Ask a different question every Monday and you get a different answer dressed up as insight. I use these 12 prompts in the same order in Claude with groas connected over MCP, so the weekly review stays comparable instead of becoming a fresh conversation.

Set the rules once, then repeat them

When to use: Before the first review. Connect groas under Customize → Connectors with https://mcp.groas.ai/mcp; I keep the workflow in a Claude Project with brand rules and targets so Monday does not start with ten minutes of context-setting.

Run this weekly review for [BUSINESS] using groas data.

Use the last 7 complete days versus the prior 7 complete days throughout. State the date ranges. Keep the same account scope, conversion definition and reporting window across answers.

Use only data you can retrieve. If a source, metric or historical comparison is unavailable, say so. Do not invent a value or treat missing data as zero.

Answer prompts 1-12 in order. Do not apply changes during the review. Keep recommendations separate from completed actions.

The adjustment that matters: Replace the [UPPERCASE PLACEHOLDERS], not the workflow. Keep four identical weeks before rearranging it. The familiar weekly checks cover budgets, conversions, search terms, spikes and placements; the value here is asking consistently, not discovering a more impressive Monday question.

Clay marketer facing twelve identical Monday checklists pinned in a row

1. Spend and pacing

When to use: First, before interpreting efficiency. Establish what actually spent before explaining what it bought.

For [ACCOUNT], compare the last 7 complete days with the prior 7 complete days: spend, conversions, cost per acquisition and revenue.

Show one row per campaign and an account total, with the same metrics in the same order. Compare spend with [WEEKLY SPEND TARGET] and flag any day more than 20% above or below [DAILY SPEND TARGET].

Show the CPA formula. No recommendations yet.

The adjustment that matters: Make the daily and weekly targets explicit. A weekly budget alone does not tell Claude which uneven days were intentional. Ask for a table, not a paragraph about “healthy pacing.”

2. Conversions and efficiency

When to use: After pacing. A cheaper CPA on half the volume is not automatically a win.

For [ACCOUNT], show conversions and cost per acquisition by campaign for the last 7 complete days versus the prior 7 complete days, using [PRIMARY CONVERSION] only.

Rank campaigns by change in cost per acquisition, worst first. Show conversion volume beside each CPA and identify campaigns where CPA cannot be calculated.

The adjustment that matters: Name the primary conversion exactly as it appears in the account. If micro-conversions creep into the answer, correct the definition before comparing campaigns. Otherwise, the tidy ranking is sorting different things.

3. Search terms, sorted by cost

When to use: Every week. This is the work I did by hand at 1am for years, and it is where waste hides in plain sight.

For [ACCOUNT], list the top 20 search terms by spend for the last 7 complete days.

For each term, show the campaign, ad group, associated keyword and match type where available, spend and conversions. Sort by spend descending.

Mark terms above [WASTE THRESHOLD] with zero conversions as negative candidates, not approved exclusions. State the proposed negative match type and account level separately.

The adjustment that matters: If a term looks relevant but has no conversions, request the prior 28 days before judging it. Keep negative mining split by match type and level; a candidate is not an instruction to exclude.

4. What changed last week

When to use: Before proposing another fix. Last week’s approved change should not get quietly undone this week.

For [ACCOUNT], list every available change-history entry from the last 7 complete days.

Show the date, who or what made the change, the campaign affected, the old value and the new value. Compare these entries with [PRIOR APPROVAL RECORD].

Flag budget, bid or targeting changes that do not match that record. Distinguish unapproved changes from changes whose approval you cannot verify.

The adjustment that matters: Supply the approval record. An empty log does not prove the cause sits outside the account. If performance moved without a recorded change, use prompt 6 to investigate the site rather than declaring the mystery solved.

Search and site: find where the movement happened

Paper-cut website map with a path dropping off between landing page, pricing and checkout

5. Organic search movement

When to use: When traffic moved and nobody can explain whether rankings or demand contributed.

For [WEBSITE], retrieve available Search Console data for the last 7 complete days versus the prior 7 complete days: clicks, impressions and average position by query and by page.

List the five queries with the largest drop in clicks and the five pages with the largest drop. Show both periods' clicks, impressions and average position beside each entry.

The adjustment that matters: If position is unavailable, request the query in Semrush or Ahrefs through groas’s SEO data, with its source and reporting window labeled separately. Do not silently substitute one measurement for another. A clicks-only answer leaves too much of the explanation missing.

6. Landing pages and site changes

When to use: When paid and organic performance move together. Check the shared destination before blaming either channel.

For [WEBSITE], take the top five landing pages by sessions in the last 7 complete days. Show sessions, conversion rate and conversions versus the prior 7 complete days.

Flag any page whose conversion rate changed by more than 20% relative to the prior period. List recorded changes to its content, offer, price, checkout or form in the same window, with dates.

Label unavailable change history explicitly.

The adjustment that matters: Investigate checkout and lead-form changes before touching a bid. I used to blame the ad when a client-side landing-page swap had tanked the page the week before. An unlogged change is a lead to investigate, not proof of causation.

7. Competitor moves

When to use: When CPA rises and the first instinct is to blame the auction. Check whether a rival actually changed its offer.

For [COMPETITOR] on Google and Meta, list ads first seen in the last 7 complete days.

For each ad, show its first-seen date, hook, offer and landing page. Where a prior-week landing-page version is available, identify changes to price, shipping promise or lead form.

Keep observed changes separate from anything you cannot compare.

The adjustment that matters: Insist on the first-seen filter. An evergreen ad list is not a weekly change report. If there is no previous page version, keep the current offer in view without inventing a before-and-after story.

AI visibility: keep the questions and denominator fixed

8. Are we cited on the same questions?

When to use: With a frozen question set. Change the questions and the comparison changes with them.

