email or a phone number. Call without confirm first — the preview shows whether someone matching already exists, so you do not create a duplicate. If a match comes back, use that contact instead unless the human tells you it is genuinely a different person.

highlevel_create_prospect