Asking an AI to "write me a page to rank for dental clinic in Austin" produces exactly what you'd expect: a correct, fluent text that is completely interchangeable with one for any other clinic in any other city. It reads well and it doesn't rank.

The difference between using a language model for SEO and using it well isn't the prompt. It's the order of the steps and the data you give it before asking it to write. A model doesn't know what your customer searches for, hasn't seen the SERP in your city, and doesn't know what makes your business different. If you don't give it those things, it makes them up.

This article is the complete method, step by step. It works the same with Claude, ChatGPT, Gemini or DeepSeek: it doesn't depend on the model, it depends on the sequence. And at every stage we point out exactly where the AI will fill gaps with plausible but false information, and what real data you need to put in front of it to prevent that.

The principle behind the whole thing: context, data, content, measurement

Most people start at the end. They open the chat and ask for the text. The correct method reverses the order and rests on four stages:

Context: who the business is and who it sells to, in their own words
Data: what people actually search for, and who is already ranking
Content: structure first, text second, copy last
Measurement: check whether it worked, on Google and inside the AI models

Each stage feeds the next. Skipping one doesn't speed up the process — it invalidates it.

Stage 1 · Context: the step almost everyone skips

From the owner's voice notes to a context document

A language model has no memory between conversations and doesn't know your business. The raw material can't be invented: it has to be extracted from what the owner tells you.

The fastest and most honest way to get it is to ask for voice notes. Have them tell you, without a script, what services they offer, why customers pick them, what kind of client calls, and which jobs carry the best margin. If you work with larger clients, always record the kickoff call: it holds far more context than you'll remember afterwards.

Transcribe those recordings and ask the model to structure them into a document covering:

Business details: name, activity, location, phone, website
Real services, each with a short description
Concrete differentiators, with verbatim examples from the recording
Which services are most profitable and why
Ideal customer and the typical cases they mention
Actual service area
Voice and tone: literal phrases that capture how they speak

This is the first critical point. Always add an explicit instruction: invent nothing, only extract and structure what the client actually says. Without it, the model will fill the gaps with "extensive industry experience" and "commitment to quality" — precisely the noise we're trying to eliminate.

That document becomes the project's permanent context. Everything you generate afterwards starts from there.

Buyer personas, but built for SEO

A buyer persona in marketing is a theoretical exercise. In SEO it serves one very concrete and very useful purpose: knowing what words your customer types into Google when they have the problem you solve.

And there's almost always more than one profile. A dental clinic can have three completely different patients: the one in pain who wants an appointment today, the one who has spent months weighing up clear aligners and compares prices without hurrying, and the one facing major implant work whose real question before deciding is whether they can trust you. They search differently, fear different things, and decide for different reasons.

For each profile, ask the model for:

Who they are and the situation in which the need arises
The real outcome they want, not the technical service
What holds them back: fears, objections, previous bad experiences
Ten or twelve verbatim phrases they'd type into Google, mixing urgency with more considered searches
The three or four arguments that weigh most in their final decision

Those search phrases are the seed for all the keyword research that follows. Keep them.

Stage 2 · Data: where AI stops being enough and tools take over

Brainstorming is a hypothesis, not a conclusion

With the context and personas loaded, the model can suggest which services deserve their own page and propose a URL tree. It's a good starting point and it saves time.

But be very clear about what that output is: hypotheses. The model has no search volume data for your city. It doesn't know whether "clear aligners Austin" has real demand or whether nobody in your area searches for it by that name. It's inferring from the context you gave it.

Two filters for that list:

Transactional intent only. Someone searching "how long do dental implants last" is researching. Someone searching "dental implants Austin cost" is about to book. Different audiences, and only the second one matters here.
No overlap. Two pages attacking the same search intent cannibalise each other and neither ranks.

Validating with real volume

This is where hypotheses become certainties or fall apart. You need real volumes, and Google Keyword Planner is free and sufficient for that. Export the results for each seed and hand them to the model so it can clean, group by intent, and assign each group to a page.

