Prompt Guidelines for Google Ads AI: 40 Working Prompts

A prompt for a Google Ads AI agent is not a search query — it is a work order to a specialist that reads your account and, on your command, changes it in seconds. The prompts that work share three habits: they name the account, they state the constraint, and they demand evidence before action. Everything below is copy-paste ready and grouped by job.

These prompts assume an agent connected through an MCP connector with real read and write tools — the setup described in what an AI agent for Google Ads is. They work in any MCP client: Claude, ChatGPT, Copilot, Gemini CLI, Cursor, OpenCode. Reads run freely; writes surface an approval in your client first, which is why you can paste these without fear — the approval loop is the safety net under all of them.

The three habits that separate working prompts from wishes

An agent that can act needs sharper instructions than an agent that can only answer. Three habits do most of the work.

Habit one — name the account

"How did we do last week?" forces the agent to guess which account you mean. If your Google sign-in can see one account, the guess is safe. If it can see an MCC with forty clients, it is not. Open with the account:

In account 123-456-7890, ...

Once per conversation is enough — the agent keeps the context. New conversation, name it again.

Habit two — state the constraint

The constraint is the sentence that makes a write safe to request. "Create the campaign paused." "Keep the budget under ₹2,000/day." "Draft the negatives but don't add them yet." An agent with write tools does what you ask — so ask for exactly the blast radius you want. Campaigns created through AdCopilot default to paused anyway, but say it: the constraint in your prompt is the one the agent optimises around.

Habit three — ask for evidence before action

Never let the first tool call be a write. "Find keywords worth pausing, show me the data, and wait" produces a table you can judge. "Pause the bad keywords" produces an approval prompt you have to reverse-engineer. Every action prompt below has the evidence step built in — keep it when you edit them.

Auditing: find the leaks before touching anything

Start here on any account you have just connected. These are pure reads — no approvals, no changes, safe on day one. For the full half-hour sequence, see your first AI Google Ads audit.

In account <account>, give me an account overview for the last 30 days:
campaigns with spend, cost, conversions, cost per conversion, and impression
share where available. Flag anything that looks structurally odd before I
ask follow-ups.

Use first on any new account. Expect a campaign table plus two or three observations — the agent reading the account, not yet judging it.

In account <account>, which search terms spent money in the last 30 days
without a single conversion? Rank by cost, show the matched keyword and
campaign for each, top 25.

The wasted-spend classic. Expect a ranked table; the top rows usually pay for the whole exercise.

In account <account>, list every ad that is disapproved or limited by policy,
with its ad group, campaign and the policy topic Google cites.

Run monthly. Expect a short list or a clean bill — disapprovals silently throttle otherwise healthy ad groups.

In account <account>, which ad groups have no enabled responsive search ad,
or only one? An ad group with a single RSA has no creative test running.

Structure check. Expect a list of thin ad groups — feed it into the campaign-build prompts later.

In account <account>, check for conflicts: are any negative keywords
blocking keywords I am actively bidding on? List the negative, its level,
and the keyword it collides with.

Run after inheriting an account. Expect either "no conflicts" or a precise collision list worth fixing the same day.

In account <account>, which campaigns were limited by budget in the last 30
days, and on which days? Show budget, spend, and lost impression share due
to budget.

Budget pressure check. Expect day-level detail — the input for the budget prompts below, not a reason to raise anything yet.

In account <account>, summarise the change history for the last 30 days:
what changed, in which campaign, and when. Group by type of change.

Run when performance shifted and nobody knows why. Expect a grouped timeline — the account's recent memory.

In account <account>, review geo targeting across campaigns: where are ads
eligible to show, and did any location spend without converting in the last
30 days? Table, worst first.

Expect a location performance table. Location creep is the quietest form of wasted spend.

Negative keywords: the highest-value low-risk writes

Negatives cannot increase spend, which makes them the right first write. Concepts and match-type detail live in negative keyword automation; the weekly cadence lives in the negative keyword sweep.

In account <account>, pull search terms from the last 14 days with at least
one click and zero conversions. Classify them: irrelevant, competitor,
research-intent, or plausibly fine. Show cost per term. Recommend negatives
— but don't add anything yet.

The weekly sweep opener. Expect a classified table and a shortlist; the "don't add yet" keeps the write for the next prompt.

From that list, add the irrelevant terms as exact match negative keywords to
the campaigns where they appeared. Show me the exact list one more time
before the call.

Follow-up to the sweep. Expect one approval per campaign showing every term — read it, then approve.