For the frozen buyer questions [BUYER QUESTIONS] across [ENGINES], compare this week with last week within each matching question-and-engine cell.

For both weeks, show valid answers, answers citing [BRAND], answers citing [COMPETITOR] and each brand's citation rate. Calculate each rate as answers citing that brand divided by valid answers.

Report missing or invalid answers separately. Do not call single-answer flips a trend.

The adjustment that matters: Require the counts beside the rate: 23 answers citing the brand out of 138 valid answers is 16.7%. If valid-answer counts differ between weeks, examine the comparable cells before interpreting movement. That is the denominator discipline behind a repeatable weekly visibility check, not a license to label every fluctuation progress.

9. Who gets cited instead of us?

When to use: Immediately after prompt 8, while the same answers are open. The rate shows the gap; the cited pages give something concrete to inspect.

Using the same [BUYER QUESTIONS], [ENGINES] and answer set from prompt 8, list the sources cited for [BRAND] and [COMPETITOR].

Show answer counts per source. Flag new sources, lost sources and sources cited in 3 or more answers.

Read the 2 most cited competitor source pages in full where accessible. State the claim or information each contains that our pages do not. Say when a page cannot be read.

The adjustment that matters: Ask for counts per source, not just domains. A source cited once is an anecdote. A source cited in 10 answers is a page worth reading. Inspect the page before prescribing the content.

10. What customers actually said

When to use: When AI answers contain language the site does not use. I spent years guessing at that language from search-term reports; Reddit threads let buyers write it themselves.

For [PROBLEM], retrieve available customer posts from the last 7 complete days on [SUBREDDITS] and reviews on [REVIEW SITES]. Identify any source you cannot access.

Quote up to 5 of the most repeated phrases customers use for the problem and up to 5 repeated complaints about [COMPETITOR]. Show repetition counts and source links.

Mark phrases absent from our [KEY PAGE]. Do not pad the list if fewer repeated phrases appear.

The adjustment that matters: Require verbatim quotes with links. If Claude paraphrases, ask again. Paraphrased customer language is agency language wearing a costume. Keep the wording that customers actually chose.

Decisions: check the answers before authorizing changes

When to use: After prompt 10, before turning observations into instructions. I use this copyable check whenever a number is about to justify an action.

Before proposing changes, check the review:

- Formula: show account ROAS as total revenue divided by total spend, and CPA as spend divided by the stated primary conversions.
- Repeatability: recalculate account ROAS from those totals. Explain any mismatch with the earlier answer.
- Denominator: show valid-answer counts for both AI-visibility periods.
- Timing: verify competitor first-seen dates and the review's reporting windows.
- History: reconcile account changes with the prior approval record.

For any failed check, restate the affected answer with its source, window and formula. Leave unresolved numbers out of the recommendations.

The adjustment that matters: Check the calculation, not the confidence. One verification write-up shows ROAS moving from 2.6x to 5.0x on identical data when the calculation changes. A polished answer still needs arithmetic.

11. Draft the changes; do not apply them

When to use: Once the checks pass. The review should end in a decision, including a justified decision to leave things alone.

From prompts 1-10 and the completed checks, propose no more than 5 changes for [ACCOUNT].

For each, state the campaign or page, exact old value, exact new value, this week's supporting data point, expected effect on [TARGET METRIC] and the rollback step.

Separate the observed evidence from the expected effect. Do not propose a rebuild or bid-strategy switch. If no change is justified, say so.

Do not apply anything yet.

The adjustment that matters: Reject a twelve-change list. Ask for the five with the largest expected effect, without pretending that effect is guaranteed. I used to approve long lists on Mondays and spend the week undoing three of them. Fewer changes make the next review easier to interpret.

12. Turn the proposals into an approval list

When to use: Next, in the same chat. For a Google and Meta workflow in one groas MCP chat, make the affected account and platform explicit on every item.

Convert the proposed changes into an approval list dated [DATE]. Number only the proposed items, up to 5.

Show the platform, account, campaign or page, old value, new value and supporting data point for each.

Wait for my explicit reply: 'approve all' or 'approve' followed by specific item numbers. Do not treat 'looks good' or 'go ahead on those' as approval.

After any approved action, show what was applied, what failed, the change-log entry, the reason and the rollback step. Do not report a proposed action as completed.

The adjustment that matters: Treat this as an instruction, not an enforced permission control. If Claude acts on a vague reply, stop, inspect the log and roll back the change before continuing. With groas, every change is logged with its reason and can be undone in one step. Insist on seeing the record.

Save the record next Monday needs

When to use: After the approval decision. Save the chat with a date in its title, then paste this final artifact. A complete Monday chat shows the workflow assembled.

Write the review record for [DATE] so I can save it for next Monday.

Include:
1. Reporting windows and data sources.
2. Spend, conversions and CPA versus the prior week, with formulas.
3. AI citation rates, citation counts and valid-answer denominators for both weeks.
4. Proposed changes with old and new values.
5. What I approved, what was actually applied and the change-log references.
6. Unresolved data gaps.
7. The 3 numbers to re-check next Monday, each with an owner.

Keep it to one page without omitting any metric that moved more than 20%. Do not claim the record was saved unless you actually saved it.

The adjustment that matters: Preserve the actual numbers and approval status, not just the summary. Claude plus groas supplies the connected marketing data and actions; groas Unlimited is $19 a month for unlimited teammates, ad accounts and websites. Neither replaces a record of what was decided.

This is the artifact I rate most useful. The twelve prompts produce the review; this one preserves the formulas, approvals and re-checks that make next week’s review comparable. A Monday ritual is easy to repeat. A decision record is harder to misremember.