Gemini Marketing Persona Prompt: A Campaign-Specific Critique Workflow

Gemini Marketing Persona Prompt: A Campaign-Specific Critique Workflow

A useful Gemini marketing persona is not a theatrical biography. It is a review contract. It tells the model which campaign decision it must support, which evidence it may use, which standards it must apply, what it must refuse to assume, and how its output will be checked by a human. If your “expert persona” keeps returning generic advice, the usual problem is not that the persona lacks personality. The prompt lacks a campaign brief, a claim ledger, a scoring rubric and a decision rule.

This guide provides a reusable workflow for reviewing a campaign before launch. The output is not “better copy” in the abstract. It is a structured critique, an issue log, a revision brief and a clear recommendation: proceed to human review, revise, or stop pending evidence.

Why marketing personas produce generic feedback

The instruction “act as a senior marketer” leaves almost every important variable undefined. A senior marketer could optimize awareness, qualified pipeline, conversion, retention, brand consistency or regulatory risk. The same creative can be acceptable for one objective and weak for another. When the prompt does not specify the decision, evidence and constraints, the model fills gaps with common advice: clarify the value proposition, strengthen the call to action and understand the audience.

Google’s official Gemini prompt-design guidance recommends clear and specific instructions, consistent structure, examples, decomposition of complex tasks and iteration. It also explains that current or obscure claims may require grounding, while calculations may require code execution. Those principles matter more than a colorful role description. A model can imitate a voice, but it cannot infer your approved claim, actual offer, channel limit or success threshold reliably.

Generic feedback usually has one or more of these causes:

1. No named decision. “Review this campaign” does not say whether the reviewer must approve the offer, select a hook, identify unsupported claims or diagnose a landing-page mismatch.
2. No evidence packet. The model sees copy but not the customer research, product facts, price, objections, past results or legal restrictions behind it.
3. No evaluation rubric. “Good” is undefined, so the answer becomes a list of broadly desirable qualities.
4. No output schema. Without fields, the model mixes observations, guesses and recommendations.
5. No uncertainty rule. The model may complete missing facts instead of marking them as unknown.
6. No human gate. The prompt treats fluent output as approval rather than as analysis for a responsible decision maker.

The remedy is to build a small review system rather than one clever sentence.

The PERSONA-7 campaign critique specification

Use seven fields. Each one prevents a different failure mode.

Field	What to provide	Why it matters
Purpose	The exact decision and stage	Prevents an unfocused review
Evidence	Approved facts, research and sources	Separates support from invention
Role boundary	What the model may and may not decide	Preserves human accountability
Standards	A scored rubric and channel rules	Makes “good” observable
Exceptions	Missing-data and uncertainty rules	Stops confident gap filling
Needed output	Fixed tables and labels	Makes results comparable
Approval rule	Thresholds and stop conditions	Connects analysis to action

The acronym is less important than the completeness of the specification. If a field is missing, record that absence before running the prompt.

1. Purpose: name one campaign decision

Bad purpose: “Improve our LinkedIn ad.”

Better purpose: “Decide whether the draft LinkedIn ad and landing-page hero are aligned closely enough to enter legal and brand review for a US B2B campaign.”

The better version identifies the assets, channel, market, stage and next gate. It does not ask the model to launch the campaign or claim that it will perform.

2. Evidence: assemble a compact source packet

Create a short packet with labelled sections:

• objective and primary conversion event;
• audience, job context and evidence for the stated problem;
• offer, price and availability;
• approved product facts;
• claim ledger showing each material claim and its source;
• channel and format constraints;
• brand rules and prohibited language;
• current creative and landing-page text;
• known objections;
• prior results, with dates and sample context, if they are legitimately comparable.

Do not place confidential customer records, personal data, unreleased financial information or protected company material into a consumer AI tool. Redact or replace sensitive examples. Your organization’s policy and the applicable product contract govern what can be processed.

3. Role boundary: assign analysis, not authority

Tell Gemini to behave as a campaign-critique analyst. It may identify mismatches, classify evidence, score the supplied materials and propose testable revisions. It may not invent customer research, approve legal claims, certify compliance, predict guaranteed results or publish anything.

This distinction is practical. A useful review can surface a problem, but accountable owners still decide whether a claim is lawful, a brand trade-off is acceptable and a campaign should launch.

