7 Ways Marketing Teams Are Actually Using Claude (With Prompts and Setup Instructions)

Most marketing teams are using Claude wrong. They paste in a brief, ask for five ad copy variants, pick the best one, and call it AI-enabled marketing. That's just faster copy editing. It doesn't compound.

The teams getting real leverage are using Claude differently — as a reasoning layer that sits on top of their actual data. They give it context about their business, their metrics, their anomalies, and ask it to think through the same frameworks a senior analyst would. The output isn't "five subject line options." It's a diagnosis, a next action, or a structured output that feeds directly into a report or a pitch.

This post covers seven specific workflows: what the setup looks like, what prompt structure works, and what to watch out for. Each is something practitioners are actively running, not theoretical.

1. RCA Framework: Stop Describing Data, Start Diagnosing It

The default way marketers use AI for data analysis is describing numbers at it. "My CPA went up 23% this week, what should I do?" Claude doesn't know your account, your seasonality, your creative rotation schedule, your audience overlaps, or which campaigns are evergreen versus flight-based. The output is generic because the input is generic.

The fix is front-loading context so Claude can run an actual diagnostic.

The context block (paste this at the start of any analysis):

ACCOUNT CONTEXT
- Primary KPI: [e.g. Purchase ROAS, target 3.5x]
- Secondary KPIs: [e.g. CPL < $18, CTR > 1.2%]
- Current flight: [e.g. Aug–Sep back-to-school, $120k budget, ends Sep 15]
- Creative rotation: [e.g. 3 video concepts, 2 static, currently in exploration phase]
- Audience structure: [e.g. Broad + ABA layered, retargeting 14d window]
- Known seasonality: [e.g. CPAs historically inflate 15–20% during this period YoY]
- Recent changes: [e.g. New pixel events added Aug 28, landing page swap Sep 1]

ANOMALY CHECKLIST (run before flagging as a performance issue)
- Is this within normal weekly variance? (±10% for this account)
- Did anything change in the last 48–72h? (creatives, bids, audiences, landing pages)
- Is the reporting window current? (attribution lag expected for this KPI)
- Is this isolated to one campaign/adset or account-wide?

Then drop in your data and ask: "Given this context and the anomaly checklist, walk through what's likely driving the change in CPA from $14 to $19 this week. Check each hypothesis before concluding."

The key word is check each hypothesis. Without that, Claude summarizes. With it, it reasons.

Add your "if X, check Y" rules. These are the institutional patterns your team has learned the hard way:

IF CPA spikes but CTR holds → suspect landing page or post-click issue
IF CTR drops and CPM holds → creative fatigue, check frequency > 3.5 on top segments
IF ROAS drops but conversion volume holds → ASP issue or product mix shift
IF CPA inflates on mobile only → check mobile landing page, payment flow
IF spend paces behind → check audience size, bid competitiveness, budget allocation

Give Claude these rules explicitly. It will apply them. When you don't, it invents its own framework that may or may not match how your account actually works.

2. Ad Copy Writing: Build a Skill File for Your Brand

Generic AI copy fails because the model doesn't know what makes your product different, what words your customers actually use, or what your team has already tested and killed. You can fix all three with a brand skill file — a structured prompt you paste at the start of every copy session.

Here's the template:

BRAND SKILL FILE — [Brand Name]

POSITIONING
- One-line: [e.g. "The only ride-hailing app where drivers keep 100% of the fare"]
- Core tension we exploit: [e.g. Drivers hate commission cuts. Riders want fair prices. We fix both.]
- Category we're reframing: [e.g. Not a taxi app — a zero-commission marketplace]

PROOF POINTS (ranked by believability)
1. [Specific, verifiable claim — e.g. "₹0 commission since day one"]
2. [Outcome-based claim — e.g. "Drivers earn 40% more per trip vs. incumbent"]
3. [Social proof — e.g. "2.1 lakh drivers in Bengaluru"]

VOICE
- We sound like: [e.g. Direct, confident, slightly irreverent — like the driver who tells it straight]
- We do NOT sound like: [e.g. Corporate, preachy, startup-generic]

WORDS WE USE
- [fair, driver-first, zero commission, what you earn, real income, straightforward]

WORDS WE AVOID
- [seamless, leverage, unlock, empower, innovative, disrupting, journey, ecosystem]
- Avoid em dashes (—), avoid bullet lists in ad copy, avoid rhetorical questions

WHAT'S BEEN TESTED AND KILLED
- "Join thousands of drivers" → low CTR, feels generic
- Fare comparison tables → too complex for top-of-funnel
- Urgency CTAs ("Limited spots") → does not match driver audience expectations

FORMAT CONSTRAINTS
- Headlines: 30 characters max, no truncation
- Descriptions: 90 characters max
- No exclamation marks unless A/B testing specifically for them
- Always end with a clear action: "Start earning" not "Learn more"

Before writing any copy, paste this block and say: "Using this brand skill file, write [5 headlines / a video script / 3 description variants] for [specific campaign objective, audience, and offer]."

