Claude Prompts for Marketing Analytics: 15 Templates for GA4, Attribution and Campaign Reads (2026)

The best Claude prompts for marketing analytics give Claude the export, the definitions, and the decision you are trying to make, and then ask it to say what it assumed before it says what it found. Marketing data is unusually easy to misread: every tool counts conversions differently, tracking breaks quietly, and a campaign can look like a winner because it took credit for sales that would have happened anyway. These 15 templates cover the weekly work (channel readouts, campaign post-mortems, email and paid search reviews) and the harder questions underneath it (why GA4 and the ad platform disagree, which channels pay back, and what a channel actually adds).

Each prompt is written for the same habit: profile and question the data first, compute second, and end with a decision and a caveat. That is the same pattern behind our Claude prompts for data analysis, applied to the specific traps of marketing numbers.

Which prompt to use for which job

• Monday morning review: the weekly channel readout, then the triage prompt for anything that moved more than usual.
• Numbers that do not match: the two-tools-disagree prompt before any meeting where someone will ask why the dashboard and the ad account show different results.
• A campaign just ended: the post-mortem prompt, with the A/B test read for any landing page or creative split inside it.
• Budget planning: CAC and payback by channel, then the budget reallocation check, then the holdout design if a big channel's contribution is in doubt.
• SEO and email: the Search Console CTR audit and the email read, both of which ignore the vanity metric (rank, opens) and judge the outcome (clicks, conversions).
• Reporting up: metric definitions once, then the monthly leadership summary every month.

A weak marketing analytics prompt, and the same prompt fixed

A typical first attempt:

Here's our marketing data. Which channel is performing best?

Claude will answer, and the answer will sound reasonable. It will also have guessed what "performing" means (clicks, conversions, revenue, or profit), which attribution the export used, whether the last week is complete, and whether "best" means cheapest per conversion or most total return. The same request, fixed:

<role>You are a marketing analyst preparing a budget recommendation.</role>
<data>Attached: channels_q3.csv, one row per channel per week, Jul 1 to Sep 30 2026. Conversions are GA4 key events, data-driven attribution. Spend is from each ad platform.</data>
<question>Which channel should get the next $5,000 a month: the one with the lowest cost per new customer, or the one with the best 90-day revenue per customer?</question>
<constraints>Organic and direct have no spend; report them but exclude them from cost comparisons. Treat any channel with fewer than 30 conversions in the quarter as inconclusive.</constraints>
<output>Check the file first and stop if anything looks wrong. Then a table by channel. Then a recommendation, and two reasons it could be wrong.</output>

It takes a minute longer to write. It removes the guesses, names the attribution model so Claude does not invent one, and turns "which is best" into a decision with a stated rule. Every template on this page follows that shape.

Attribution: what Claude can and cannot tell you

Attribution is where marketing analysis most often goes quietly wrong, and no prompt can fix it from inside one export. Three facts are worth knowing before you ask Claude to compare channels.

First, the model decides the answer. GA4 reports with data-driven attribution by default, and in 2023 Google removed the first-click, linear, time-decay and position-based models from GA4 and Google Ads, leaving data-driven and last-click variants. A channel that starts journeys (social, content, podcasts) looks weaker under last-click and stronger under data-driven. If two people compare channels using exports with different models, they will disagree and both be right.

Second, every ad platform grades its own homework. Meta's default attribution setting counts conversions within 7 days of a click and 1 day of a view; Google Ads defaults to a 30-day click window. Add up what each platform claims and the total is usually larger than your actual sales. The two-tools-disagree prompt exists to make that gap explicit instead of arguing about it.

Third, attribution measures credit, not cause. A retargeting campaign shown to people already on their way to checkout will collect a lot of credit and add little. The only way to measure what a channel adds is to turn it off for a comparable group and compare, which is what the holdout prompt designs. Claude is good at planning that test and reading its result; it cannot infer incrementality from an attribution report, and the prompts tell it not to try.

Same position, eight times the clicks: a CTR example

The Search Console prompt is on this page because rank alone hides the most fixable problem in organic traffic. Our own numbers show it. In the 28 days to October 31, 2026, this site's two largest pages ranked at nearly the same average position and got very different results: the Claude data analysis pack had 1,715 impressions, 54 clicks and a 3.15% click-through rate at position 6.5, while our glossary entry on RSI in AI had 1,252 impressions, 5 clicks and a 0.40% rate at position 7.1.

