AI prompts for sponsored articles optimized for SEO – ready-to-copy templates for Claude (Opus 4.6, Sonnet 4.6), ChatGPT (GPT-5.x), and Gemini. Each prompt enforces a natural heading structure and reduces typical AI speech patterns in English text.
TL;DR: The guide covers a base one-shot prompt, a 4-stage prompt chain, and prompts for the lead, FAQ, meta tags, style editing, and internal linking. Every prompt comes with a mechanism explanation – why each element is built the way it is. There’s also a full list of forbidden “AI-speak” phrases with alternatives, ready to drop into the constraints section.
Prompts work better with a richer knowledge base. If you’re using Claude – add skills to the project (e.g., a copywriting skill, an SEO skill – Claude has a skill for creating skills, just tell it to use one and describe what you need) or files with brand and portal guidelines. In ChatGPT, you can hook in instructions via Custom Instructions or in GPTs. AI models generate noticeably better sponsored articles when they have access to industry data – e.g., through API integrations with DataForSEO (SERP analysis, People Also Ask, search volume) or Ahrefs (competitor analysis, keyword difficulty). Without that data, the prompt relies entirely on the model’s internal knowledge, which may be outdated or generic.
What makes a simple prompt produce weak sponsored articles?
The prompt “write a sponsored article about [product]” doesn’t give the AI any of the information it needs to write a good text. The model doesn’t know the search intent, target audience, ratio of educational to promotional content, or the tone of the portal where the article will be published.
The result is text that:
- Opens with a sentence that fits 100 different articles (“There are many solutions on the market…”)
- Mentions the brand in every other sentence – or not at all
- Has no SEO-friendly heading hierarchy – no long-tail phrases in subheadings
- Uses empty descriptors with no mechanism: “innovative”, “comprehensive”, “advanced solution”
A well-built prompt forces specific behavior on the AI: content ratios, heading style, brand mention limits, list of forbidden phrases. The more context you give, the less editing you’ll do after generation.
Prompt structure – what layers it consists of and why
Every prompt for AI models (Claude, ChatGPT, Gemini) consists of layers. Below are 5 layers, each with a specific function.
Layer 1 – role
Sets the style, vocabulary, and level of detail in the response. A prompt with the role “experienced SEO copywriter specializing in sponsored articles” generates different text from a prompt without a role – even with an identical task.
Layer 2 – context
Information about the brand, audience, publication portal, keyword phrases, and article goals. This is input data the AI has no way to source on its own.
Layer 3 – task
A precise description of what the AI should generate. With format, length, heading structure, and specific sections.
Layer 4 – constraints
A list of bans and requirements. Forbidden phrases (ideally with replacements – bad phrase -> good phrase), max frequency of brand mentions, content ratios, sentence style. This layer eliminates the typical problems in generated text.
Layer 5 – output format
Specification for headings (H2/H3), formatting (bold, italic), language, and extras (meta tags, anchor text suggestions).
How prompt format differs between models
| Element | Claude (Opus / Sonnet 4.6) | ChatGPT (GPT-5.x) | Gemini |
|---|---|---|---|
| Structure | XML tags: <role>, <task>, <constraints> | Markdown headings or numbered lists | Prefixes: Input:, Instructions: |
| Instruction style | Calm, direct. No CAPS or exclamation marks – Claude 4.x “overtriggers” on aggressive language | Direct, numbered steps | Brief, with examples |
Base prompt (one-shot) – sponsored article with natural linking
Below is the main prompt for generating sponsored articles. A version for Claude (XML tags) and one for ChatGPT (markdown). Both prompts enforce the same structure and rules.
Sponsored article ratios
Claude version – one-shot
<role>
Experienced SEO copywriter specializing in sponsored
and link-building articles. You write substantively,
concretely. Tone: an expert giving advice
to a friend in the industry.
</role>
<context>
Brand: [BRAND NAME]
Industry: [INDUSTRY]
Product/service to promote: [DESCRIPTION - 2-3 sentences
covering what it does and who it's for]
Target URL (anchor link): [URL]
Anchor text: [ANCHOR PHRASE]
Publication portal: [PORTAL NAME / TYPE OF PORTAL]
Target audience: [WHO: age, role, knowledge level.