In account <account>, do searches containing "free", "jobs", "careers",
"salary" or "diy" spend money here? If yes, propose phrase match negatives
per campaign and wait for my approval.

Pattern blocking. Expect either "no spend on those patterns" or a per-campaign proposal with the spend evidence attached.

In account <account>, how much did searches containing competitor names cost
in the last 30 days, and did any convert? Give me the case for and against
blocking them — numbers, then your recommendation.

Competitor terms are a judgement call. Expect an argument with data on both sides, not an action.

In account <account>, are brand searches leaking into my non-brand
campaigns? Check search terms in generic campaigns for my brand name and
propose negatives to route brand traffic to the brand campaign.

Structure hygiene. Expect a leakage table and a campaign-level negative proposal that keeps CPCs honest in both campaigns.

In account <account>, review the existing negative keywords in <campaign>.
Is anything plausibly blocking traffic I want — too-broad phrase negatives,
or negatives added for a problem that no longer exists?

The reverse audit. Expect a shortlist of negatives to reconsider — over-blocking is invisible in every other report.

Campaign builds: from brief to reviewable draft

Builds are multi-step writes, and the paused default is what makes them calm: nothing spends until you enable it. The full walkthrough is launch a campaign with AI; Performance Max has its own build.

In account <account>, create a paused search campaign called
"<name> - Search - <month>" with a daily budget of ₹1,500 and manual CPC.
Don't enable it. Confirm what you created with its ID.

The scaffold. Expect one approval for create_campaign, then a confirmation with the new campaign ID.

In that campaign, create an ad group "<theme>" with a default CPC bid of
₹40, then add these keywords in phrase match: <list>. Keep everything
paused-safe: the campaign stays paused.

Expect two approvals — ad group, then keywords. The campaign-level pause keeps the whole tree from spending.

Draft a responsive search ad for ad group "<theme>" landing on <URL>.
12 headlines under 30 characters, 4 descriptions under 90 characters,
drawing on this value proposition: <two sentences>. Show me the copy for
edits BEFORE creating anything.

Copy first, tool call second. Expect the full headline/description set in the chat, then — after your edits — one create_responsive_search_ad approval.

In account <account>, add sitelink assets for <page 1>, <page 2>,
<page 3>, <page 4> and link them to campaign <campaign>. Draft the sitelink
text and descriptions for my review first.

Expect drafted sitelinks, then asset-creation and linking approvals. Repeat the shape for callouts and structured snippets.

Set geo targeting on campaign <campaign> to <cities/regions>. Show me the
final list of locations before the call.

Expect a target list to sign off, then one set_geo_targets approval. AdCopilot adds location targets; location exclusions stay a Google Ads UI job.

Set an ad schedule on campaign <campaign>: Monday to Friday 9:00 to 20:00,
Saturday 10:00 to 16:00, off Sunday. Confirm the schedule back to me in
plain words before applying.

Expect the schedule restated, then one approval. Schedules silently default to "always on" when nobody sets them.

Here is a campaign brief: <paste brief>. Build the whole thing paused in
account <account>: campaign, ad groups by theme, phrase match keywords, one
RSA per ad group, sitelinks. Walk me through it step by step, asking
approval at each write, and give me a final summary of everything created.

The full build. Expect a sequence of approvals in dependency order and a closing inventory — then review in the Google Ads UI before enabling.

Budgets and bids: spend changes, evidence attached

These are the prompts where the approval loop earns its keep. Every one asks for the case before the change. What AdCopilot can move: campaign budgets, campaign status, ad group default CPC bids, keyword CPC bids, device bid adjustments. It cannot switch an existing campaign's bidding strategy — that stays a deliberate decision in the Google Ads UI.

In account <account>, review budget pacing: for each enabled campaign, show
daily budget, average daily spend over the last 14 days, and whether it is
losing impression share to budget. No changes — just the table.

The pre-decision read. Expect a pacing table that makes the next two prompts obvious.

Raise the daily budget of campaign <campaign> from ₹1,500 to ₹2,000. Before
the call, show me its last-14-day cost per conversion so I can sanity-check
that I'm scaling something that works.

A single change with its evidence. Expect the metric, then one update_campaign approval showing the new budget.

In account <account>, mobile feels weak. Show conversion rate and cost per
conversion by device for the last 30 days, and if mobile genuinely
underperforms desktop by a wide margin, propose a device bid adjustment.
Wait for my approval.

Expect a device table and either "leave it" or a specific percentage proposal via set_device_bid_adjustment.

