SEO with AI: from keyword research to content that ranks

SEO has a reputation as a dark art, but the vast majority of the work in it is ordinary: list what people search for, sort it into groups, figure out why they're searching for it, write a text that answers it, and check a hundred small details. None of that is rocket science. There's just an enormous amount of it, which is why companies pay agencies to do it.

That mechanical layer is exactly what AI can now handle in an afternoon — and that's exactly where it stops. Deciding which topics to bet on, the experience you fold into the text, and the point of view that sets you apart from twenty identical articles can't be delegated.

We'll go through seven phases: from first phrases through intent and a content brief to writing, on-page checks, the new era of AI search, and measurement. Every phase comes with copy-paste prompts — just fill in the brackets. And one rule sits above the whole thing: AI doesn't know search volumes or your rankings. Numbers always come from a tool, never from a chat.

A typical scenario

Petra runs marketing at a small company that sells ergonomic chairs. The site has seventeen pages, the blog has been dormant since last year, and organic traffic rests entirely on branded queries — people who already know the company. An agency quoted her an audit, keyword research, and six articles a month; there's no budget for that.

Petra blocks off two afternoons. She spends the first on research: from fifty seed phrases, AI helps her build a list of several hundred variants, she sorts them into twelve topic clusters, and for each one she works out whether people are searching for information or looking to buy. She checks the volumes in a tool and drops eight clusters — either too competitive, or aimed at a different customer.

The second afternoon she builds content briefs: a breakdown of what's on page one of the SERP, which questions keep recurring in the topic, what structure makes sense. She writes the articles herself, but from a brief and with material in hand, so a first draft takes two hours instead of a day. Half a year later she has fifteen pieces covering the whole buying journey, from “my back hurts at my desk” to “which chair for a taller person” — and inquiries have started coming in from people who'd never heard of the company before.

Phase 1: keyword research — from seeds to clusters

Keyword research is the most mechanical of all SEO activities, which is exactly where you save the most time. The goal: a list of topics where you know who's searching for them and why — not a spreadsheet with a thousand rows nobody ever opens.

Seed phrases: start with what customers actually say

Seed phrases are your starting points — the plain words your product or service gets described in. The best source isn't a tool, it's your customers' actual sentences: inquiry emails, chat questions, questions asked in sales meetings, competitor reviews. People search the way they talk about a problem, not the way you word it in your price list.

I run a [line of business] and sell [product/service] for
[target audience] in [region]. Here's what customers ask me most:

[paste 10–20 real questions from emails, chat, or meetings]

Generate a list of 50 seed phrases such a person might search with.
Split them into groups:
1. Phrases describing the problem (the person doesn't yet know a
   solution exists)
2. Phrases describing a type of solution (searching for a category,
   not a brand)
3. Phrases describing a specific product or spec
4. Comparison and decision phrases

Use everyday language a layperson would type into a search box,
including informal and shortened forms. No jargon unless people
actually use it themselves. For each phrase, note who you picture
searching it.

It returns a broad starting set, and a fifth of it is usually stuff you wouldn't have thought of yourselves — especially in group 1, phrases in the customer's words rather than the seller's. Watch out: the model sometimes drifts into vocabulary nobody actually uses in your market, and sometimes invents phrases nobody searches at all. Treat the set as hypotheses, not data.

Expansion and long tail

Long-tail phrases are longer, more specific queries with lower volume and much less competition. For a small site, this is the one battlefield where you can actually win: you won't rank for “office chair,” but you can rank for “office chair for someone with lower back pain” — and a person searching that precisely is much closer to a decision.

Handle expansion with a follow-up request: for each seed phrase, generate fifteen longer variants across four angles — a question, an added situation or persona, an added spec, and a comparison. Ask for the output as a table with columns for the phrase, the variant type, who the model pictures searching it, and what stage of the decision that person is at; the format of your content then follows from that last column. Add the instruction “don't generate variants that are just the same words shuffled,” or you'll get hundreds of near-duplicates. And don't forget the sources AI can't replace: the search box's autocomplete, the “people also ask” block in results, and your own site's internal search.