PROBLEM: what hurts them. GOAL: what they want to achieve]
Primary keyword: [PHRASE]
Supporting phrases (max 3-5): [PHRASE 1], [PHRASE 2], [PHRASE 3]
</context>
<task>
Write a sponsored article of [1500-2500] words.
Structure:
1. Lead (first paragraph): max 40 words, 2-3 sentences.
Start with a concrete problem or surprising fact.
DO NOT mention the brand in the first paragraph.
2. Main content: split into sections with H2 to H5 headings where needed.
Headings - only the first letter of the first word in each sentence
capitalized (no title case). (e.g., Example heading. And next. -> correct;
Example Heading. And Next -> wrong)
Phrase subheadings as questions or benefits,
not as labels (not "Types of X" but "Which X
works best for [Problem]").
Each section should vary in length, paragraph count, and sentence count.
Where it fits, add a bulleted list (but not everywhere) and use formatting
like bold or italic.
Make sure each section has a different visual structure (bad: paragraph-paragraph-paragraph
in every section | Good (example): paragraph, paragraph-paragraph-list-paragraph,
paragraph-paragraph, paragraph, paragraph-paragraph-paragraph, paragraph-list)
3. Brand mention: max 1 per 300-400 words and no more than 3 in the article.
The brand appears as a natural solution to the problem
described in the section. Not as advertising.
4. Anchor text link: weave naturally into the content,
max 2 times in the article. The keyword phrase should appear in the same paragraph
or ideally the same sentence as the link (but not necessarily in the link itself).
5. FAQ section: 5-6 questions phrased like Google search queries.
Answers: 2-4 sentences. First sentence of the answer = direct answer
to the question. Brand mentioned in max 2 answers.
6. Conclusion: max 3 sentences. CTA based on benefit,
not "check out our offer". The conclusion shouldn't repeat content from
the article - it should complement it or serve as a CTA.
</task>
<constraints>
FORBIDDEN PHRASES (never use):
- "In today's world", "On the market", "In the era of", "In times when"
- "crucial", "innovative", "comprehensive", "wide range"
- "perfect solution", "not only... but also"
- "it's worth noting", "interestingly", "moreover", "furthermore"
- "In this article, you'll learn", "Below you'll find"
STYLE:
- Active voice instead of passive ("we built" not
"the product was built")
- Every sentence with a keyword must carry real information
- Mix sentence lengths: short (4-8 words), medium (10-15),
longer (up to 25 words). No two consecutive sentences of
the same length.
- Paragraphs: max 4 sentences, ~40-60 words; vary length
and number of paragraphs per section
- Formatting: bold for important concepts (2-3 per section),
italic for quotes and foreign terms, bold+italic for the most important phrases
- If a sentence repeats the previous sentence's idea in different
words - delete it
DATA:
- Don't make up statistics, prices, or research findings.
No data = write nothing, ask the user, or insert a placeholder [XX%] or [XX USD]
- Don't use em dash (—). Use "-"
</constraints>
<output_format>
Format: text with markdown headings (## for H2, ### for H3)
Language: English
At the end, add:
- Meta title (max 60 characters, keyword phrase near the start,
only first letter capitalized)
- Meta description (max 155 characters, keyword phrase
in the first 60 characters)
- 3 anchor text suggestions for internal portal links
</output_format>ChatGPT / GPT-5 version – one-shot
You're an experienced SEO copywriter. You specialize
in sponsored and link-building articles. You write
substantively, concretely. Tone: an expert
giving advice to a friend in the industry.
## Inputs
- Brand: [BRAND NAME]
- Industry: [INDUSTRY]
- Product/service: [DESCRIPTION - 2-3 sentences: what it does, who it's for]
- Target URL: [URL]
- Anchor text: [ANCHOR PHRASE]
- Portal: [PORTAL NAME]
- Target audience: [WHO, PROBLEM, GOAL]
- Primary keyword: [PHRASE]
- Supporting phrases: [PHRASE 1], [PHRASE 2], [PHRASE 3]
## Task
Write a sponsored article, 1500-2500 words.
Structure:
1. Lead: max 40 words. Problem or surprising fact (real, not fabricated).
No brand.
2. Content: sections with H2 to H5 headings. Subheadings as questions
or benefits (not labels). Headings - only the first letter of the first
word in each sentence capitalized (no title case). (e.g., Example heading.
