AI Marketing Workflows: 10 Prompts for Your Own Stack

How these workflows run

MCP (Model Context Protocol) is a standard way for an AI client to call tools in other apps. If it is new to you, start with what MCP is.

You connect each marketing tool once. After that the model can pull reports and, on a paid plan, make changes.

Each workflow below gives the connections it needs, a prompt, and the tools the model calls. The tool names are from PopMCP's servers. Another MCP server will name them differently, but the prompts carry over.

All ten work with read access. Workflows 5 and 8 have an optional write step.

None of them runs on a schedule. PopMCP does not schedule prompts, so "last week's report" is something you ask for on Monday. For a number you check at the same time every day, a dashboard is the better tool.

10 AI marketing workflows, with prompts

1. Last week's Meta and Google Ads report

Pull last week's spend, clicks, conversions and cost per conversion
for every active campaign in Meta Ads and Google Ads. Compare each
with the week before. One table, sorted by spend. Then list the
three biggest changes.

Needs Meta Ads and Google Ads. The model calls meta_ads_get_account_insights, broken down by campaign, and runs a GAQL query through google_ads_search.

A conversion is whatever each account is set up to count. Keep the two platforms in separate rows and do not add them into one total.

2. Search Console content decay

Compare clicks by page for the last 3 months with the same 3 months
last year. List pages that lost more than 30% of their clicks. For
each one, show the queries that dropped most and the current
average position. Rank by clicks lost.

Needs Google Search Console. The model calls gsc_compare_periods with the same-dates-last-year option, then gsc_queries_for_page for each losing page.

Search Console keeps 16 months of data, so a year-over-year check on a 3-month window fits. Google's API exposes at most 50,000 rows a day per search type, sorted by clicks. A very large site gets its top pages, not every page.

3. Klaviyo flow audit

List every live flow with its trigger type and number of messages.
For the last 90 days add recipients, open rate, click rate and
revenue. Flag flows with no revenue and flows with a click rate
under 1%. Which of these do we not have: welcome, abandoned
checkout, browse abandonment, post-purchase, winback?

Needs Klaviyo. The model calls klaviyo_get_flows and klaviyo_query_flow_values. Message counts come from klaviyo_get_actions_for_flow, which it finds through klaviyo_search_tools.

Klaviyo's flow report needs a conversion metric, such as Placed Order. If you track more than one, say which. Klaviyo lists six flow trigger types: added to list, added to segment, metric, date property, price drop and low inventory.

4. A GA4 drop, traced to a cause

GA4 sessions dropped this week. Find the day it started. Break the
change down by channel, device, landing page and country. Then pull
daily spend from Meta Ads and Google Ads for the same days. Which
segment explains most of the drop, and what did you not check?

Needs GA4, Meta Ads and Google Ads. The model calls ga4_run_report several times, then meta_ads_get_account_insights and google_ads_search.

GA4 has its own anomaly detection, which flags an unusual value in a metric's time series. This prompt goes one step further and lines the date up against ad spend.

The model's cause is a hypothesis. The last line of the prompt is there so it tells you what it skipped.

5. HubSpot lead triage

Show contacts created in the last 7 days with lifecycle stage Lead.
Group them by original source. For each, show job title, company
and company size. Pick the 10 I should call first and give one line
of reasoning for each.

Needs HubSpot. The model calls hubspot_contacts_search, then hubspot_associations_list and hubspot_companies_get for company size.

The ranking is the model's opinion from the fields it read. It is not your HubSpot lead score unless you ask it to include that property.

Optional write step: Create a follow-up task due tomorrow for each of the 10. That calls hubspot_tasks_create and runs as soon as the model calls it, so read the list first. Writes need a paid plan or the trial.

More CRM prompts are in AI CRM with MCP.

6. Okendo review mining

Read the 200 most recent reviews across the store. Group the
feedback into themes such as sizing, quality, shipping and price.
Quote two short reviews per theme. Then draft 5 ad headlines that
reuse the customers' own words.

Needs Okendo. The model calls okendo_list_reviews, which returns published reviews, most recent first.

For a per-product view, okendo_list_product_reviews and okendo_get_product_reviews_keywords work on one Shopify product at a time. Connect Shopify as well if you want it broken down by product.

7. Striking-distance queries to a keyword plan

List queries where we rank between positions 5 and 20 with at least
100 impressions last month. For the top 20, get Google Ads keyword
ideas and average monthly searches. Mark each as a new page, a page
update or a paid test.

Needs Search Console and Google Ads. The model calls gsc_striking_distance_queries, whose default range is positions 4 to 20. Then google_ads_generate_keyword_ideas and google_ads_generate_keyword_historical_metrics.

8. Google Ads search terms with spend and no conversions

Pull search terms from the last 30 days with spend and zero
conversions. Group them by theme. Suggest negative keywords and a
match type for each. Do not change anything.

Needs Google Ads. The model runs a GAQL query on search_term_view through google_ads_search.

The report is a read. Adding the negatives is a write: google_ads_mutate_campaign_criteria at campaign level, or google_ads_mutate_ad_group_criteria for one ad group. It needs a paid plan and runs when the model calls it. Pasting the list into Google Ads yourself is the safer first run.

9. Email campaign review

For every Klaviyo email campaign sent last month, show subject
line, send time, recipients, open rate, click rate, unsubscribes
and revenue. Which subject lines and send times did best? Suggest
two things to test next month.

Needs Klaviyo. The model calls klaviyo_get_campaigns, klaviyo_get_messages_for_campaign and klaviyo_query_campaign_values.

With only a few campaigns in the month, treat "best send time" as a guess to test.

10. Monthly client report for agencies

For the Acme workspace, summarize last month across Meta Ads,
Google Ads, GA4 and Klaviyo: spend, revenue, ROAS, sessions and
email revenue against the month before. Write 5 plain bullets for
the client and 3 next actions for our team.

Needs one workspace per client. With PopMCP each client's connections and usage data sit in their own workspace, and every workspace is reached from the same URL.

Name the client in the prompt. When more than one connection of a provider is authorized, the model picks one by its ID, and list_connections shows which workspace each belongs to.

Ask it to state which accounts the numbers came from. See MCP for agencies.

How to set it up

1. Pick one report you already build by hand. You know what the right answer looks like, so you can check the model.
2. Connect the tools in PopMCP. Each provider is authorized once in a workspace. Expect some credential work: Google Ads needs your own Google OAuth client and a developer token, GA4 and Search Console need your own Google OAuth client, Meta Ads takes a System User access token, Klaviyo a private API key and HubSpot a Service Key.
3. Add https://app.popmcp.com/mcp to your AI client. It opens a PopMCP sign-in and you tick which connectors that client can reach. There is no token to paste. See the Claude steps or the ChatGPT guide.
4. Run the prompt and check the numbers. Compare the first few answers with the source tool.

Claude supports custom connectors on Free, Pro, Max, Team and Enterprise plans, and Free is limited to one. ChatGPT developer mode is available to Pro, Plus, Business, Enterprise and Education accounts on the web.

What makes the answers reliable

Name the date range and the metric. "Last 7 days, cost per purchase" can be checked against the ad account, and "how are ads doing" cannot.

Ask for the table before the commentary, because a table is quick to check. Then ask what it did not check and which accounts it read.

Keep a human on the writes. The MCP spec says there should always be a human in the loop who can deny a tool call.

PopMCP does not hold a write for approval, so that check is your AI client's. Claude shows a tool approval request, and ChatGPT developer mode asks for confirmation on write actions by default. Use Claude's "Allow always" only for read tools you trust.