Same site, same part of the results page, eight times the click rate. The difference is the query. People searching for prompt templates want the templates and have to click to get them. People searching "what does RSI mean in AI" often get the answer from the results page itself and never need to click. A CTR audit that only looks at position would tell you to rewrite both titles; the prompt asks Claude to look at the queries and separate a snippet problem from an intent the page cannot win a click for. Only the first is worth a rewrite.

Getting the data to Claude

There are three ways to work, and they suit different questions:

• Attach an export. Download CSVs from GA4, the ad platforms, Search Console or your email tool and attach them with code execution switched on. Claude then computes numbers by running code, and you can ask to see and rerun it. This is the default for anything numeric.
• Paste a schema. If the data lives in a warehouse, paste the table definitions and ask for SQL. Claude cannot see the warehouse, so it writes the query and you run it. The data analysis pack covers this workflow in depth, including long schemas and join fan-out.
• Connect a tool. Analytics and ad tools increasingly expose their data through MCP servers, which let Claude query them directly inside a conversation. It is convenient, and it makes one rule more important: ask Claude to state the property, date range and filters of every query it ran, because a connected tool makes it easy to get a confident answer from the wrong slice.

Whatever the route, aggregate before you share. Most marketing questions are answered at channel, campaign, cohort or week level, and aggregated data keeps both the file small and personal data out of the conversation.

Five ways marketing numbers lie

1. Tracking changes read as performance changes. A new consent banner, a tag manager edit, or a redesigned thank-you page can halve reported conversions overnight. That is why the triage prompt checks the plumbing before the campaign.
2. Open rates. Mail privacy features load images automatically for many recipients, which inflates opens. Judge email on clicks and conversions.
3. Averages instead of margins. A channel's average cost per acquisition says little about what the next dollar will buy there. Budget decisions need the marginal view the reallocation prompt asks for.
4. Small samples. Thirty conversions split across six creatives is five per creative, which is noise. The post-mortem and A/B prompts both ask for the sample size behind each claim.
5. Partial periods and seasonality. Comparing a week with a holiday in it, or a month three days in, against a full one. The weekly readout asks Claude to check completeness first.

For the wider marketing workflow (positioning, campaign planning, content), use the ChatGPT prompts for marketing pack; for customer research and survey synthesis, the market research prompts; for organic search work beyond the CTR audit, the SEO prompts. When a readout format works for your team, save it as a reusable template in Prompt Builder so every week's report asks the same questions of the data.

You are a marketing analyst preparing the weekly channel review. Attached is an export with one row per channel per week: [columns, e.g. week, channel, sessions, conversions, revenue, spend]. Before any analysis, tell me the date range, whether the last week is complete, and any channel whose definition looks like it changed mid-series. Then give, per channel: this week versus the trailing 4-week average for spend, conversions, cost per conversion and revenue. Flag only changes larger than the normal week-to-week range you can see in the data. End with three sentences for the team: what moved, the likeliest reason, and what we check next.

Our [ad platform] reports [N] conversions for [campaign] in [period], GA4 reports [N], and our CRM shows [N] new customers from it. Explain every reason these three numbers can differ, ordered by how much of the gap each typically explains: attribution model, click and view windows, which touch gets credit, consent and ad blockers, cross-device journeys, time zones, conversion counting (once versus every), and deduplication. For each, tell me the exact setting or report I should open to measure its share of the gap here. Do not pick a winner; tell me which number to use for which decision.

Here is a list of every source / medium / campaign combination from the last [period] with sessions: [paste]. Find: inconsistent spellings and capitalisation of the same source, paid traffic tagged as organic or referral, internal links carrying UTM parameters, campaigns with no medium, and anything that will land in GA4's Unassigned channel. Group the fixes into a cleanup table with the current value, the corrected value and the sessions affected, then write a one-page UTM naming convention we can hand to everyone who builds links.

Campaign: [name, goal, dates, budget, channels, audiences, creatives]. Results: [paste metrics by channel and creative]. Write a post-mortem: did it hit the goal it was set, which channel and creative carried it, which spent money without return, and which results are too small to conclude anything from (say the sample size you would need). Separate what we learned about the audience from what we learned about execution. End with three changes for the next campaign, each tied to a number above.