The difference in output quality between this and a bare request is significant — not because Claude is more creative but because it has guard rails. It won't drift into generic benefit statements because the "words we avoid" list catches them before they're suggested.

Update the skill file quarterly. When something new gets killed in testing, add it. When a new proof point outperforms, promote it.

3. N-Gram Analysis for Google Search Terms

N-gram analysis — breaking search queries into 1-, 2-, and 3-word sequences and comparing CPA/ROAS by n-gram — is the single highest-signal analysis most PPC teams are running inconsistently or not at all. It finds patterns in what's converting and what's burning budget that query-level review misses entirely.

The manual process: export your Search Terms report from Google Ads, paste the data into Claude, and run this prompt.

Prompt:

Here is my Google Ads Search Terms report for the last 30 days.
Columns: search term, impressions, clicks, conversions, cost, conversion value.

Run a 1-gram, 2-gram, and 3-gram frequency analysis.
For each n-gram:
- Count how many queries contain it
- Sum impressions, clicks, conversions, and cost across those queries
- Calculate CPA (cost/conversions) and ROAS (conv. value/cost)

Then give me three ranked lists:
1. HIGH-VALUE N-GRAMS: Appear in 3+ converting queries, ROAS > account average. These should go into exact match or phrase campaigns.
2. WASTE N-GRAMS: Appear in 3+ queries, zero conversions, cost > $[threshold]. These should become negatives.
3. AMBIGUOUS N-GRAMS: High spend, low conversion count — need human review.

[paste your search terms data below]

Claude handles the aggregation, surfaces the patterns, and gives you a structured output you can act on immediately. A 5,000-row search terms export that would take 3–4 hours of manual pivot table work takes under 2 minutes.

What to do with the output:

• High-value 2-grams → build a new ad group targeting them in phrase or exact match with tightly themed copy
• Waste 1-grams → add as phrase or exact match negatives at the campaign level (check for cross-campaign impact first)
• Ambiguous n-grams → look at the query-level data to understand context before deciding

This is most powerful when run monthly with a consistent threshold. The first run is diagnostic. The second run shows whether your structural changes worked.

4. AEO: Query Fanning and Keyword Research for Answer Engine Optimization

Answer Engine Optimization is the practice of optimizing for queries that AI search engines (Google AI Overviews, Perplexity, ChatGPT) answer with cited sources. The queries that get AI-answered are specific, intent-clear, and often underserved by traditional keyword research tools because they don't have high search volume — but they have high conversion intent.

Step 1: Query fanning with Claude

Start with your core topic. Ask Claude to fan it out:

My core topic: [e.g. "UTM parameters for marketing"]

Generate 40 specific questions a marketing practitioner would type into an AI search engine about this topic.
Include:
- How-to questions (procedural intent)
- Comparison questions ("X vs Y")
- Troubleshooting questions ("why does X happen")
- Definition questions ("what is X")
- Scenario questions ("when should I use X")

Format as a plain list. No grouping yet.

You'll get 40 queries that represent actual user intent rather than volume-optimized keyword variants. These are the questions your content needs to answer completely enough that an AI will cite it.

Step 2: Cross-check volume and competition

Take your 40 questions into Google Keyword Planner. Filter for: monthly searches > 50, competition = Low or Medium. You're looking for questions with real search volume that haven't been fully colonized by high-DA content farms.

Then open Google Trends and compare 3–5 of the shortlisted queries. You want queries with flat or rising trend lines — declining queries mean the AI has already answered them well enough that users stopped searching.

Step 3: Validate AEO eligibility

Search each target query in Google. If it triggers an AI Overview, the query is AEO-eligible. If not, it's still a standard SEO target.

For AI Overview queries: the citation usually comes from content that answers the question directly and completely in the first 300 words, with a clear factual structure (definitions, numbered steps, or comparison tables). Ask Claude to draft that structure before you write the full piece.

Prompt:

Target query: [paste query]
My content brief: [paste what the page is about]

Write the first 300 words of this page optimized for AI Overview citation.
Requirements:
- Answer the query directly in the first sentence
- Include the key definition, process, or comparison the query implies
- Use a structure (numbered steps or a brief table) that AI can extract as a clear answer
- Match the reading level of a marketing practitioner
- Do not pad with intro paragraphs or generic context

5. Tech SEO Audit and Page Speed Automation

Manual tech SEO audits are time-consuming and inconsistent. The version that actually runs regularly is the automated one, which means you need a setup that doesn't require a developer to touch it each week.

The lightweight version: a Claude Projects session with a standing system prompt that gets fed new crawl data each Monday.

System prompt for your Claude Project:

You are a technical SEO auditor for [site name]. 