Clustering: turning several hundred phrases into a handful of pages

The most common beginner mistake is writing one page per phrase. Search engines today understand that “best ergonomic chair” and “which chair is best for the office” mean the same thing — and two pages about the same thing compete with each other. Clustering is the process that turns several hundred phrases into twelve or twenty topics, each one a single page.

Here's a list of keyword phrases for the topic [topic]:

[paste the list of phrases, several hundred lines is fine]

Group them into topic clusters based on what the person is actually
searching for — not shared words. Two phrases belong in the same
cluster if the same page would satisfy both of them.

For each cluster, return:
- the cluster name (the topic in one sentence)
- the primary phrase that best represents the cluster
- every phrase that belongs to it
- what type of page the topic calls for (article, category, product,
  FAQ, comparison) and which step of the buying decision it belongs to

At the end, list separately any phrases that don't fit anywhere, and
explain why. Don't drop anything — every phrase from the input must
end up either in a cluster or in that leftover list.

It returns a content plan in raw form. The last paragraph of the prompt matters: without it, the model quietly drops part of your input and you won't notice. The leftover list, too, often holds the most interesting topics.

Volume and competition: where AI stops

This is where a lot of people trip up: a language model has no idea how many times a month something gets searched. Ask it for search volume and you'll get a number — plausible, confident, and made up. Volume, difficulty, and who's ranking on page one belong to tools: Google Keyword Planner, Search Console, or, for a country-specific market, whatever the local equivalents of Ahrefs and Semrush are for larger budgets.

The workflow runs in the right order because of that: AI builds the clusters, you pull the real numbers from a tool, and you make the call. Priority is set by three things — how many people search it, how strong the competition is, and how close the topic sits to your money. A cluster with low volume right next to a purchase beats a cluster with high volume on the edge of your field.

Build the plan by feeding the clusters, with numbers from your tool, back into the chat along with a description of your site and how many pieces you can realistically produce each month, and have it sequence them into a six-month schedule — with a reason for each position and an estimate of how long before anything shows results. Add the instruction “be conservative, not optimistic, and don't assume I'll leapfrog established sites on the hardest phrases.” Without it, models love to promise page one within three months, which is fantasy for competitive queries.

Phase 2: search intent — content that matches why someone is searching

The most expensive SEO mistake isn't picking the wrong keyword. It's picking the right keyword with the wrong content format: an online store puts up a category page for a query that's really a question, or a blog writes an article where the person wants to buy right now. Neither one ranks, no matter how well it's written.

Four types of intent

Informational — the person wants to know something (“why does my back hurt at a desk”). The answer is an article, a how-to, an FAQ; sales pressure hurts here.

Transactional — the person wants to buy (“buy an ergonomic chair”). The answer is a product, a category, a form; a long intro hurts here.

Commercial investigation — the middle step people forget about: the person already knows they want to buy, but not what (“best ergonomic chair,” “reviews”). The answer is a comparison, a ranked list, a spec table. This is where the money decision gets made, and this is usually where the content gap is.

Navigational — the person is looking for a specific site. The only type content can't win, only being known can.

Reading intent from the SERP

Intent isn't guessed, it's read off the search results. Run the phrase and look at what Google actually shows: if page one is full of how-to guides, it wants a how-to. If it's full of store category pages, an article won't get in. If it's a mix, the intent is mixed.

I searched for [phrase] and here are the first 10 results (title,
domain, page type, and description snippet):

[paste the results]

Analyze them for me:
1. What search intent do the results imply — informational,
   commercial investigation, transactional, navigational, or mixed?
2. What content format does the search engine seem to favor
   (length, page type, structure)?
3. What do the results have in common, and what sets the top three
   apart from the rest?
4. What page format would you recommend to have a shot at ranking
   here?

Work only from what's in the pasted list. Don't guess at the content
of pages you can't see, and don't estimate rankings or traffic.

It returns a structured read of the SERP without missing anything. The last paragraph is mandatory: without it, the model starts describing the content of pages it's never seen, and you end up building a brief on fiction.