And next. -> correct; Example Heading. And Next -> wrong) Each section should vary
in length and number of paragraphs or sentences. Where it fits, add a bulleted
list (but not everywhere) and use formatting like bold or italic. Make sure each
section has a different visual structure (bad: paragraph-paragraph-paragraph in every
section | Good (example): paragraph, paragraph-paragraph-list-paragraph,
paragraph-paragraph, paragraph, paragraph-paragraph-paragraph, paragraph-list)
3. Brand: max 1 mention per 300-400 words and max 3 per article. Naturally as
a solution to the problem described in the section.
4. Link: anchor text woven naturally, max 2 times. The keyword phrase should appear
in the same paragraph or ideally the same sentence as the link (but not necessarily
in the link itself).
5. FAQ: 5-6 questions like Google queries. Answers 2-4 sentences.
First sentence = direct answer.
Brand in max 2 answers.
6. Conclusion: max 3 sentences, CTA as a benefit.
## Forbidden phrases
Never use: "in today's world", "on the market",
"in the era of", "in times when", "crucial", "innovative", "comprehensive",
"wide range", "perfect solution", "not only... but also",
"it's worth noting", "interestingly", "in this article, you'll learn"
## Style requirements
- Active voice, not passive
- Every sentence with a keyword = real information
- Mix sentence lengths (short 4-8 words, medium 10-15,
longer up to 25), paragraph lengths, and paragraph counts per section
- Paragraphs: max 4 sentences
- Bold for important concepts, italic for quotes and foreign terms,
bold+italic for the most important phrases
- Don't repeat the same idea in different words
- Don't make up data. None = don't write, ask the user, or insert placeholder [XX%]
- Don't use em dash (—). Use "-"
## Output format
Markdown with headings. At the end:
- Meta title (max 60 characters, only first letter capitalized)
- Meta description (max 155 characters)
- 3 anchor text suggestionsSave this prompt as a template. Before each use, you only fill in the fields in square brackets [...]. One prompt template removes the need to write from scratch every time.
Why should the prompt include a “forbidden phrases” section? (AI-overused phrases with replacements)
AI models have statistical preferences – certain phrases show up in their responses far more often than in human-written text. Without an explicit ban, those phrases will end up in your article.
An explicit ban list in the constraints section makes the model actively avoid those constructions. Without the list, the model has no reason to skip them – they appear often in its training data.
The list below is split by category. Every phrase has an alternative – a replacement that carries the same information without AI-speak (language typical of AI models). Copy the whole list into the constraints section of your prompt.
Full list of forbidden phrases with alternatives (AI-overused phrase list)
FORBIDDEN PHRASES - NEVER USE (with alternatives):
GENERIC TIME-BASED OPENERS:
❌ "in today's world" → ✅ skip, start with a concrete fact
❌ "in this day and age" → ✅ skip, start with the problem
❌ "in times when..." → ✅ skip or give a specific date/period
❌ "in a rapidly changing..." → ✅ skip
❌ "in the era of..." → ✅ skip
❌ "in the world of..." → ✅ skip
❌ "in the digital age" → ✅ skip
❌ "over the years" → ✅ give a specific period: "over the
past 3 years", "between 2022 and 2025"
EMPTY ADJECTIVES AND DESCRIPTORS:
❌ "crucial" / "key" / "the key to" → ✅ "important",
"useful" or just explain WHY something matters
❌ "the foundation" / "fundamental" → ✅ describe the specific
function or role
❌ "essential" / "indispensable" / "must-have" → ✅ "useful",
"needed" or explain what happens without it
❌ "innovative" → ✅ describe what exactly is new and how it works
❌ "comprehensive" → ✅ "complete" or list 3-4 specific elements
❌ "wide range" → ✅ list specific options
❌ "advanced solution" → ✅ describe how it works
❌ "perfect solution" → ✅ "proven solution"
or describe in what scenario it fits
❌ "really works" → ✅ give data, case study, mechanism
❌ "changed our approach" → ✅ describe the specific change
AI COMPARATIVE CONSTRUCTIONS:
❌ "not only..., but also..." → ✅ two separate, short sentences
❌ "it's not just X, it's Y" → ✅ two separate sentences
❌ "it's not just aesthetics" → ✅ describe both qualities separately
❌ "blending functionality with aesthetics" → ✅ describe
functionality and looks in separate sentences
❌ "aesthetics with functionality" → ✅ same as above
FILLER TRANSITIONS:
❌ "it's worth noting" → ✅ skip, state the fact directly
❌ "it's worth highlighting" → ✅ skip
❌ "interestingly" → ✅ skip, give the fact
❌ "what's more" → ✅ skip
❌ "moreover" / "furthermore" → ✅ skip
❌ "of course" → ✅ skip
❌ "certainly" / "without a doubt" → ✅ skip or give proof
❌ "in conclusion" → ✅ "in short", "to wrap up"
or just write the closing sentence without an intro
❌ "in fact" → ✅ skip
❌ "essentially" → ✅ skip
❌ "in practice" → ✅ give a concrete example instead of announcing one
SELF-REFERENTIAL PHRASES (the text talking about itself):
❌ "In this article, you'll learn..." → ✅ just write the content
❌ "Below you'll find..." → ✅ just write the content
❌ "This guide covers..." → ✅ just write the content
❌ "Tips that really work" → ✅ describe the tips,
don't claim they work
❌ "Concrete, proven information" → ✅ give the information
❌ "Practical tips" → ✅ give the tips
MARKETING JARGON:
❌ "gets the job done" → ✅ describe what it does and the effect
❌ "the unsung hero" → ✅ describe the function directly
❌ "attention to detail" → ✅ list the specific details
❌ "real" / "really" / "real-world" → ✅ "concrete",
"actual", "measurable"
❌ "groundbreaking advance" → ✅ describe what changed,
with numbers or before/after comparison
❌ "the power of transformation" → ✅ describe the change with specifics
❌ "but that's not all" → ✅ skip, move to the next
point
❌ "significantly" → ✅ give a number or scale: "by 30%", "doubled"You can extend the list. After generating 3-5 articles, you’ll notice which phrases the model repeats despite the ban. Add them to the list. Typical “slip-throughs” for English Claude: “certainly”, “without a doubt”, “the foundation”. For ChatGPT: “it’s important to”, “keep in mind”.
Pro-tip for Claude: If you have conversation search and memory enabled, you can ask Claude to search across conversations from a given project or about a given topic and compile a list of the corrections, alternatives, or issues you’ve raised most often, which you can then add to your knowledge base or base prompt 😉
Lead prompt – three formulas for the first paragraph
The lead is the first 40 words of the article. By default, AI generates openings that are generic and could fit any topic. The prompt below forces 3 specific formulas.
<task>
Write 3 lead variants for an article on:
"[ARTICLE TOPIC]"
Target audience: [DESCRIPTION]
Each variant: max 40 words, 2-3 sentences.
The first sentence should contain the bolded and italicized keyword: [phrase]
Variant 1 formula (Known - Gap - Promise):
- Sentence 1: a fact the reader takes as true
- Sentence 2: information that challenges or expands that fact
- Sentence 3: a promise to fill the knowledge gap
Variant 2 formula (Problem - Consequence - Solution):
- Sentence 1: a specific problem of the audience (with a number
or detail)
- Sentence 2: the consequence of that problem
- Sentence 3: a hint at the solution
Variant 3 formula (Fact - Context - Question):
- Sentence 1: a specific fact or statistic
- Sentence 2: why this matters to the reader
- Sentence 3: a question that pulls the reader in
</task>
<constraints>
- No variant starts with "In today's world",
"On the market", "Did you know", "Imagine", "In a digital age",
"In the era of", "Over the years", etc.
- No variant mentions the brand
- First sentence of each variant: max 10 words
- Don't promise more than the article delivers
</constraints>An editorial tip from practice: after generating 3 variants, delete the first sentence of each. The second sentence is almost always the better opener – it gets to the point faster.
FAQ prompt optimized for featured snippets
An FAQ section is one of the simpler tools for getting visibility in Google – both in featured snippets and in the People Also Ask section. But without instructions, AI generates generic questions (“Is X good?”) and answers without “substance” (generic, recycled).