A/B test on [page]: control had [visitors] visitors and [conversions] conversions, the variant had [visitors] and [conversions], over [days] days. Compute both conversion rates, the relative lift, a 95% confidence interval for the difference, and the p-value, and show the calculation. Then tell me whether the test ran long enough to cover full weekly cycles, whether traffic mix shifted between arms, and whether we peeked early. Give a one-line decision: ship, keep running (for how long), or call it a tie.

Here are, per acquisition channel for [period]: spend, new customers, average first-order value, gross margin [%], and monthly retention or repeat revenue per customer: [paste]. Compute blended and per-channel customer acquisition cost, months to pay back CAC from gross margin, and 12-month gross profit per customer. State every assumption, especially how you split spend that is shared across channels and how you treat organic customers. Flag any channel whose result rests on fewer than 30 customers.

I want to move [amount] of monthly budget from [channel A] to [channel B] because B has a lower cost per acquisition. Here is weekly spend and conversions for both over the last [N] weeks: [paste]. Using the spend variation in this data, estimate whether each channel shows diminishing returns, and what the last dollar spent in each channel actually bought, not the average. Tell me what the data can and cannot support, and propose a staged test (amounts and weeks) that would answer the question with less risk than moving it all at once.

Here is our search terms report for the last [period]: [paste search term, match type, clicks, cost, conversions]. Group the terms by intent (buying, researching, job seekers, competitors, irrelevant). For each group give spend, conversions and cost per conversion. List negative keyword candidates with the match type to use, and flag any term that looks irrelevant but converted, so we do not block it by accident. Keep brand and competitor terms in separate sections.

Here is a Google Search Console export of pages with impressions, clicks, CTR and average position for the last 28 days: [paste]. Find pages whose CTR is well below what other pages at a similar position get on this same site. For each, look at the queries it ranks for: [paste query rows], and say whether the gap is likely the title and description, a search intent the page does not match, or a results page where the answer is shown without a click. Suggest a new title and description only where the problem is the snippet.

Email results for the last [N] sends: [paste date, audience size, delivered, opens, clicks, unsubscribes, conversions, revenue]. Treat open rate as unreliable because of mail privacy protection and image proxies, and judge each send on click rate, conversions per thousand delivered and unsubscribes. Identify which subject lines, send days, and segments perform best, and say which differences are large enough to act on given the list sizes. Suggest the next two tests to run.

Here is a cohort table of customers by acquisition channel and signup month, with how many were still active or purchasing in each later month: [paste]. Compare retention curves by channel, not just month one. Tell me which channel brings the customers who stay longest, whether cheap channels bring short-lived customers, and whether any difference could be explained by when each channel was active (seasonality) rather than the channel itself. Then rank channels by retained customers per unit of spend.

I want to know how many conversions [channel or campaign] actually adds, rather than how many it gets credit for. Our setup: [regions, audience size, weekly conversions, budget]. Design an incrementality test: geo holdout or audience holdout, with how to choose the holdout group, the test length, the minimum detectable effect at our volumes, and what we measure. Write down in advance what result would make us cut, keep or grow the channel. List the ways the test could be contaminated and how to guard against each.

Write precise definitions for these metrics as our company should use them: [list, e.g. lead, MQL, SQL, CAC, ROAS, conversion rate]. For each give the formula, the source system, the time window, what counts and what is excluded (test accounts, refunds, existing customers, internal traffic), and two edge cases with the correct handling. Flag every place where marketing and finance commonly define the metric differently, and ask me the questions you need answered to settle it.

[Metric, e.g. conversions from paid social] moved from [usual level] to [new level] on [date]. Give me a triage list in the order to check it, starting with the cheapest causes to rule out: a tag or pixel change, a consent banner or cookie change, a site release that broke a form or thank-you page, a change in conversion counting or attribution settings, bot or spam traffic, a campaign paused or a budget cap hit, then a real change in demand. For each, say what to open and what I should see if it is the cause.

Turn this month's marketing data into a one-page summary for leadership: [paste key metrics versus target and versus last month, plus notes]. Lead with the single number that matters most against target and whether we are on track. Then three points: what worked, what did not, and what we are changing, each with one supporting number. Include one honest caveat about data quality or attribution. No channel jargon, no more than 250 words, and no number that is not in the data I gave you.