Almost everything else follows from the intent type. For informational content, the first screen decides — the answer should come right away, not after ten paragraphs of wind-up. For commercial investigation, the table and the honesty decide: nobody trusts a comparison where your product comes out best at everything. For a transactional page, what matters is what the person needs to decide, not a paragraph about your company's history. And when the SERP shows guides and products side by side, do both and link them together.

Phase 3: the content brief — a plan to write from

A brief is the assignment for writing: what the piece needs to cover, which questions it has to answer, how it's structured, and what makes it different. Write without a brief and you write from yourself — and you usually skip half of what people actually care about.

Analyzing the SERP competition

I'm about to write a page on [topic], primary phrase [phrase],
identified search intent: [informational / commercial investigation
/ transactional].

Here are the headings (H1, H2, H3) from the first five results
ranking for this phrase:

[paste the outlines — headings copied from each page, labeled by
domain]

Give me a breakdown:
1. Which topics every page covers (the required minimum)
2. Which topics only one or two cover (possible differentiators)
3. Which questions a reader would naturally have that none of the
   pages answer
4. What the typical structure is and why that order makes sense
5. Where those pages are weak — where a heading promises more than
   the section delivers

Don't try to tell me my page will be better. I want an objective
breakdown.

It returns a map of the topic that shows both the required minimum and the room to stand out. Point 3 is the most valuable — a gap nobody else fills is the cheapest way to get linked to, and to get cited by an AI search assistant.

Questions the piece has to answer

A good SEO piece is, at its core, a set of answers: a reader who arrives with a question and leaves without an answer goes back to the results — and that's the strongest signal your page fell short. Have it list twenty-five questions your reader is asking about the topic, split into three groups: the ones they'd say out loud (and would type into a search box), the ones in their head that they wouldn't ask (doubts and concerns), and the ones that only occur to them after reading the basics. The second group is the one that pays off — unspoken concerns are what actually drives demand, and competitor content usually skips them.

The brief as an assignment

Put together a content brief for an article on [topic].

Inputs:
- primary phrase: [phrase], secondary phrases: [list]
- search intent: [type]
- what the competition covers: [paste the SERP breakdown output]
- reader questions: [paste the previous prompt's output]
- my edge that the competition doesn't have: [own data, experience,
  a real case, access to people in the field]

The brief should include:
1. One sentence on what the reader should know or be able to do
   after reading
2. A two-level heading outline in an order that makes sense to
   the reader
3. For each section, 2–4 bullets on what it must cover, plus a
   length estimate
4. Which phrases should naturally appear in which sections
5. Which of my advantages go where, so the piece isn't just a
   summary of other people's work
6. What must NOT be in the article (to keep it from bloating)

Don't write the actual text. I want a plan.

It returns a one- to two-page assignment you can write from quickly. The last point saves more than it looks like it would — without it you get pieces nobody finishes reading, covering three topics at once.

Phase 4: writing with AI without SEO spam

This is where it's decided whether you end up with a site that ranks or a pile of text that looks like content and that nobody reads. The line is simple: the difference between “write me an article about X” and “help me write an article from this brief, using this material of mine.”

Quality over keyword density

Stuffing a phrase into a text twenty times was a technique that worked around 2010. Today it's a quality red flag: search engines understand synonyms and context, and a text that repeats “ergonomic office chair” over and over reads the same way to them as it does to a person — off. The rule: use the phrase naturally in the title, the first paragraph, and one heading; anywhere else, only where a person would actually say it. Related terms cover the rest on their own.

E-E-A-T: the experience a machine doesn't have

The acronym stands for experience, expertise, authoritativeness, and trustworthiness. In practice it comes down to ordinary things: it's clear who wrote the piece and why a reader should trust them, it contains specific things from real practice instead of general truths, and claims have a source.