The difference between bad and good FAQ:
| Element | Bad FAQ | Good FAQ |
|---|---|---|
| Question | “Is a memory foam mattress good?” | “How many years does a memory foam mattress last?” |
| Answer | “Yes, memory foam mattresses are an excellent choice…” | “7-10 years on average. The lifespan drops to 4-5 years at humidity above 60%…” |
| Intent | None – nobody types this into Google | Specific – informational, with numbers |
<role>
SEO specialist creating FAQ sections for Google featured snippets
and People Also Ask.
</role>
<task>
Write an FAQ section for an article on: "[TOPIC]"
Industry: [INDUSTRY]
Generate 6 questions and answers.
Question types:
- 3 informational (starting with "How", "What", "Why")
- 2 comparative ("Is X better than Y", "Difference between X and Y")
- 1 about cost or time ("How much does", "How long does")
Answer rules:
- Sentence 1 = direct answer (not a restatement of the question)
- Sentence 2-3 = context, mechanism, or example with a number
- Optional sentence 4 = caveat or condition
- Brand naturally in max 2 answers as an example
of a solution
</task>
<constraints>
- Don't write "It depends on many factors" - if the answer
depends on context, give 2-3 scenarios with numbers
- Don't start the answer by repeating the question
- Don't use: "It's worth noting", "Interestingly", "Furthermore", etc.
- Questions must sound like something someone types into Google,
not like textbook topics
- Headings: only the first letter of the first word in each sentence
capitalized (no title case). (e.g., Example heading. And next. -> correct;
Example Heading. And Next -> wrong)
</constraints>Prompt for meta title and meta description
Meta tags affect CTR in search results. Two rules the prompt has to enforce:
- Meta title – keyword near the start, max 60 characters, only first letter capitalized (no title case)
- Meta description – keyword in the first 60 characters (Google often truncates longer descriptions), max 155 characters
<task>
Generate 3 variants each of meta title and meta description
for an article on: "[TOPIC]"
Primary keyword: [PHRASE]
Search intent: [informational / transactional / comparative]
Meta title (3 variants):
- Max 60 characters with spaces. Give the character count.
- Keyword near the start
- Only the first letter of the first word in each sentence capitalized
(no title case). (e.g., Example heading. And next. -> correct;
Example Heading. And Next -> wrong)
- Variant 1: with a number
- Variant 2: with a question
- Variant 3: with a benefit
- Variant 4: clickbait-style
Meta description (3 variants):
- Max 155 characters with spaces. Give the character count.
- Keyword in the first 60 characters
- CTA or a promise of specific value
- Don't use "Learn more" or "Check out our article", etc.
</task>Modifying the base prompt for specific industries
The base prompt needs adjustment depending on the industry. Below are specific snippets to add to the <constraints> section for industries where Google applies extra quality requirements (known as YMYL – Your Money or Your Life).
“For example, our systems put even greater weight on content aligned with E-E-A-T for topics that could significantly impact the health, financial stability, or safety of people, or the welfare or wellbeing of society. We call these ‘Your Money or Your Life’ (YMYL) topics.” – Google
Source: https://developers.google.com/search/docs/fundamentals/creating-helpful-content#eat
Medicine and health
Add to <constraints>:
- Don't diagnose or suggest treatment
- End every health-related tip with: "Consult a doctor
before making any decisions"
- Provide active substance names alongside brand names
- Don't fabricate, always do online research and fact-checking. If you don't have
information you can back up with a link to an authoritative source, don't write it
(don't fabricate links; when citing a link, verify the article's content actually
contains the information; if it doesn't - don't write about it or include the link)
- The article must meet Google E-E-A-T standards for YMYL
content - signals of author experience and expertiseFinance and insurance
Add to <constraints>:
- Don't present rates of return as guaranteed
- Mark every numerical example: "illustrative example"
- Add a disclaimer about investment risk
- Don't use phrasing that suggests certainty of profitE-commerce / physical products
Add to <constraints>:
- Compare technical parameters instead of writing "best
on the market" (Don't make up parameters - if missing from the user's prompt,
ask or look it up online by the specific model name)
- Provide concrete specs: dimensions, materials, weight
- Tone: shopping guide, not a review
- Don't use superlatives without supporting data (Don't fabricate data,
always rely on the user's prompt or online research)Hospitality / HoReCa
Add to <constraints>:
- Express costs per portion / guest / month
- Use restaurateur language (covers, table turnover,
food cost), not corporate-speakGoogle E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) is a quality framework that isn’t a direct ranking factor, but does influence how Google’s algorithms assess content. For YMYL industries (health, finance), E-E-A-T expectations are higher. The prompt should enforce expertise signals: specific examples, source data, caveats, and limitations (but based on real data – ideally provided by the user).
