Once GoodLads is connected, your assistant can read your Google Ads accounts and reason about them with our playbooks behind it - the same ones the product holds its own recommendations to. It still cannot change anything. This page is the five things worth asking it, and the exact words to paste.
Not connected yet? Connect it here - one URL, one approval, about a minute:
https://goodlads.cc/api/mcp
Start every conversation the same way
An MCP session begins outside any account, so the first thing your assistant needs is which one you mean. Name it, and name the window you care about. If your client does not show GoodLads' ready-made prompts, open with this and it will behave the same:
Use the GoodLads tools. Call list_accounts, then list_playbooks, before you analyse anything.
That second call is the one that matters. list_playbooks and get_playbook hand your assistant GoodLads' method for the lever in question - the evidence to pull, the bar an idea has to clear, and the honest mechanics of the change. Without it you get plausible advice about a live budget, which is the expensive kind of wrong.
The five to start with
1 of 5
It is Monday. What should I change this week?
Run a GoodLads review of my Google Ads account and give me the three changes worth making this week, ranked by the money at stake. Cite the figure behind each one, and skip anything that is already on the board or already applied.
In a client that shows MCP prompts, pick GoodLads account review instead of pasting this.
What it does behind the scenes
list_playbooks, to see which levers exist before deciding which apply
list_campaigns, to put the attention where the spend is
list_board_ideas and list_implemented, so nothing you already have comes back as news
get_playbook for each candidate lever, then get_campaign_context for the campaigns carrying real spend
What you getA short ranked list. Each item names the figure that triggered it, the one thing it changes, the test that would settle it, and a link to the card in your dashboard.
Watch forAsk it to say which levers found nothing. An honest empty answer on six of nine levers is the normal result, and it tells you where the account is already clean.
2 of 5
Where is this account leaking money?
Find the search terms taking real money and returning nothing across the whole account. Name each term with its own spend and its own return, total the monthly waste, and tell me which ones you would exclude and where.
In a client that shows MCP prompts, pick Where is this account leaking money instead of pasting this.
What it does behind the scenes
get_playbook("negative_keywords"), which is account-scope for a reason: the same wasted query usually fires in several campaigns at once
list_search_terms with no campaignId, the only view that shows one term competing in two ad groups
list_implemented, to skip terms already excluded
What you getNamed terms with their own numbers and a monthly waste total, not a campaign average dressed up as a finding.
Watch forA good answer says plainly that a negative blocks a term rather than rerouting its traffic somewhere better, and never proposes negating a brand term. If yours does either, it skipped the playbook.
3 of 5
We are about to make a change. Is it right?
We are about to raise the daily budget on [campaign name] by 40%. Pressure-test that against the account's own numbers before we do it: is it supported, is it material, has it been tried here, and what would settle it?
In a client that shows MCP prompts, pick Pressure-test an idea before spending on it instead of pasting this.
What it does behind the scenes
get_playbook for the lever the idea belongs to, so it is judged on the same bar our own ideas are
get_campaign_context, which carries the budget-and-bidding diagnostic over 30, 60 and 90 days
list_implemented, because the fastest way to know is that it was tried in March
What you getA verdict - do it, test it first, or drop it - with the test that would settle it, sized to the campaign's real daily volume, and what would change its mind.
Watch forThis is the one to run on your agency's recommendations too. The point is to break the idea, not to endorse it, so if the answer only agrees with you, say "argue the other side".
4 of 5
Did what we shipped last month actually work?
What has GoodLads applied to this account in the last 90 days, and did any of it actually work? Give me the measured outcome for each one, and be blunt about the ones that did nothing.
What it does behind the scenes
list_implemented, which carries each applied change with its measured outcome
list_board_ideas, for what is still waiting on a decision
get_campaign_detail on the campaigns involved, when a number needs checking
What you getA verdict per change rather than a list of activity, and a short read on which lever is earning its place in this account.
Watch forA change applied last week has no verdict yet. An assistant that grades one anyway is guessing.
5 of 5
I run this account for a client and the update is due.
Write this month's update for [client]: what the spend bought, what we changed and what it did, and what we are testing next. Pull the last 30 days and the 30 days before that so the comparison is real. Numbers from the account only - no adjective you cannot back with one.
What it does behind the scenes
list_campaigns twice, once at days 30 and once at days 60, so the assistant can do the subtraction itself
list_implemented for what changed and how it measured
list_board_ideas for what is queued next
What you getA draft you edit, instead of an hour in a spreadsheet assembling one.
Watch forNothing in GoodLads computes a period-over-period delta for you yet, which is why the prompt asks for two windows explicitly. Without that, an assistant will happily describe a trend it never measured.
In Claude
Claude connects over OAuth, so the tools are there in every chat once you have approved access. Where your Claude client shows MCP prompts, GoodLads adds 12 of them as ready-made commands - you pick one instead of writing a brief:
GoodLads account review - The full pass GoodLads makes over an account: every lever, in order, with the evidence each one needs and nothing padded.
Where is this account leaking money - The waste pass: search terms taking real money and returning nothing, and terms competing against themselves in two ad groups.
Pressure-test an idea before spending on it - Take an idea - yours, an agency's, or one on the GoodLads board - and try to break it against the account's own numbers.
One per lever - 9 more, to run a single playbook against a single account.
Where it does not, nothing is lost: paste the prompts above. They are the same instructions the commands carry.
In Claude Code and Cursor the same prompts appear as slash commands. Useful if you already keep a terminal open, less so if you do not - none of this needs a code editor.
In ChatGPT
ChatGPT reaches the same tools through a developer-mode connector, but it does not surface MCP prompts, so there is nothing to pick. Open with the line above, name the account, and it will call list_playbooks on its own. If an answer comes back light on figures, the usual cause is that it skipped get_playbook: ask it which playbook it used, and it will go and read one.
What it can read
Every account you have connected to GoodLads, and for each one: campaigns with spend and impression-share diagnostics, one campaign's full briefing (budget diagnostic, ad groups, the ad copy running today, geography, search terms, keyword quality), account-wide search terms, keywords with Quality Score, your deep research report, your idea board, and everything GoodLads has already applied with its measured outcome. Anything not covered by those, it can pull straight from the Google Ads report registry.
And the method for each of the 9 levers GoodLads knows: limited by budget, ad-group bidding targets, app campaign targets, geography, keyword-ad combinations, messaging angle tests, ad-group splits, new campaigns, negative keywords. See Skills for what each one does.
Getting good answers out of it
Name the account and the window. "Last 30 days on Acme" beats "recently", and the tools take a lookback of up to 365 days.
Ask for the figure, not the adjective. "Underperforming" is an opinion; "ROAS 1.2x against an account 3.4x" is a finding you can act on.
Say that nothing is an acceptable answer. The playbooks already say it, but repeating it stops an assistant padding three ideas out of one.
Ask which playbook it used. The fastest way to tell a grounded answer from an improvised one.
Push back. These are hypotheses, not verdicts. "Argue the other side" is a fair instruction and it has the data to do it.
What it will not do
Change anything. The server is read-only by construction: there is no tool on it that can touch a campaign, an ad, a budget or a keyword, whatever your assistant is asked to do. When it talks you into a change it hands you a link to your GoodLads dashboard, where you apply it yourself with the usual one-click approval and a recorded way back. That is deliberate - bring-your-own-AI should never mean an AI we did not write editing a live campaign.
It also costs you nothing extra. Your assistant does the reasoning and pays for its own tokens, so there is no usage limit on the endpoint and no second bill from us.