The first letter matters most today, because it's the one thing AI genuinely can't generate. Experience is what you personally saw, measured, or got wrong — a number from your own operation, the sentence “for three out of ten customers, the problem turned out to be desk height, not the chair,” a photo of the thing you're writing about. The move: write down three to five things only you know before you build the brief, and insist they make it into the text. A draft from AI won't supply them — it can only work them in.

A draft from your own material

Here's my content brief and my raw material for the article:

BRIEF:
[paste the brief]

MY MATERIAL (notes, real-world experience, numbers, cases):
[paste your own material — bullet points and rough is fine]

Write a first draft of the text following the brief. Rules:
- build it on MY material; add general statements only where
  needed for clarity, and flag them with a comment
- where the brief calls for content I don't have material for,
  write a standalone line TODO: [what to fill in] — don't invent
  filler
- no throat-clearing openers like “in today's fast-paced world”
- no summary paragraphs that just repeat what came before
- short paragraphs, concrete verbs, numbers wherever the material
  has them
- use the primary phrase naturally, don't force repeats
- write in [language], tone to match this sample: [paste a
  paragraph of your own writing]

List all the TODOs at the end.

It returns a draft built on your material and a list of gaps. Spot-check a few paragraphs against your source material — models like to “upgrade” a claim into a stronger version than the original had (“helped” quietly becomes “demonstrably solves”). Then rewrite the draft in your own words; the process from the tip on talking to AI, not shooting blind helps here.

Checking for generic AI phrasing

Before you publish, run the text through one more pass: have it flag clichés (“in today's world,” “plays a key role”), unsupported superlatives, empty summary sentences, and above all sentences that could sit in an article about literally anything else. Add “don't rewrite anything, just flag and explain” — if you let the model fix the text itself, you'll just trade one set of generic phrases for another. That last category is the single best quality test there is.

And a final rule for this phase: verify facts and numbers. A model will invent a study, a percentage, and a year so convincingly it looks completely real. The method is in the tip on verifying facts with AI; for SEO content it matters even more, because a mistake stays live on the site for years.

Phase 5: on-page — title, description, headings, links

On-page is a set of small things that don't win anything on their own, but whose absence costs you rankings. And it's work AI does reliably, because it's about format, not knowledge.

Title and meta description

The title tag is the strongest on-page element and, at the same time, an ad in the search results. It has to contain the primary phrase, fit within the displayed length (roughly sixty characters), and give someone a reason to click. The description doesn't factor into ranking, but it drives click-through — and click-through is the whole point.

Write 10 title tag variants and 5 meta description variants for
this page:

Topic: [topic]
Primary phrase: [phrase]
Target reader: [description]
What the page offers beyond the competition: [your edge]
Page type: [article / category / product / comparison]

Rules:
- title under 60 characters including spaces, primary phrase as
  close to the start as possible
- description 140–155 characters, ending on a reason to click
- variants must differ in angle, not just reshuffled words
- at least two framed as a question, two as a number or list, and
  two as a promise of a specific outcome
- no clickbait the page doesn't deliver on, no all caps

For each variant, give the character count and one sentence on
who it's aimed at.

It returns a set to choose from. Recount the character counts yourself — models get them wrong and love to hand back a title twenty characters longer than they claim.

Heading structure

The rules are simple and get broken constantly: one H1 per page, then H2s for main sections and H3s for subsections, no skipping levels. A heading should describe what's in the section, not be clever — someone scanning the page with their eyes decides whether to stay based on the headings. Check it with a single request: paste the finished text and ask for a list of whether there's exactly one H1, where a level gets skipped, which headings don't describe what's underneath them, and which sections deserve to be split. Add “don't change my text, just list findings,” or you'll get a rewritten article instead of a list.

Internal linking

Internal links are the most underrated SEO tool you have, and also the one fully under your own control. They do three things: send visitors further into the site, show the search engine which pages relate to each other, and distribute the site's authority across pages. Link text should describe the destination (not “click here”), links belong in the body text, not just the footer, and above all — a new page has to be linked from old ones. An article nothing links to is an orphan on your site.