Prompt chain – generating an article in 4 stages
One prompt for an entire 2000-word article gives worse results than a series of shorter prompts. Reason: AI loses coherence after 800-1000 words – it forgets the instructions from the start of the prompt. In a prompt chain, each stage has one task and a short output.
Stage 1: research and structure
<role>
SEO analyst.
</role>
<task>
Keyword: "[PHRASE]"
Industry: [INDUSTRY]
Target audience: [DESCRIPTION]
Do this:
1. Identify search intent: informational, transactional,
comparative, navigational
2. Suggest 6-8 questions people actually type into Google
for this phrase (People Also Ask style)
3. Suggest article structure: H2 to H5 headings
(phrased as questions or benefits, not labels)
4. List 3-5 long-tail phrases related to the main keyword
5. Suggest an angle for the article that sets it apart from
standard content on this topic
Format: numbered list, short description for each point.
</task>Stage 2: outline with brand placement
<task>
Based on the analysis from Stage 1:
[PASTE OUTPUT FROM STAGE 1]
Create a sponsored article outline.
Brand: [BRAND]
Product: [PRODUCT DESCRIPTION]
For each H2-H5 heading, write:
- Section topic (1 sentence)
- Reader value (1 sentence)
- Brand mention in this section: YES/NO.
If YES: in what role (problem solver /
example / data source)
- Length: [50/100/150/200/250 words]
Rule: brand in max 3 sections of the entire article.
Not in the first paragraph, not in the first section.
</task>Stage 3: section-by-section generation – importantly, generate within the same conversation
<task>
Write the section "[SECTION HEADING]" of the sponsored article.
Article outline:
[PASTE OUTLINE FROM STAGE 2]
This section:
- Topic: [FROM OUTLINE]
- Value: [FROM OUTLINE]
- Brand: [YES/NO - if YES, role]
- Maximum length: [XXX words]
Requirements:
- Valuable/Helpful/Interesting information, focused on reader value
- Subheading as a question or benefit
- Max 4 sentences per paragraph
- If brand: weave in as a solution to the section's problem
- Forbidden phrases:
FORBIDDEN PHRASES - NEVER USE (with alternatives):
GENERIC TIME-BASED OPENERS:
❌ "in today's world" → ✅ skip, start with a concrete fact
❌ "in this day and age" → ✅ skip, start with the problem
❌ "in times when..." → ✅ skip or give a specific date/period
❌ "in a rapidly changing..." → ✅ skip
❌ "in the era of..." → ✅ skip
❌ "in the world of..." → ✅ skip
❌ "in the digital age" → ✅ skip
❌ "over the years" → ✅ give a specific period: "over the
past 3 years", "between 2022 and 2025"
EMPTY ADJECTIVES AND DESCRIPTORS:
❌ "crucial" / "key" / "the key to" → ✅ "important",
"useful" or just explain WHY something matters
❌ "the foundation" / "fundamental" → ✅ describe the specific
function or role
❌ "essential" / "indispensable" / "must-have" → ✅ "useful",
"needed" or explain what happens without it
❌ "innovative" → ✅ describe what exactly is new and how it works
❌ "comprehensive" → ✅ "complete" or list 3-4 specific elements
❌ "wide range" → ✅ list specific options
❌ "advanced solution" → ✅ describe how it works
❌ "perfect solution" → ✅ "proven solution"
or describe in what scenario it fits
❌ "really works" → ✅ give data, case study, mechanism
❌ "changed our approach" → ✅ describe the specific change
AI COMPARATIVE CONSTRUCTIONS:
❌ "not only..., but also..." → ✅ two separate, short sentences
❌ "it's not just X, it's Y" → ✅ two separate sentences
❌ "it's not just aesthetics" → ✅ describe both qualities separately
❌ "blending functionality with aesthetics" → ✅ describe
functionality and looks in separate sentences
❌ "aesthetics with functionality" → ✅ same as above
FILLER TRANSITIONS:
❌ "it's worth noting" → ✅ skip, state the fact directly
❌ "it's worth highlighting" → ✅ skip
❌ "interestingly" → ✅ skip, give the fact
❌ "what's more" → ✅ skip
❌ "moreover" / "furthermore" → ✅ skip
❌ "of course" → ✅ skip
❌ "certainly" / "without a doubt" → ✅ skip or give proof
❌ "in conclusion" → ✅ "in short", "to wrap up"