When you ask for the grouping, insist on a condition almost nobody sets: only variants with a real semantic difference. Singular and plural, with or without a preposition, or reordered words that mean exactly the same thing are not variants. It only counts if the user's intent or context changes. Without this instruction you'll end up with an inflated list of duplicates that you'll later try to force into the text.

Measuring properly: three devices, not one

And here's the difference between estimating and knowing. Search volume tells you how many people search for something. It doesn't tell you what position you're in, which is the only thing that determines whether the work paid off.

Local SEO has a nuance that breaks most general-purpose tools: there is no such thing as "your position". There are three, and they're usually very different from each other:

Desktop: the classic organic result
Mobile: where most local searches happen, and where the Map Pack fills the entire screen
Map Pack: the three businesses Google shows in the map block, which is what drives the phone call

The RankuTracker rank tracker measures all three separately, with CSV import so you can upload the keywords you've just validated. You can launch it from the tracking panel and see from day one whether the page you're about to write starts from nothing or from page two.

Analysing whoever is already ranking

Before writing a single line, search your target keyword in incognito and open the top three results. Not to copy them: to understand what they answer, what structure they use, and what gap they've left.

Ask the model to analyse those three URLs and tell you what headings they use, why you think they rank, which intents they leave unresolved, and what you could do better given your business context. If the model can't access the URLs, paste the visible text of each page.

That analysis covers content. For the geographic side — which is what dominates in local — RankMap simulates searches from different points across the city and shows you where you appear and where a competitor replaces you. That's information no model can give you, because it requires querying Google from real coordinates. And with competitor monitoring you see who holds those positions consistently, not on a single day.

Stage 3 · Content: structure, text and copy, in that order

The heading hierarchy is the most important step

The H1, H2 and H3 outline is the index that tells Google what the page is about and in what order it answers the user's questions. If the structure is wrong, perfect text is worthless.

The rules that work:

The H1 contains the exact keyword or a very close natural variation
The H2s use real semantic variants from the cluster, not the main keyword repeated
The H3s complement their parent H2, also with real variants
No heading without real search demand behind it
Section order follows the user's mental journey, from arrival to hiring decision

An example of the difference. "Our services" has no search behind it and answers no specific question. "Single-tooth and full-arch dental implants in Austin" contains cluster keywords, resolves a specific doubt, and differentiates by treatment type.

Review this outline by hand before moving on. This is the moment to add or remove sections. Once approved, it doesn't change.

The text, with the keyword list in front of you

Now the text. And here comes the most common mistake of all: if you only ask for "write the SEO text", the model produces something that reads well but ignores a large share of the secondary keywords, especially the lower-volume ones — which are exactly the ones with least competition.

The fix is to hand it the full list and demand a check at the end:

Main keyword and the exhaustive list of secondaries, explicit in the prompt
Natural language, without hammering keywords in
Zero filler phrases like "we are leaders in" or "we have extensive experience"
The real differentiators from the context document, not generic ones
A final checklist marking each keyword as integrated or not, plus an added sentence in the most natural section for any that were left out

That checklist is the trick that changes the outcome. It forces the model to consciously review every keyword before handing you the text as final.

The copy layer: from ranking to converting

The previous text is optimised for Google. Now it needs optimising for the person reading it. These are two different objectives, worked separately, even though the final result satisfies both.

This pass improves the opening, so it hooks within the first two sentences; the CTAs, which should be specific rather than a generic "contact us"; the differentiator sections, so they read as real reasons; and the closing, which has to build trust and reduce friction.

With one non-negotiable condition: if any heading is adjusted, no keyword can be lost. Few changes, kept short, always preserving what they contain.

And this is where the Stage 1 context document pays for itself. When asking for copy, tell the model to use the voice and tone section with the owner's literal phrases. That's what stops the text sounding like AI — because deep down it isn't entirely, it comes from how a real person speaks. Google rewards content with first-hand experience — that's the E-E-A-T criterion — and readers notice it even when they can't explain why.