The request is simple: paste a list of every page on the site (URL, title, one sentence on the content) plus the text of the new article, and ask which existing pages should link to the new one, and which the new one should link to in turn — always with a specific spot in the text and suggested link text. Add the limit “no more than five links, and don't suggest one just because the pages share a word.” Without it, the model links everything to everything, and the article turns into a directory page.

And one small thing people forget: alt text describes what's in the image, not wherever else you can squeeze in the primary phrase. Compress your images — a slow page loses visitors before they ever see the text.

Phase 6: AI search in the new era — how the machine cites you

This is where SEO changed the most. A large share of people today don't ask a search engine, they ask an assistant — and get the answer straight away, with two or three links underneath it. Whether your site is one of them now matters as much as your ranking in the results.

What changed

A classic search engine rewards you for being the answer. An AI assistant rewards you for being a citable source: a clear, checkable claim on the page, with it obvious who's saying it and how they know. The difference is the unit — it used to be the whole page that got evaluated, now it's a paragraph pulled from the page. The practical consequence: answer directly, and answer early. A section that opens with a direct answer and only then expands is far easier to cite than one that reaches its point at the end.

llms.txt: your site's content, machine-readable

llms.txt is a plain text file at the root of a site that describes, in a machine-readable way, what's there. It emerged as a counterpart to robots.txt: where robots.txt says where bots may go, llms.txt says what they'll find. The format is plain markdown — the site name, a summary paragraph, and sections of links with one-line descriptions. It isn't a standard anyone is required to honor, and its effect is hard to measure; but the cost is minimal.

Structured data and robots.txt

Structured data in JSON-LD format is a piece of code embedded in the page that tells a machine what's what: this is an article, this is the author, this is the date. A search engine builds rich results from it; an AI assistant pulls context from it. Useful types: Article for pieces of writing, Person or Organization for the author, BreadcrumbList for navigation, FAQPage for question-and-answer sections, Product for products. The one rule that matters: structured data has to match what's actually visible on the page. Describing an author who isn't on the page, or a rating you made up, is a fast track to a penalty.

And then there's a decision no site can avoid these days: let AI crawlers in, or block them. There's no universally right answer — a publisher who lives off traffic may want to protect their content, while a company that wants to be recommended wants the opposite. In the second case, it's worth naming the bots explicitly in robots.txt, because an explicit entry makes the intent legible to anyone who opens the file.

Case study: this site

On Produktivní.cz, the new era of SEO is deployed for real, and the intent is unambiguous: be a cited source in AI answers.

First, robots.txt. Besides the general rule for all bots (allow everything except the API surface), the AI crawlers are named explicitly — GPTBot, ClaudeBot, Claude-Web, PerplexityBot, Google-Extended, and CCBot — all of them allowed. There's a comment in the code above it explaining the intent, so nobody accidentally removes it a year from now.

Second, llms.txt. It isn't written by hand; it's generated from the same content as the site, so it can't go stale. It contains the site name, a summary paragraph (who's behind the site, what's on it, what language it's in, how often it's updated), then sections for the guide's chapters, a selection of tips with a link to the full list, AI news, the English version, and contact information. Each entry has a one-sentence description — that's exactly what makes the file useful, because a machine can tell what a page is about without opening it.

Third, structured data. Every page carries JSON-LD: WebSite with a defined search action, Article for pieces of writing (title, description, language, canonical URL, publish date, category, author, publisher, preview image), and BreadcrumbList for navigation. The author is defined once as a Person with their own identifier and links to other sites, and every other reference points back to that one record — so a machine can see that the same person writes the content here and on those other addresses. That's the concrete form of “who's saying this,” which is what decides things in the new era.

You can have your own version built with this prompt:

I want to create an llms.txt file for my site, following the format
from llmstxt.org.

My site: [address]
What the site does and for whom: [description]
Who's behind it: [author/company, why they're relevant]
Content language and how often it's updated: [description]
Site structure:

[paste a list of pages: address | title | 1-sentence description]

Assemble the file in markdown:
- a level-1 heading with the site name
- a blockquote (>) summary paragraph that gives a machine context:
  what the site is, who it's for, why it's trustworthy
- level-2 headings for the site's logical sections
- in each section, links in the form [title](address): one-sentence
  description
- for long sections, link to the full list first, then a selection
- contact info at the end

Write the descriptions factually, no marketing language. Don't
invent anything — use only what's in my input.