or just write the closing sentence without an intro
❌ "in fact" → ✅ skip
❌ "essentially" → ✅ skip
❌ "in practice" → ✅ give a concrete example instead of announcing one
SELF-REFERENTIAL PHRASES (the text talking about itself):
❌ "In this article, you'll learn..." → ✅ just write the content
❌ "Below you'll find..." → ✅ just write the content
❌ "This guide covers..." → ✅ just write the content
❌ "Tips that really work" → ✅ describe the tips,
don't claim they work
❌ "Concrete, proven information" → ✅ give the information
❌ "Practical tips" → ✅ give the tips
MARKETING JARGON:
❌ "gets the job done" → ✅ describe what it does and the effect
❌ "the unsung hero" → ✅ describe the function directly
❌ "attention to detail" → ✅ list the specific details
❌ "real" / "really" / "real-world" → ✅ "concrete",
"actual", "measurable"
❌ "groundbreaking advance" → ✅ describe what changed,
with numbers or before/after comparison
❌ "the power of transformation" → ✅ describe the change with specifics
❌ "but that's not all" → ✅ skip, move to the next
point
❌ "significantly" → ✅ give a number or scale: "by 30%", "doubled"
</task>You repeat this stage for every section. It sounds slower than one prompt for the whole article, but the result is clearly better – especially in articles over 1500 words.
Stage 4: audit and meta tags
<task>
Analyze the finished article:
[PASTE FULL ARTICLE]
1. Check (and fix on the spot):
- How many times does the brand [BRAND] appear? In which sections?
- Are there sentences with forbidden phrases?
- Are there fabricated statistics without a placeholder? (Do fact-checking)
- Does each H2-H5 heading work as a standalone sentence
(a reader scanning the headings can follow the article)?
- Do two consecutive sentences/paragraphs/sections have similar length
or similar number of elements like sections or list bullets?
2. Meta title (3 variants, max 60 characters, keyword "[PHRASE]"
near the start, only first letter capitalized)
3. Meta description (3 variants, max 155 characters, keyword
in the first 60 characters)
4. Suggested fixes (if the audit found problems)
</task>Stage 4 serves as verification. Most people skip it. Yet that’s exactly the stage that catches: too many brand mentions, forbidden phrases that “slipped through”, and fabricated statistics.
Prompt for editing AI style in finished text
Even text from a good prompt can have typical AI traits: uniform sentence length, no tension, predictable section structure. The prompt below corrects those patterns in finished text.
<role>
Text editor. Task: improve the naturalness of the style,
preserving 100% of the substantive content.
</role>
<task>
Text to edit:
[PASTE TEXT]
Do this:
1. Vary sentence lengths: short (4-8 words), medium (10-15),
longer (up to 25). No two consecutive sentences of similar length.
Also vary paragraph length, and the number of paragraphs, lists, and list
items per section.
2. Find sentences that repeat the same idea in different words -
keep only the better version. Don't repeat the same information across
different sections.
3. Switch passive voice to active where possible.
4. Remove redundant adverbs: "very", "extremely",
"particularly", "really", "significantly".
5. Add 1 ultra-short paragraph (1 sentence) every 400-500 words
to break the visual rhythm.
6. Check that each section has a different visual structure
(bad: paragraph-paragraph-paragraph in every section | Good (example):
paragraph, paragraph-paragraph-list-paragraph, paragraph-paragraph,
paragraph, paragraph-paragraph-paragraph, paragraph-list).
DO NOT CHANGE: data, numbers, proper nouns, anchor texts,
links, headings.
</task>Prompt for internal linking
A sponsored article without an internal linking strategy is a wasted opportunity – both for the portal and the brand. Internal links strengthen the site’s topical structure (topic clusters) and pass SEO value between pages. You can paste the portal URL with sitemap info, attach the sitemap in the message – or attach a WordPress posts export.
