How to use Claude and ChatGPT for advanced keyword research

How to use Claude and ChatGPT for advanced keyword research without ending up with a list of 200 plausible phrases and zero search volume: this is the difference between using language models correctly and incorrectly in SEO. Models are very good at language, grouping and formulating the questions people ask; they do not have access to search data and, when asked directly, invent numbers. This guide shows the workflow we use at the agency: the model proposes, the data confirm.

The instructions below work in both models; the differences between them for this type of task are small, and we note them where they matter. The general comparison is in the article on Claude or ChatGPT.

How to use Claude and ChatGPT for advanced keyword research

What each tool does well

Task	Language model	Traditional keyword tool
Search volumes	does not know them, approximates them	measured data
Grouping by intent	fast and sensitive to subtleties	rigid, based on text matches
Variations and synonyms	very good, including in Romanian	limited to its own database
People's real questions	plausible, need checking	extracted from real data
Keyword difficulty	cannot estimate it	calculates it from links
Article structure from keywords	very good	nonexistent

The practical conclusion: the model is good at structure and language, the tool is good at numbers. Anyone using only the model is left without volumes; anyone using only the tool is left without intent and without customers' language.

Step 1: context before any instruction

The quality of the result depends on what you say about the business before asking for keywords. An opening message we reuse, with the fields filled in for each client:

I am responsible for marketing at a company in Romania that offers [service/product] for [type of customer], in [city/country]. Typical customers are [description: size, problem, budget]. The specific offering: [3-5 services with indicative prices]. Main competitors: [names]. The site already has pages for [list]. Language: Romanian, with diacritics. Do not invent search volume figures; when you do not know, say so.

The last sentence matters: without it, both models add "estimated" volumes that look real. In Claude, the context can be saved in a project and reused; the organization method is described in the article about projects in Claude.

Step 2: the instructions we use

Expanding the list

Generate 60 search phrases a potential customer would type when they need [service], in Romanian, as people actually search (without diacritics where likely, with common mistakes if any). Divide them into: informational (wants to understand), commercial (compares), transactional (ready to buy), local (with city). Mark those you consider rare.

Classifying an existing list by intent

Here is a list of phrases exported from Search Console, with impressions. Classify each phrase by intent and say, for each one, what type of page would best answer it (service page, article, comparison, frequently asked questions, city page). Then group the phrases that can be covered by the same page.
[list]

Thematic grouping and page plan

Group these phrases into themes. For each theme: a page title (sentence case, under 65 characters), the H2 subheadings it would have, the questions it must answer and the content type. Flag themes that overlap with existing pages: [list of pages].

People's questions

What questions does a customer ask before buying [service]? Formulate 30, as search queries, from introductory ones (what it is, how much it costs) to advanced ones (comparisons, risks, contract). Mark those for which the answer would differ in Romania compared with the rest of the world.

Analyzing existing results

Paste the titles and first paragraphs of the first five Google results for the target phrase and ask: which intent they satisfy, what they have in common, what is missing from all of them and what angle could add something new. This is the most useful way to use the model: it finds gaps you would miss when reading in a hurry. The concept of information gain that Google looks for is explained in the article on information gain score.

Step 3: validation with data

1. Search Console: filter the performance report by each theme; phrases with real impressions are confirmed, along with their current position. The reporting guide is in the complete guide to Search Console.
2. Google Ads Keyword Planner: indicative volumes for phrases with no history on the site; free, with broad ranges when there is no active campaign.
3. A paid tool (Ahrefs, Semrush), if you have one: volumes, difficulty and competitors' phrases.
4. Google Trends for seasonality and comparisons between wordings, as we show in the article on Google Trends in your SEO strategy.
5. Manual searching: for the 20 main phrases, look at what Google displays; if the results are a different type from the page you are planning, the intent was classified incorrectly.

Phrases without volume are not automatically discarded: very specific phrases ("prestashop online store seo audit") have low volume but clear intent and low competition. Our rule: we keep anything with impressions in Search Console or volume in the planner, plus long phrases with obvious commercial intent, and eliminate generalities with no data.

A complete example

For an accounting firm in Cluj, the model produced 60 phrases; 22 had impressions in Search Console (the site was already appearing, at poor positions, for "accountant cluj", "accounting firm cluj" and variations), 14 had volume in the planner without impressions on the site, and 24 appeared nowhere. Grouping produced six themes: general services, prices, accounting for sole traders (PFA), accounting for a new limited liability company (SRL), payroll, changing accountants. The site had a single services page; the plan became five new pages and a frequently asked questions section, with titles and subheadings proposed by the model and corrected by us. Of the 24 phrases without data, four were retained for their intent ("accountant who answers the phone" says something about the customer), and the rest were eliminated. The result after four months: the PFA page at position 4 for the main phrase, with 60 clicks per month, up from zero.

Keywords for local pages

For local businesses, the model is useful for something tools do poorly: wording variations with a city and neighborhood, in real language ("dentist Alba Iulia center", "dentist near the Citadel", "emergency dentist alba iulia"). Ask it for a list of nearby towns that someone in that area would type, then check in Search Console which have impressions; for small towns, Google's planner has no data, and Search Console is the only confirmation. How to build city pages from these lists is described in the guide to the benefits of local SEO.

Competitors' keywords

The model cannot see competitors' data, but it can analyze their pages: paste the titles, subheadings and descriptions from a competitor's five service pages and ask for the list of phrases they target and the gaps they do not cover. Combined with the competitor's keyword report from a paid tool, the analysis shows where you can compete with better content; without a tool, it at least shows the structure used by ranking pages. This is a form of intent analysis, not spying: everything you read is public.

The limits we respect

• Models do not know volumes; any volume figure in a response is invented, even if the model presents it as an estimate.
• Training data have a cutoff date; trends from recent months are missing, and the model's web search helps only partially.
• Wordings may be "too correct": people search for "seo company" without diacritics and with mistakes; explicitly ask for real variations.
• The model repeats familiar structures; by the third list for the same field, the results look similar. Change the angle (by customer type, by purchasing stage, by objection).
• Do not paste client data (email lists, reports containing names) into conversations; use anonymized exports.

Frequently asked questions

1. Can I completely replace a paid tool with a language model? No, if you need volumes and difficulty. Models do not have access to search data and, when asked directly, produce numbers that sound credible without being measured.
2. How do I check an AI-generated list? Run it through Search Console and Keyword Planner. Keep what has real impressions and discard the rest. The model is good at proposals, not confirmations.
3. Which instruction gives the best results? The one describing the business, customer and specific offering, not the one asking for "keywords for SEO". The more specific the context, the fewer generalities you get.
4. Claude or ChatGPT? For grouping and Romanian language, the results are comparable; Claude tends to follow the instruction "do not invent numbers" more strictly, while ChatGPT's integrated web search is more readily available for quick checks. We use both, with the same context.

The validated list becomes the brief for writers, in the format we describe in the article on the SEO brief for writers, while page titles and descriptions follow the patterns in the SEO title examples and meta description examples. For those who want to write the article with the model, not just plan it, the steps are in the guide to writing an article with Claude.

Sources

• Claude documentation, writing instructions: the principles for clear instructions with context and examples.
• Google Search Central, helpful and trustworthy content: the quality criteria that pages planned from these lists must meet.

Source references

• Documentația Claude: scrierea instrucțiunilor
• Documentația Claude
• Google Search Central: conținut util și de încredere