How to Vibe Your SEO: The Prompt-Driven Process

Vibing your SEO means running your search work as a repeatable, prompt-driven loop — a fixed sequence you feed to ChatGPT or Claude, with a human checking every point where the machine can lie to you — and this post hands you the exact prompts for each step. It’s not a one-time audit; it’s a process you run for every page, then again when Google shows you what you missed.

Copy the prompts below, swap the [BRACKETS] for your details, paste. This isn’t a rundown of what to check on a finished site — that’s a separate guide, linked where it belongs. Here’s the operating manual: the order you work in, and the prompt for each move.

What does it mean to vibe your SEO?

You know vibe coding: you describe the site and the AI writes it. Vibing your SEO aims the same move at the work after the build — research, writing, tags, schema, the monthly check of what is and isn’t ranking. The build is a one-shot job; SEO is a loop that never ends, which makes it the better fit for prompts you run on repeat. And yes, AI-built sites rank fine once this is done — see our post on whether AI-built websites rank on Google.

The catch, and the reason a human stays in the loop: AI invents things. Ask for search volumes and it hands you confident numbers it doesn’t have; ask for your prices and it fills the blank with plausible fiction — the same hallucination habit that ships fake plugins into vibe coded sites. So the rule under every step is the same: the machine drafts, you supply the facts and check them.

Here’s the loop — nine steps, each with its prompt. Run them in order the first time; after that you mostly live in steps one, three, and nine.

Find the queries people actually type

SEO starts with words — your customer’s words, which are almost never the ones you’d use for your own business. This prompt pulls real queries and sorts them by intent, so you see who’s researching and who’s ready to pay. It can’t invent search volumes, so you’ll check the list against Google autocomplete, which is free.

You are an SEO strategist. My business is [WHAT YOU DO] in [CITY OR REGION],
selling [PRODUCT/SERVICE] to [WHO MY CUSTOMER IS].

List 40 searches a real person types into Google, or asks an AI assistant,
when they have the problem I solve. For each: the query in the customer's
plain words, and the intent (informational, commercial, transactional, or local).

Sort from "just researching" to "ready to pay." Do NOT invent search volumes —
you don't have that data. Flag anything you're guessing so I can check it
against Google autocomplete myself.

Cluster the queries into pages

A list of 40 queries isn’t a plan until you group it so each page answers one thing. This is the most common vibe coded mistake — one bloated Services page chasing nine intents and ranking for none. The prompt turns the list into a simple site map, one intent per URL.

Here is a list of queries with their intent: [PASTE THE LIST FROM STEP 1].

Cluster them into website pages. Rules:
- one intent per page: if two queries want the same answer, they share a page
- a "ready to buy" query becomes a service or product page, not a blog post
- a "still learning" query becomes a blog post or guide

Return a site map table: page title, URL slug, main query, 3-5 related queries,
page type. Keep slugs short, lowercase, and free of dates or prices.

Write an answer-first brief

Don’t draft straight from a keyword. Write a brief first, because that’s where you set the angle and list the facts only you know. The token that matters is [NEED FROM YOU]: it makes the model ask for your real numbers instead of quietly inventing them.

Write a content brief for a page targeting "[YOUR MAIN QUERY]." My reader is
[WHO], and the one thing they want to know is [THE QUESTION, IN ONE SENTENCE].

Include:
- the one-sentence answer for the first two lines of the page (answer-first)
- an H2 outline, each heading a question the reader actually asks
- the concrete facts the page needs (prices, timelines, specifics) — list what
  you need from me, don't invent them
- 3-5 internal pages to link to, and why
- 5 FAQ questions real people ask

Mark every spot where you need a fact from me with [NEED FROM YOU].

Draft it without sounding like a robot

Now the model writes, and the voice guardrails earn their keep. AI prose defaults to a corporate-brochure hum that makes a reader’s eyes glaze. Ban it out loud, and repeat the no-invented-facts rule, because drafting is where a stray statistic sneaks in.

Draft the page from this brief: [PASTE BRIEF].

Voice rules — follow them exactly:
- one expert talking to one customer; first person, plain American English
- short sentences, concrete over abstract, specifics not adjectives
- answer the question in the first two sentences
- banned words: unleash, elevate, seamless, robust, leverage, dive in,
  "in today's digital landscape," "whether you're a X or a Y"
- no stat, price, or date I didn't give you in the brief. If a claim needs a
  number I didn't supply, write [CHECK: ...] instead of guessing

About [WORD COUNT] words. Use the brief's H2s verbatim.

Nail the on-page basics

The body’s written; now dress it for search. Title, meta description, headings, internal links, canonical — the elements that tell Google what the page is and win the click. This prompt hands them over without touching your copy. The full finished-site list — sitemaps, robots.txt, seventeen items — is in the on-page checklist.

Here is my finished page copy: [PASTE COPY]. Main query: "[YOUR MAIN QUERY]."