<article>
[ARTICLE]
</article>
<task>
Portal: [PORTAL] (check the sitemap)
Portal industry: [INDUSTRY]
Based on the article content and the portal's sitemap entries, suggest:
1. 3-5 outbound links FROM the article (to portal content):
- Anchor text (2-4 words, natural in sentence context)
- Topic and link of the target article from sitemap/export
- Article section to place the link in
2. 3 inbound links TO the article (from other portal content):
- Anchor text
- Topic and link of the source article
Rules:
- Anchor text reads naturally in the sentence (not "click here",
not "more information")
- The anchor contains a phrase related to the target article topic
- Max 1 link per 200-300 words of text
- Don't link from the lead or from the FAQ
</task>Short prompts for micro-tasks
Not every task needs an elaborate prompt. Below are one-sentence instructions for common micro-tasks.
Turning a label-style heading into a question:
Turn this heading into a question someone would actually type into Google: "[HEADING]"Shortening a paragraph:
Shorten to max 40 words while keeping the main idea: "[PARAGRAPH]". Remove any repeated information.Naturally weaving in an anchor text:
Weave the phrase "[ANCHOR]" as a link into the sentence so it reads
like a natural part of the text: "[SENTENCE]"Image alt text with a keyword:
Alt text (max 125 characters) for an image: [IMAGE DESCRIPTION].
Include the phrase: [PHRASE]. The alt describes what's in the image,
it's not advertising.Turning a list into a paragraph:
Rewrite as flowing prose, max 60 words: [LIST]7 mistakes that hurt the quality of generated articles
1. No role in the prompt
A prompt without a role generates “general” text – no specialization, no industry vocabulary, no appropriate level of detail. Adding the role “SEO copywriter with experience in the restaurant industry” changes the output from the first paragraph: industry terminology shows up, concrete scenarios appear, and the tone fits.
2. Target audience described as “anyone interested”
The more general the audience, the more generic the text. Instead of “people interested in the topic”, write: “restaurant owners in the US, 30-50 years old, looking for ways to lower operating costs, technical knowledge: basic”.
3. No forbidden phrases list
Without an explicit ban, AI will use its statistically preferred phrases. A constraints section with phrases to exclude is required in any prompt generating English text.
4. One prompt for the whole article instead of a chain
AI generates coherent text for about 800-1000 words. Beyond that threshold, it loses context – instructions from the start of the prompt have less and less effect. A prompt chain (4 stages) solves this.
5. No format specification
Without output_format, AI decides on formatting itself. Sometimes it adds headings, sometimes not. Sometimes it uses lists, sometimes paragraphs. Spell out the format: “## for H2, ### for H3, bold for concepts, max 4 sentences per paragraph”.
6. Too many keywords
A prompt with 15 phrases produces text that Google’s Helpful Content Update algorithm flags as keyword stuffing. Limit: 1 primary keyword + max 5 supporting phrases. The rest should emerge naturally from the topic.
7. No instructions about data
AI models generate statistics, prices, and references to studies that don’t exist – but sound credible. The instruction “don’t make up data, use a placeholder [XX%]” lowers the risk. It doesn’t eliminate it – every number in generated text needs manual verification.
FAQ
Does Google penalize AI-written sponsored articles?
Not for using AI itself. Google penalizes low-quality content – regardless of who wrote it. An AI article that’s substantive, answers real questions, and was reviewed by a human is treated the same as a hand-written one. Problems start when the text is generic, duplicates existing content without added value, or contains hallucinations (fabricated facts).
Which AI model writes sponsored articles best in English?
It depends on budget and process. Claude Opus 4.6 best handles complex instructions in XML tags and generates coherent longer texts. ChatGPT (GPT-5.x) is faster for iterations and multi-turn edits. Gemini handles human-sounding English well. So it’s worth writing the article in, e.g., Claude, then humanizing it in Gemini.
Is a prompt chain better than a single prompt?
Yes, measurably. A single prompt for a 2000-word article loses coherence after 800-1000 words – the model “forgets” the initial instructions. A prompt chain gives you: better coherence, more accurate execution of instructions, and the ability to fix one section without regenerating the whole thing.
Prompts tested on Claude Opus 4.6, Claude Sonnet 4.6, and ChatGPT (GPT-5.x). XML formats are optimal for Claude, markdown versions for ChatGPT and Gemini. As of March 2026.