It returns a finished file ready to upload to the site root. If the site is built in code, generate it directly from the content — a hand-written llms.txt goes stale within a month, which is worse than not having one.

Writing so AI cites you

Five habits. Answer right at the start of a section. Use specific numbers and names — “most companies” cites poorly, “four out of five clients” cites well. Write under your own name. Cite sources for claims that aren't your own. And keep content current, including a visible last-updated date.

Phase 7: measurement and upkeep

SEO without measurement is guesswork, and the one tool you need is free: Search Console. It shows which queries you show up for, how often, at what position, and how many people clicked. Queries with lots of impressions and low click-through are pages that just need a better title.

Here's a Search Console export for [period] — queries, impressions,
clicks, click-through rate, and average position:

[paste the data]

And here's a list of my pages with their topics: [paste]

Analyze:
1. Queries with high impressions and low CTR — which ones are worth
   rewriting the title and description for, and why
2. Queries at positions 5 through 15 — which are closest to moving
   up, and what it would take
3. Queries I show up for but don't have a dedicated page for
4. Pages that dropped significantly over the period

Don't calculate anything that isn't in the pasted export, and don't
estimate missing numbers. Where the data isn't enough to draw a
conclusion, say so.

It returns a prioritized list of fixes, usually including a few quick wins that pay off more than a new article would. The last paragraph is mandatory — without it, the model fills in an “approximate” traffic figure and you end up planning around numbers that don't exist.

The other half of upkeep is updating. An old piece that still ranks is the cheapest content you have: adding current numbers and a section on what's changed is an hour of work with the effect of a brand-new article.

Common mistakes

• Asking AI for search volume and difficulty. The model doesn't know the numbers and will invent them. Volume belongs in a tool; clusters and phrasing belong in chat.
• One page per phrase. You end up with twenty pages about nearly the same thing, competing with each other, and none of them good enough.
• Ignoring search intent. A perfect article on a transactional query will never rank. Format is read off the SERP, not guessed at.
• Letting AI write the whole piece and publishing it as-is. You get content indistinguishable from twenty others, giving readers no reason to stay.
• Publishing unverified facts. A made-up study or percentage stays on the site for years and damages the whole site's trustworthiness, not just one page.
• Publishing and walking away. A piece with no internal links from existing pages never gets found. Half the work happens after you hit publish.

The best tools

• Google Search Console — the one source of truth for what you show up for and who clicks.
• Google Keyword Planner or a local equivalent — search volume and difficulty for your market; numbers AI doesn't know.
• An assistant with projects — with a project you keep your site structure, brand voice, and brief stored, so you don't have to explain who you are every single time.
• Deep research with citations — for mapping a topic and the competition when you're entering a field; the process is in the tip on market research.
• A prompt library — SEO prompts get reused constantly, so saved and named ones save you more time than a paid tool would.

What you get out of it

• Time: keyword research and clustering, two days of manual spreadsheet work, becomes an afternoon; a content brief goes from hours to twenty minutes.
• Money: what you'd otherwise pay someone external to do is, in large part, mechanical work. Running it yourself costs a fraction of a recurring outside contract — and the domain knowledge stays with you.
• Quality: a brief built on real reader questions produces a piece that answers what people actually care about, not what you felt like writing.
• Peace of mind: a plan for the next six months means you're not asking “what do I write about” every single month.

Pro tip

Do a content audit once a year: list every page on the site, pull Search Console data for each, and have it sorted into four buckets — leave it alone, update it, merge it with another page, or delete it. The last two buckets are usually surprisingly large, and clearing them out lifts site performance more than three new articles would.

And a closing rule that survives every algorithm update: AI handles the mechanics, you supply what nobody else has. A text that's entirely generatable is also entirely replaceable — no matter how well it's optimized.