Give me the on-page elements — don't touch the body:
- a title tag under 60 characters, keyword near the front
- a meta description, 150-158 characters, written like an ad for the page
- a check of my H2s: each one unique, each using words a searcher would type?
- 3-5 internal links to add, with exact keyword anchor text (never "click here")
- one self-referencing canonical URL

Generate the schema, then validate it

Structured data hands a machine your facts in a form it can’t misread — an article, an FAQ, a local business. One prompt outputs all three as JSON-LD. Two hard rules: mark up only what’s visible on the page, and never let the model invent an address, phone, or rating to fill a blank. Then validate, because bad schema earns a penalty, not a rich result.

Generate JSON-LD for this page as three separate
<script type="application/ld+json"> blocks:

1. Article — headline, author, datePublished, publisher
2. FAQPage — using ONLY these visible questions and answers: [PASTE FAQ PAIRS]
3. LocalBusiness — name [BUSINESS NAME], address [FULL ADDRESS],
   phone [PHONE], url [URL], areas served [CITIES]

Describe only what is visible on the page. Use my real details above — do not
invent an address, phone, rating, or review. Leave any field blank rather than
filling it with a placeholder.

Run the output through Google’s Rich Results Test or a schema validator before it’s live.

Add the AI-answer layer

ChatGPT, Perplexity, and Google’s AI Overviews answer plenty of questions and quote pages that state a complete fact in one liftable sentence. Quick pass, not a rebuild: make the opening answer the question on its own, and plant a few standalone facts with your name on them. The deep version — entity consistency, crawler access, the schema that helps — is its own job, covered in getting cited by AI.

Review this page for AI-answer readiness: [PASTE COPY]. Tell me:

1. Does the first paragraph answer the query completely enough that an AI could
   quote it alone? If not, rewrite those two sentences.
2. Pull 3 sentences that state a complete fact on their own — a price, a
   timeline, a definite claim a chatbot can lift with my name on it. If there
   aren't three, tell me what facts to add.

Don't add anything I didn't say. Consistency beats cleverness here.

Ship it fast and accessible

A page nobody can load doesn’t rank or get cited, so run one speed-and-access pass before publish — speed is a live ranking factor, spelled out in the Core Web Vitals targets in the checklist.

Here is my page HTML: [PASTE HTML]. List the fastest speed and accessibility
wins, in priority order:
- images missing width/height, or missing alt text
- more than two font families loaded
- render-blocking scripts the page doesn't need
- body text under 16px, or low contrast against its background
- headings that skip a level (an H2 straight to an H4)

Give me the fix for each, shortest first. Don't rewrite the page — just the list.

Measure, then run the loop again

The step everyone skips. Six to eight weeks after a page is live, Search Console shows the queries you’re almost ranking for — the ones on page two, one edit from page one. Feed those back in as new sections or pages, and the loop restarts with better inputs.

Here is a Search Console export for one page: [PASTE QUERIES, CLICKS,
IMPRESSIONS, AND AVERAGE POSITION].

Find me the opportunities:
- queries in positions 5-15 (close, not there) — what section or sentence would
  push each one up?
- queries with impressions but almost no clicks — title or meta problem, and how
  would you rewrite it?
- queries I rank for but never targeted — is one worth its own page?

Base every suggestion on the numbers I pasted. Don't assume traffic I didn't show you.

Frequently asked questions

Can AI do my SEO for me?

No, not the whole of it, and anyone selling hands-off “AI SEO” is selling you trouble. AI is great at the labor — drafting, sorting queries, generating tags and schema — and bad at the judgment: what’s true, what your prices are, whether a page is any good. Let it type; you keep the facts and the final read.

Will AI-written SEO content get penalized by Google?

Not for being AI-written. Google’s guidance targets quality, not authorship — it demotes unhelpful content no matter what produced it. The real risk is the filler an unguided model spits out: generic paragraphs, invented stats, thin pages. The guardrails in step four exist to keep you clear of that line.

How is this different from your SEO checklist?

The checklist is what you verify on a finished site — titles, sitemap, canonicals, seventeen items you tick off once. This is the repeatable process that produces those pages, a prompt per step. Use this loop to build and grow; use the on-page checklist to inspect what you built. Different tools, same shop.

Which AI tool is best for this?

Any strong general model runs this loop — the prompts are model-agnostic on purpose. We reach for Claude when the writing has to sound human and ChatGPT when we want it to browse and pressure-test a claim, but the prompt matters more than the logo on it. Whatever you use, verify the output the same way.

How often should I run this process?

Run all nine steps once per new page. Then run the short version — step one and step nine — monthly: pull Search Console, find the queries stuck on page two, feed them back as new sections or pages. SEO isn’t a launch you finish; it’s a subscription you pay in attention.

Run the loop, or hand it off

That’s the process: nine steps, nine prompts, one human keeping the machine honest. Save them, run them on your next page, and steal freely — the same way you can take our free website prompts and build today. Want the wider system — briefs, cadence, internal-linking across a whole site? It’s in our SEO playbook.

And if you’d rather spend your attention on the business than the loop, we run it for you. SEO optimization is a one-time job from $1,800; when AI search is the target, our SEO and AI-search audit is $1,200 one-time with ongoing help from $950 a month. Same prompts, same checks — we just hold the pen.