Stage 4 · GEO and structured data: getting found by Google and by AI

FAQs designed so a model will cite you

GEO — Generative Engine Optimization — is the equivalent of SEO when the thing answering isn't a search engine but a model. When someone asks ChatGPT "which dental clinic would you recommend in Austin for implants?", you want to be in that answer.

The way to achieve it is having content in question-and-answer format that a model can quote without needing extra context. A well-built FAQ block does two jobs at once: it enriches the page for Google and feeds models with specific answers.

What separates a GEO-useful FAQ from filler:

Questions someone about to hire would genuinely ask, mixing urgency, technical doubts, price and process
Conversational answers, the way you'd explain it to a friend
Based on real context, inventing no data and no prices
Each answer mentions the geographic area naturally
Each answer is self-contained: a model can quote it on its own and it still makes sense
Between 60 and 100 words per answer

Structured data and the file AI models read

Structured data is the language that speaks directly to Google. It tells it unambiguously what kind of business you are, where you are, and what you do. For a local business with service pages you need at least LocalBusiness on the home page and FAQPage on every page with frequently asked questions. Always validate before publishing: if there are errors, paste them to the model and it fixes them in seconds.

And there's a newer layer that almost nobody has yet: the llms.txt file, designed so language models can understand your site's structure and content. RankuTracker generates it automatically from your domain's indexed URLs.

The stage the method usually forgets: checking whether it worked

This is where almost every AI workflow stops. They publish the page and assume the method worked. But writing so an AI will cite you and checking whether it cites you are two different things, and only the second is a fact.

You have three measurements to close:

Google position, on desktop, mobile and Map Pack separately, with the rank tracker. This tells you whether the structure and keywords worked.
Presence in the AI Overview, that block Google shows above the organic results and which, for many searches, is the only point of contact. You can check whether you appear and with which pages.
Citation inside the models. AI Visibility sends the questions your customer would ask to ChatGPT, Perplexity and Gemini, and tells you whether they mention you, where you sit in the list, and which competitors appear in your place.

That last measurement is what makes GEO actionable. Without it you're writing FAQs and hoping. With it you know which questions cite you and which don't, and you can rewrite exactly those.

And don't forget the foundation everything else in local rests on: if your name, address and phone number don't match across directories, no amount of content fixes it. The NAP audit checks it, and review management covers the signal that weighs most in the customer's final decision.

Which model to use

The question always comes up, and the honest answer is that it matters less than you'd think. Claude, ChatGPT, Gemini and DeepSeek all do this job well if you give them the right context and data, and all four write mediocre pages if you don't. The sequence outranks the model.

That said, there is one nuance that does matter, and it's not about writing but about measurement: the models you want to appear in aren't necessarily the one you write with. Your customer asks ChatGPT, Perplexity or Gemini. That's why RankuTracker measures citation across those three, not across whichever model you drafted with.

As for the platform's own AI features — review analysis, summaries, suggestions, llms.txt generation — they run on DeepSeek, whose quality-to-cost ratio lets us include them in every plan instead of selling them as an add-on.

In summary

The complete method, no shortcuts:

Real context extracted from the owner's recordings, with an explicit instruction to invent nothing
Buyer personas oriented towards the phrases people actually type into Google
Page ideas as hypotheses, never as conclusions
Validation with real volume and grouping by intent, discarding false variants
Position measured on desktop, mobile and Map Pack, before and after
Analysis of the top three to find the gap, not to copy
Heading hierarchy approved before writing a single line
Text with the keyword list in front of you and a coverage checklist at the end
A copy layer in the owner's real voice
Self-contained FAQs so models can quote you
Validated structured data, plus llms.txt for the AI models
Measuring the result on Google, in the AI Overview, and across ChatGPT, Perplexity and Gemini

Stages 1 to 3 are handled by a well-directed language model. The data and measurement stages need to query Google for real, from real locations and across all three devices. No chat solves that part — and it's precisely where you find out whether the page you just wrote is worth anything.

If you'd rather start with the measurement, which is what tells you whether the rest is worth doing, you can create an account and launch your first tracking run in a few minutes.

How to Build a Page That Ranks on Google with AI: The Complete Step-by-Step Method