Site profile:
- Platform: [e.g. Next.js 15, deployed on Vercel]
- Primary crawl tool: [e.g. Screaming Frog, Ahrefs Site Audit]
- Weekly crawl exports will be pasted here each Monday
- Core Web Vitals baseline: LCP < 2.5s, INP < 200ms, CLS < 0.1

Your weekly checklist on receiving crawl data:
1. Flag any new 4xx errors — compare to last week's list
2. Flag any pages that moved from 200 to 3xx — report destination
3. Check canonical discrepancies (canonical pointing to non-indexable URL)
4. Flag any page with title tag > 60 characters or < 30 characters
5. Flag meta descriptions missing or > 160 characters
6. Check for pages with duplicate H1s or missing H1
7. Flag any new orphan pages (internal links = 0)
8. Check Core Web Vitals deltas from baseline — flag regressions > 15%

For each issue, output:
- Severity (Critical / Warning / Info)
- URL(s) affected
- One-sentence description of the issue
- Recommended fix

Group by severity. Skip issues we've already flagged in previous weeks unless they're unresolved.

Once this project is set up, Monday audit = paste crawl export, get prioritized issue list in 90 seconds.

For page speed specifically: export your Core Web Vitals data from Google Search Console (Performance → Core Web Vitals → Export), paste it into the project with the prompt "Compare this week's CWV data to baseline. Flag any URLs where LCP, INP, or CLS regressed more than 15%." Claude will isolate the regressions and let you focus attention rather than scanning a full table.

6. Metabase to Claude: Pull the Data You Actually Need

The standard workflow: you need a number, you ask your analyst, they pull it in two days, you get a table that half-answers your question, you ask for a cut, two more days. The bottleneck isn't the analyst's skill — it's the queue and the back-and-forth to clarify what you actually wanted.

The Metabase MCP connects Claude directly to your Metabase instance. Once connected, you can ask questions in natural language and Claude writes the SQL, executes it, and returns the result.

Setup (one-time, takes 15–30 minutes):

1. Install the Metabase MCP server (available on the Claude MCP marketplace)
2. Provide your Metabase instance URL and API token
3. Let Claude index your table schema (it reads the database structure, not your data)

What it looks like in practice:

Pull me daily installs broken by channel (organic, paid, referral) for the last 30 days.
Also show 7-day rolling average next to each day.

Claude writes the SQL, runs it against Metabase, formats the output. If your schema uses non-obvious table names (which most do), you can add a schema reference to your system prompt:

Our key tables:
- events.installs → app installs, columns: date, user_id, channel, campaign_id
- marketing.spend → daily spend by channel, columns: date, channel, spend_usd
- users.signups → account creation, columns: created_at, user_id, source
Campaign IDs join to marketing.campaigns on campaign_id.

With that context, Claude can write accurate joins without guessing.

What this is useful for:

• Quick gut-checks before a meeting ("what's our D7 retention for the Aug cohort?")
• Exploratory analysis where you're not sure what you're looking for
• One-off reports that don't justify a proper dashboard

What this is not useful for: production reporting, dashboards other people rely on, or analysis where audit trails matter. For those, keep using your analyst and proper BI workflows.

7. Artifacts with Claude for Executive Presentations

Executive presentations are usually built the wrong way: analyst creates a data file, strategist builds slides from it, slides get reviewed, someone asks a question, analyst re-pulls, slides update. Three loops, two days gone.

Claude Artifacts (available in Claude.ai and Cowork mode) let you build a self-contained interactive HTML presentation that pulls from your data and updates on demand. It's not a replacement for a polished deck — it's the right format for a working session where the exec will ask questions and you'll need to move through scenarios in real time.

What an artifact session looks like:

You paste in your data (or connect via MCP). You ask Claude to build an artifact. Claude creates a self-contained HTML page you can open in a browser, share via URL, and update with a single follow-up message.

Here is our Q3 performance data by channel: [paste data]

Build an executive artifact with:
- Header: Q3 Marketing Performance Review
- Summary row: total spend, total revenue, blended ROAS
- Table: each channel with spend, revenue, ROAS, YoY delta
- A simple bar chart showing ROAS by channel
- A "so what" section with 3 bullets: what worked, what didn't, what we'd do differently

Make it clean, no decoration. Should work in any browser, no login required.

The artifact is interactive — the exec can open it on their phone between meetings. If they ask "can you add a conversion volume column," you add it in one message and the artifact updates.

For board-level presentations: keep using slides. Artifacts are for internal working sessions, not for delivering polished narratives to external audiences who expect production quality.

What Actually Determines Whether This Works

The marketers getting leverage from Claude share one thing: they've invested time front-loading context. The brand skill file, the RCA framework, the schema reference for Metabase — none of these are complex to build. They each take 30–60 minutes the first time. After that, every session starts from a higher base and the marginal quality improvement per query compounds.

The failure mode is treating Claude like a search engine. You wouldn't Google "why did my CPA go up" and expect a diagnosis. You need to bring the information. Once you do, the quality of reasoning is high enough that senior practitioners trust the output as a starting point for their own analysis — not a finished answer, but a structured first draft of a diagnosis or a document that would have taken two hours to produce manually.

That's the real gain: compressing the time between "I have data" and "I have a direction." Not replacing the thinking, but removing the friction that slows it down.

Want to build UTM parameters for your campaigns, shorten and track your links, or create QR codes that log analytics? All of these are free on MarketerTools — no account required.