4. Standards: score the campaign against observable criteria

Use a 100-point scorecard:

Criterion	Weight	Passing evidence
Audience-problem specificity	15	Named audience, context and problem supported by supplied evidence
Offer clarity	15	Reader can identify what is offered, for whom and what happens next
Claim support	20	Material claims map to sources or are labelled hypotheses
Message continuity	15	Ad promise, landing hero and call to action describe the same next step
Differentiation	10	Difference is specific and supportable, not a generic superlative
Channel fit	10	Length, format and attention pattern fit the named placement
Friction and objection handling	10	Major hesitation is addressed without hiding conditions
Safety, privacy and reviewability	5	No sensitive input, disguised evidence or autonomous approval

Set two independent gates. A campaign needs at least 80/100 overall, and claim support must score at least 16/20. A high total cannot compensate for unsupported claims.

5. Exceptions: define how uncertainty appears

Require four labels:

• Supported: directly supported by the supplied packet.
• Inference: a reasonable interpretation, not an established fact.
• Unknown: information is missing.
• Conflict: two supplied sources disagree.

Tell the model to cite the packet section for every material observation. If it cannot, it must use Unknown. This makes the output auditable and gives the human reviewer a clean research list.

6. Needed output: force a decision-ready structure

Request five outputs in order:

1. a one-sentence decision recommendation;
2. the completed scorecard with evidence references;
3. an issue log ranked by decision impact;
4. a revision brief with owner and acceptance test;
5. unanswered questions and required evidence.

For machine-readable workflows, Gemini’s official documentation recommends structured output for complex JSON schemas rather than relying on prompt wording alone. For a manual workflow, a Markdown table is easier to review.

7. Approval rule: connect the score to a real gate

Use three outcomes:

• Proceed to human review: total at least 80, claim support at least 16, and no critical conflict.
• Revise: the proposition is coherent, but one or more fixable criteria fail.
• Stop pending evidence: a material claim, audience premise, price or availability statement lacks support.

The model recommends a route. The named human owners—typically marketing, product, brand and legal where applicable—approve the next step.

Copyable Gemini persona prompt

Replace bracketed text. Keep the evidence packet after the instructions, because long-context guidance generally works better when the task clearly points back to the supplied context.

ROLE
You are a campaign-critique analyst. Evaluate only the supplied evidence.
Do not invent research, facts, benchmarks, customer quotations or legal conclusions.
Do not approve, publish or predict guaranteed campaign performance.

PURPOSE
Support this decision: [exact decision].
Campaign stage: [draft / pre-test / pre-legal / post-test].
Market and channel: [market, audience, placement].

EVIDENCE RULES
For every material observation, cite the packet section.
Label it Supported, Inference, Unknown or Conflict.
If evidence is missing, say what is needed; do not fill the gap.

SCORING
Score 0 to the listed maximum:
- audience-problem specificity 15
- offer clarity 15
- claim support 20
- message continuity 15
- differentiation 10
- channel fit 10
- friction and objections 10
- safety, privacy and reviewability 5

DECISION RULE
Proceed to human review only if total >= 80, claim support >= 16,
and there is no critical conflict. Recommend Revise for fixable failures.
Recommend Stop pending evidence when a material claim or offer fact lacks support.

OUTPUT
1. One-sentence recommendation.
2. Scorecard: criterion, score, evidence, failure.
3. Issue log: severity, observation, status label, source, decision impact.
4. Revision brief: change, rationale, owner, acceptance test.
5. Missing evidence and questions.

EVIDENCE PACKET
[Paste labelled packet here]

TASK
Apply the rules above. Be specific to this campaign. Do not give generic marketing advice.

Worked example: from generic praise to an actionable stop decision

Assume a fictional analytics product is running a LinkedIn campaign for finance managers.

Objective: book qualified demonstrations.

Audience evidence: five interviews mention that month-end variance explanations take too long. This is directional research, not a market estimate.

Approved product fact: the product imports a defined CSV format and generates a draft variance narrative for human review.

Unsupported draft claim: “Close your books 50% faster with AI.”

Ad: “Close your books 50% faster. See every variance instantly.”

Landing hero: “Automated finance intelligence for every team.”

Call to action: “Start free,” although the current offer is a booked demonstration.

A generic persona might praise the benefit and suggest a stronger call to action. PERSONA-7 should reach a different conclusion.

The 50% claim is Unknown because the packet contains no controlled measurement. “Every variance instantly” conflicts with the actual CSV import and draft-review workflow. The ad promises speed; the landing page switches to broad intelligence; the call to action describes a free trial that does not exist. The likely score is below 80, and claim support is below 16. The correct recommendation is Stop pending evidence, not “make the copy more engaging.”

