AI prompts for marketing metrics: 6 systems to copy

AI prompts should reduce wasted work before they touch a sales number. In about 60 minutes, you'll set up six reusable prompt systems for customer research, enquiries, email clicks, landing pages, content briefs, and weekly reports. You can paste them into the AI assistant you already use. Each system includes the source material, a check, and one metric to watch.

What you'll build

You'll create a small prompt library, not a folder of clever questions. Each prompt has five parts: the job, approved information, rules, a fixed answer shape, and a check a person completes before anything reaches a customer.

That structure follows current guidance from OpenAI, Anthropic, and Google. In plain words, tell the system exactly what success looks like, show it the material it may use, and make the answer easy to inspect. Examples help when you need the same shape every time.

Build one prompt first. Test it on five old jobs whose correct outcome you already know. Anthropic calls this an evaluation, which simply means a repeatable test. Keep the prompt only when it passes more of your checklist than the current process.

1. Mine customer language without inventing it

Start here if your website sounds polished but customers describe the problem differently. Gather 15 to 30 recent reviews, call notes, enquiry emails, or survey answers. Remove names, phone numbers, and any private details before pasting them into an AI tool.

The output is a message bank: exact phrases grouped by problem, desired result, hesitation, and buying trigger. It gives you language to test. It does not prove that every customer thinks the same way.

You are organizing customer research for [BUSINESS TYPE].

Use only the source comments between <comments> tags. Do not invent quotes or combine several comments into one quote.

Return a table with these columns:
1. Theme
2. Exact customer quote
3. What the customer appears to want
4. Where we could test this language: headline, email, FAQ, or sales reply

Include a theme only when at least two comments support it. Mark uncertain interpretations as "check manually."

<comments>
[PASTE ANONYMIZED COMMENTS]
</comments>

Check every quoted phrase against the original. Then add the approved themes to the brief you use for the other five prompts. If you are also working on search visibility, connect this message bank to the AI recommendations visibility setup.

2. Draft a useful reply to a new enquiry

A fast reply hurts trust when it answers the wrong question. This prompt turns a new enquiry into a short draft while keeping prices, availability, and promises inside approved limits.

Give it your service area, opening hours, price rules, booking process, and the customer's message. Never paste payment details or sensitive personal information. The number to watch is qualified reply rate: the share of enquiries that answer your question or take the requested next step.

Draft a reply to the customer enquiry below.

Goal: help the customer decide whether the next step fits them.
Tone: plain, calm, and specific. Maximum 120 words.

Rules:
- Use only the approved facts.
- Do not promise timing, price, or results that are not stated.
- Answer the customer's direct question first.
- Ask one necessary qualifying question.
- End with one next step.
- If a fact is missing, write [NEEDS HUMAN ANSWER].

<approved_facts>
[SERVICE AREA, HOURS, PRICE RULES, BOOKING STEPS]
</approved_facts>

<enquiry>
[CUSTOMER MESSAGE]
</enquiry>

Have a person approve every reply until five varied tests pass without an unsupported promise. If enquiries also need to enter the software that stores customers and sends follow-ups, use the lead-to-CRM automation build after the wording is stable.

3. Generate email angles you can actually test

Do not ask for “ten catchy subject lines.” That request has no customer, offer, proof, or destination. Feed the prompt one approved offer and one page, then ask for distinct reasons to click.

Click rate is the percentage of delivered emails where someone clicks a link. It tells you whether the message and destination connect, but it does not tell you whether those visitors bought. Track the sale or booking separately.

Create four email test angles for this offer.

Audience: [WHO RECEIVES IT]
Approved offer: [OFFER AND LIMITS]
Destination page promise: [WHAT THE LINKED PAGE DELIVERS]
Approved proof: [FACTS, REVIEWS, OR NONE]

For each angle, return:
- the customer concern it addresses
- one subject line under 45 characters
- one preview line under 80 characters
- a body of 80 to 120 words
- one link phrase that accurately describes the destination

Make each angle meaningfully different. Do not create urgency, discounts, proof, or scarcity that is not supplied. Finish with a checklist of claims a person must verify.

