SEO Prompts: From Query to Published Page, Step by Step

Ask a chatbot for the monthly search volume of "plumber Lyon" and you may get a very plausible number out of thin air. Unless it's connected to a data tool, a language model knows neither current volumes nor your site's rankings. It can, however, reason very well about data you give it.

That's the whole logic of SEO prompts that work: you pull the material from Google, Search Console or your keyword tool, and the AI does the analysis and writing. The path below follows a page from its query to its update, using the example of a Lyon plumber and that of an accounting firm.

What the AI does well, and what you need to bring it

If an answer contains a number you didn't provide, treat it as false until proven otherwise. Our article on hallucinations details how to limit this reflex of the model.

Step 1: read the intent in the real results

Type your query into Google, in private browsing, and copy the titles and snippets from the first page. Then paste them into this prompt. The results show you the format Google considers useful for this search, and nothing replaces that observation.

Target query: [QUERY]
Here are the titles and snippets from the first page of results:
[PASTE THE RESULTS]

Analyze only this data:
1. What is the dominant intent (get information, compare, buy, find a local provider)? Is there a secondary intent?
2. Which format dominates (guide, list, service page, local listing, comparison)?
3. Which subtopics come up in at least three results?
4. What does no result cover, even though [TARGET] would need it?

Conclude in one sentence: which type of page is most likely to answer this query? If the data doesn't let you decide, say so.

For "water heater leaking what to do," our plumber will probably see emergency guides dominate. A commercial "Lyon emergency repair" page wouldn't have the right format, and a practical guide pointing to the service would be better.

Step 2: a brief fed by your experience

In the old way, you'd ask "write a 1,500-word SEO article on a leaking water heater." Likely result: a decent text, identical to those that already exist, with nothing to prove that the author has ever seen a water heater.

In the new way, you go through a brief, and you inject what only the professional knows:

You're preparing the brief for a page targeting the query "[QUERY]".
Intent analysis: [PASTE THE RESULT OF STEP 1]
Author: [PROFESSION, EXPERIENCE]
Reader: [EXACT SITUATION AT THE MOMENT THEY SEARCH]
My field observations: [3 TO 5 FACTS, COMMON MISTAKES, REAL CASES]

Write a brief with:
- the page's promise in one sentence
- an H2/H3 outline where each heading announces what the reader will learn
- for each section, the question addressed and the field observation to use
- the subtopics not to cover, because they're outside the intent
- the proof to add (photo, real price, example)

Don't invent any observation. If a section lacks material, write "to be completed by the author."

The outline will say nothing new if your observations are empty. That's where you make the difference: the accounting firm will put in, for example, the three mistakes it corrects every year for new sole proprietors (auto-entrepreneurs).

A note on generated content: as of this writing, Google says it evaluates the quality of content rather than how it was produced, and its anti-spam rules target mass publication of pages with no value, whatever the tool. Check its current guidelines. In practice, a page rewritten from what already exists adds nothing, whether it comes from an AI or not.

Step 3: title and meta tags under constraints

A title tag is judged on its ability to make people want to click while staying faithful to the page. This prompt forces the AI to count and justify its choices. Recount behind it anyway, because models are bad at counting characters.

Page: [SUMMARY OR TEXT OF THE PAGE]
Main query: [QUERY]
Competing titles: [3 TO 5 TITLES]

Suggest 5 title tags of 60 characters maximum and 3 meta descriptions of 150 characters maximum.
Rules:
- the query appears naturally, as early as possible
- each title highlights a different benefit and stands apart from the competitors
- no promise the page doesn't keep, no words in capitals, no "ultimate" or "complete"
- give the character count after each suggestion

Finish by recommending one option, with your reason in one line.

Step 4: internal linking from your real URL list

This is an underrated use. Export the list of your pages with their titles, then let the AI look for logical connections, without allowing it to make up an address.

Here is the list of my pages (URL and title):
[PASTE THE LIST]

New page to link: [URL AND SUMMARY]

Suggest:
- 3 to 5 existing pages from which to add a link to the new page
- 2 to 4 pages the new page should link to

For each link, give the exact URL from the list, a descriptive anchor of 2 to 6 words and the reason that's useful to the reader. Use no URL that isn't in the list. If you can't find a relevant link, don't suggest one.

If your list runs to hundreds of lines, work by topic: our guide on long documents explains how to proceed without losing information along the way.

Step 5: diagnosing a page that's slipping

Export from Search Console the queries for a page over two comparable periods, and give them along with the current text. Watch out for seasonality: a guide to boilers can't be compared with itself between January and July.

Page: [URL AND CURRENT TEXT]
Queries, clicks, impressions and position, period A: [DATA]
Same data, period B: [DATA]

1. List the queries that lost the most, separating drops in position from drops in impressions.
2. Formulate hypotheses for each case, specifying what is observable and what remains assumed.
3. Indicate the parts of the text to update, complete or remove.
4. Suggest 3 actions ranked by effort and expected impact, without promising a gain in rankings.

A correlation between two periods doesn't prove the cause. Treat the answer as a list of leads to check, not a verdict, and test your prompt on a page whose history you already know before relying on it.

To start, pick a single page that lost clicks this quarter, export its queries from Search Console and run the step 5 prompt. The result will tell you whether the method is worth extending to the rest of the site.