The revision brief might say:

Change	Rationale	Owner	Acceptance test
Replace “50% faster” with a supported workflow statement	No speed study exists	Product marketing	Every factual phrase maps to an approved product fact
Align ad and hero around draft variance narratives	Restores message continuity	Campaign lead	A reviewer can summarize the same promise from both assets
Change CTA to “Book a workflow review”	Matches available offer	Demand generation	Click leads to the stated booking flow
Add qualification note about CSV format and human review	Prevents overstatement	Product + legal	Limitation appears before conversion

A revised ad could state: “Turn an approved variance-data export into a reviewable first-draft narrative. Book a workflow demonstration.” That sentence is less spectacular, but it is specific, supportable and aligned with the actual offer.

Add two examples before using the prompt at scale

Google’s guidance recommends few-shot examples because examples teach format and boundaries. Supply one acceptable observation and one unacceptable one.

Acceptable example: “Message continuity: 8/15. Supported. The ad promises a demonstration, while the landing CTA offers a newsletter. Source: Creative A and Landing CTA. Decision impact: conversion path is inconsistent.”

Unacceptable example: “The audience will probably love this because finance teams want AI.” This is an unsupported prediction and has no packet reference.

Use examples from your own review policy, not confidential historic campaigns. Make the format consistent with the requested output.

A five-pass human review workflow

Pass 1: evidence hygiene

Remove sensitive data, confirm that each source is current and label any historical result with its market, channel and date. Separate product facts from campaign hypotheses. A hypothesis can be tested; it should not masquerade as evidence.

Pass 2: prompt preflight

Check the seven PERSONA fields. If Purpose, Evidence or Approval rule is blank, do not run the prompt. The missing input will produce an attractive but weak answer.

Pass 3: independent model critique

Run the prompt without first telling the model what outcome you prefer. Save the model version, prompt version, input version and output. Re-running an edited prompt without version control makes comparisons unreliable.

Pass 4: human challenge

For every high-severity issue, ask: Is the observation supported? Is the source appropriate? Is the recommendation within the model’s boundary? Could following it create a new brand, legal, accessibility or privacy problem? Record accepted and rejected model suggestions.

Pass 5: controlled test

The scorecard determines readiness for human review, not market success. Use an appropriate experiment to test the revised hypothesis. Define the primary metric, guardrail metric, sample rule and stop condition before seeing results. Do not claim causality from an uncontrolled before-and-after comparison.

Diagnose a weak critique in five minutes

When output is generic, do not immediately lengthen the persona. Audit the system:

Symptom	Likely cause	Repair
Advice could apply to any campaign	Purpose too broad	Name one decision and one stage
Model invents audience needs	Evidence packet absent	Supply research or mark the premise Unknown
Long prose with no priority	Output schema absent	Require scorecard and issue log
Unsupported confidence	No uncertainty labels	Enforce Supported/Inference/Unknown/Conflict
Every campaign passes	Thresholds too weak	Add independent claim-support gate
Review is accurate but unusable	No revision owner or test	Add owner and acceptance test

The best prompt improvement is often a missing business input, not additional adjectives.

Where Gemini tools can help—and where they do not

Grounding can help retrieve current public information when that is appropriate, but retrieved text still needs source review. Code execution can help calculate a score or compare variants, but a calculation is only as valid as its inputs. Structured output can improve consistency, but a valid JSON object can still contain an unsupported claim.

Do not confuse tool use with governance. Decide which sources are allowed, how freshness is checked, who reviews the output and what happens when sources conflict. For material campaigns, preserve a claim-to-source ledger outside the chat.

Final campaign-critique checklist

Before accepting a Gemini critique, confirm:

• the decision, channel, market and campaign stage are explicit;
• the audience problem has evidence rather than an invented persona;
• every material product or performance claim has a source;
• ad, landing promise, offer and call to action are continuous;
• uncertainty appears as a label, not hidden in fluent prose;
• the scorecard has weighted criteria and independent stop conditions;
• the issue log ranks decision impact;
• every proposed revision has an owner and acceptance test;
• sensitive data were excluded or handled under approved policy;
• human reviewers retain approval responsibility;
• testing follows a predeclared metric and stop rule.

Build the capability, not just the prompt

Campaign critique sits inside a larger workflow: evidence collection, offer design, channel planning, content production, controlled testing and human governance. The Generative & Agentic AI in Marketing programme is the relevant MTF pathway for learners who want to deepen those connected capabilities. Use the template in this article as a portfolio artifact: complete it with a fictional or properly authorized campaign, show the evidence labels, record rejected suggestions and explain the final human decision.

Sources

• Google AI for Developers: Prompt design strategies
• Google AI for Developers: Long context
• Google AI for Developers: Using tools with Gemini
• NIST AI Risk Management Framework

This guide was prepared on 16 September 2026. AI products and interfaces change; verify the current official documentation and your organization’s policy before operational use.