Send the current version and one new angle to comparable groups. Change the angle, not the offer and page at the same time. Otherwise you will not know what caused the difference.

4. Find the leak on a landing page

This prompt is a critic, not a page writer. Give it the page copy, the advertisement or email that sends visitors there, and five real customer objections. Ask it to identify mismatches before it rewrites anything.

Watch form completion rate: completed forms divided by visits to the page. A low number can come from weak copy, a long form, slow loading, or the wrong visitors. The prompt can inspect words. It cannot diagnose a broken page or poor traffic by itself.

Audit this landing page for message gaps.

Visitor source message:
[PASTE THE AD OR EMAIL]

Known customer objections:
[PASTE FIVE VERIFIED OBJECTIONS]

Landing page copy:
[PASTE PAGE COPY]

Return a table with:
1. Page section
2. Specific mismatch or unanswered objection
3. Evidence from the supplied material
4. Smallest proposed edit
5. What a person must verify

Do not rewrite the full page. Do not invent customer concerns, evidence, prices, guarantees, or testimonials. Rank no more than five edits by likely importance and explain the ranking in one sentence each.

Make the top edit, record the date, and leave the other variables alone long enough to compare results. One clean change beats a full AI rewrite because you can tell whether it helped.

5. Turn approved research into a content brief

AI can organize research quickly, but it can also state an unsupported guess with confidence. Supply the sources yourself. The prompt's job is to assemble a writing plan and expose missing evidence.

Track brief preparation time and the number of unsupported claims caught during editing. For a wider tool setup, compare this workflow with the AI writing tools for SEO.

Build a content brief for the target question: [QUESTION]

Reader: [WHO THEY ARE AND WHAT DECISION THEY FACE]
Business goal: [BOOKING, SALE, OR EDUCATION]

Use only the material inside <sources>. For every factual claim, name the source beside it. If the sources do not answer an important reader question, put it under "Research still needed."

Return:
- the direct answer in 40 words
- reader questions in decision order
- an outline with the purpose of each section
- a claim-and-source table
- terms that need a plain-English explanation
- five checks for the editor

Do not draft the article.

<sources>
[PASTE NOTES AND SOURCE URLS]
</sources>

Reject any brief that fills “Research still needed” with invented answers. Do the missing research, add the source, and run it again.

6. Explain a weekly report without changing the numbers

Owners need the reason behind a change, but a language model should not be allowed to quietly repair missing figures. Give it one approved table and require it to separate facts from possible explanations.

The metric here is reporting time. Accuracy is a gate, not a score. Every figure in the answer must match the supplied row. The full wiring is in the automated weekly marketing report.

Explain this weekly marketing report for a business owner.

Rules:
- Copy numbers exactly as supplied.
- Do not calculate a comparison unless both required numbers are present.
- Label each explanation as "possible reason," not fact.
- Do not claim one marketing activity caused a result.
- If data is missing, say "comparison unavailable."

Return exactly four sections:
1. What changed: maximum three bullets
2. What needs attention: maximum two bullets
3. Possible reasons to check: maximum three bullets
4. One decision for next week

<approved_report>
[PASTE THE APPROVED TABLE]
</approved_report>

Compare the answer with the source row before sending it. If one number differs, stop the workflow and fix the prompt or data handoff.

The part that breaks

Most prompt libraries fail because nobody owns the source material. A carefully worded instruction cannot rescue an old price list, missing customer context, or a spreadsheet with the wrong week. Assign one owner to each prompt and put an “information checked” date above the approved facts.

The second failure is changing the prompt after every odd answer. Build five test cases instead: an ordinary request, a short one, a messy one, one with missing information, and one that should be refused. Score the same five cases whenever you edit the prompt. That is how you separate an improvement from a lucky answer.

Upgrade path

Put the six prompts in one shared document with an owner, version number, approved inputs, five test cases, and the metric being watched. Review one prompt each month. Retire anything nobody uses.

Then automate only the boring handoff: place approved source data into the prompt and route the draft to a review queue. Your AI tool creates the draft. A person still checks the facts and decides whether it goes out.

Get the next build from the prompt systems library, or browse all Runbook articles.