In campaign <campaign> (manual CPC), which keywords have strong conversion
rates but low impression share? Propose keyword-level bid increases capped
at 20%, show current and proposed bids, and wait.

Expect a keyword table with current → proposed bids; approve the update_keyword calls per keyword or decline individually.

In account <account>, which enabled keywords spent more than ₹2,000 in the
last 30 days with zero conversions? Show me the list with spend, then pause
the ones I confirm.

Expect the list first, your confirmation, then paired update_keyword status approvals. Pausing is reversible; that is why it is safe to delegate.

Campaign <A> is budget-capped and converting; campaign <B> spends without
converting. Propose a reallocation: how much off B, onto A, and why. Two
separate budget changes, each with its own approval.

Expect a reasoned split and two update_campaign approvals — decline either half independently.

Reporting: answers you can forward

Reads only — no approvals anywhere in this group. The agent queries the account live and formats what it finds.

In account <account>, write my Monday summary: last 7 days vs the 7 before —
spend, conversions, cost per conversion, by campaign. Three lines of
narrative on what moved and the single most likely reason.

Expect a compact table plus genuinely short narrative. The anchor of the weekly routine.

In account <account>, compare this month to last month at account level:
spend, clicks, CTR, conversions, cost per conversion. Note anything that
moved more than a fifth in either direction and say why, checking change
history if needed.

Expect a two-column comparison with called-out movers — and a change-history cross-reference when something jumped.

Deep-dive campaign <campaign> for the last 30 days: performance by ad
group, by device, by day of week. Where is the money going, and which cell
of that breakdown is weakest?

Expect three small tables and a verdict identifying the weakest slice.

In account <account>, keywords ranked two ways for the last 30 days: top 10
by conversions, top 10 by cost with zero conversions. One table each.

The winners/losers pair. Expect two clean tables — feed the losers into the pause prompt above.

In account <account>, how much of total spend went to Performance Max in
the last 30 days, and what can you actually see about where it went —
asset group performance, listing groups, anything queryable?

Expect an honest partial answer. PMax hides channel detail from everyone; what is visible is worth knowing precisely.

Turn the last 30 days of account <account> into a client-ready summary:
five bullet points, plain language, no jargon, numbers rounded. Then a
"what we'll do next week" line based on what you found.

Expect prose you could paste into an email — check the "next week" line matches what you actually intend.

Housekeeping: the tidy-account prompts

Small jobs that keep an account legible. Mostly reads with the occasional low-stakes write.

In account <account>, audit naming: list campaigns and ad groups whose
names don't follow "<pattern>". Propose renames, and apply the ones I
approve.

Expect a nonconforming list with proposals — rename approvals are the lowest-stakes writes there are.

In account <account>, find structural dead weight: ad groups with no
enabled keywords, no enabled ads, or no traffic in 90 days. List them with
status and last activity.

Expect a dead-weight list. Pause what you confirm — AdCopilot cannot delete anything, so cleanup means pausing, and that is reversible by design.

In account <account>, which campaigns are missing sitelinks, callouts or
structured snippets? Show the asset coverage gap per campaign.

Expect a coverage matrix — feed the gaps into the sitelink build prompt above.

In account <account>, list the distinct final URLs my enabled ads point to,
by campaign. Flag anything pointing at a page that looks off-pattern —
wrong domain, http, tracking parameters gone wrong.

Expect a URL inventory with flags. Follow up outside the agent by actually loading the flagged pages.

In account <account>, do any campaigns have an ad schedule or geo exclusion
that looks like a leftover — a schedule from a past promo, an exclusion
that contradicts the targeting? List anything suspicious with its setting.

Expect a suspicion list, each row naming the campaign and the setting to reconsider.

In account <account>, review device bid adjustments across campaigns. Which
adjustments exist, and does the last 30 days of device data still support
each one?

Expect a table of adjustment vs current evidence — stale adjustments outlive the data that justified them.

Summarise everything you changed in account <account> this conversation:
every write, in order, with the values. I'm pasting this into my log.

End every working session with this. Expect a clean changelog — it should match the approvals you granted and the connector's own audit trail.

Make the library yours

Copy the ten prompts you will actually reuse into a note, replace <account> with your customer ID, and adjust the currency amounts to your budgets — that is the five-minute version of this page. Then run the auditing group top to bottom on your own account; it costs nothing and usually finds the first month of work.

If you are not connected yet: Start a free pilot — a Google sign-in gets you a connector address to paste into whichever client you use. Seven days, full toolset, up to five accounts, no card. Then start at the top of this page and work down.