Tech for Retail 2025 Workshop: From SEO to GEO – Gaining Visibility in the Era of Generative Engines

Back to blog

SEO prompts: writing instructions that produce verifiable results

GEO

Discover Incremys

The 360° Next Gen SEO Platform

Request a demo
Last updated on

18/11/2026

Chapter 01

Example H2
Example H3
Example H4
Example H5
Example H6

SEO prompts are the instructions you give an assistant such as ChatGPT to carry out a search optimisation task: expanding a keyword list, preparing an outline, writing tags, analysing an export. The quality of the output depends almost entirely on the instruction. This page is for SEO leads and content teams who want reusable prompts, task by task, with the checks that go with them.

 

It goes deeper into one point of our page on AI and SEO, which sets out what the machine speeds up and what humans keep. Here there is a single subject: writing the instruction.

 

What an SEO prompt is, and what it is not

 

A prompt is not a magic formula to be copied. It is a short specification, written for a fast worker who knows neither your market, nor your data, nor your rules.

 

A specification, not a question

 

Asking "write an article on local search" produces generic copy: the assistant fills every gap in information with the most common answer. A useful SEO prompt takes that freedom away. It states who you are writing for, which intent you are answering, which material to use, which limits to respect and in what form to return the result.

 

Usage is already established in teams: 71% of marketers use ChatGPT (Chad Wyatt, 2026). Mastery is far less so: 10% of AI users master prompt engineering (Digit-Formations, 2026). The gap between the two explains a large share of the interchangeable content published today.

 

What the assistant does not know

 

Three limits determine what a prompt can ask for:

 

  • It does not know your data: search volumes, positions, conversions and existing pages must be supplied to it, otherwise it invents or ignores them.
  • It can state something false with confidence: a figure, a source, a definition.
  • It tends towards the average: without material of your own, it reproduces what the web already says.

 

A well-written prompt compensates for these limits: it supplies the data, forbids invention and imposes a format that makes the output checkable. It does not remove the need for review; it makes review faster.

 

The anatomy of an SEO prompt

 

Whatever the use, an effective prompt brings together the same five components. Leaving them out does not stop you getting an answer; it stops you being able to rely on it.

 

ComponentWhat it specifiesExampleWhat happens if it is missing
ContextThe site, the persona, the intent, the page's objective"B2B site, reader: logistics manager, intent: compare"Copy for everyone, and so for no one
MaterialThe data and sources to use, and those alone"Here is the query export and the current page"Invented or generic facts
TaskA single action, expressed by a verb"Group these queries by intent"An output that mixes several jobs
ConstraintsLengths, tone, terms to use or avoid"Title of 60 characters at most, no superlatives"Outputs to be retouched one by one
FormatThe form of the result"Table with four columns"Continuous text, hard to compare

 

Add two safety instructions to all your prompts. The first: "if a piece of information is missing, write 'missing data' instead of inventing it". The second: "do not add any figure that does not appear in the material supplied". These two sentences avoid most factual errors.

 

Finally, work in stages. A prompt that asks for the research, the outline and the writing at once gives a mediocre result on all three. Chain short instructions together, validating each output before the next: the list of questions, then the outline, then one section.

 

Prompts for keyword and intent research

 

Keyword research is the first area of use, provided you use it to expand and classify, not to measure: the assistant does not know search volumes.

 

Expand a topic by question and by persona

 

Start from a topic and a reader, and ask for wordings, not volumes. A prompt of this kind works:

 

  • "You are preparing a keyword research exercise. Topic: {topic}. Reader: {persona}, in {sector}. List 30 questions this reader asks, from discovery to decision. Sort them into three stages: understand, compare, implement. Word them as the reader would type them. Do not give any search volume."

 

You get leads to check in your tools: that is where you assign a volume, a difficulty and a priority. Repeat the prompt for each persona: the questions of a marketing director and those of a technical lead hardly overlap. Keep the wordings as they come out, even when clumsy: those are the ones you will find, in close variants, in your query reports and in the questions put to assistants.

 

Group and qualify an existing list

 

When you already have a list, from Search Console or a tool, the assistant helps structure it:

 

  • "Here is a list of queries: {list}. Group those that express the same need, so that a single page can answer them. For each group, give: the main query, the variants, the intent (informational, comparison, transactional, navigational) and the suitable type of page. Return a table. Flag ambiguous queries instead of classifying them by default."

 

Check two things: that grouped queries really do show similar results in the search engine, and that no group duplicates a page you already have. How to read the intent behind a query is detailed in our article on search intent.

 

Prompts for the brief and the outline

 

The brief is the stage where a good prompt saves the most time, because it fixes what the writing will have to contain.

 

Prepare a brief per persona

 

Supply the query, the persona and what you know about competing pages, then ask for a framework, not copy:

 

  • "Target query: {query}. Reader: {persona}. Here are the titles of the best-ranked pages: {titles}. Prepare a brief: the intent to satisfy, the questions the page must answer, the concepts and entities to cover, the evidence to gather before writing (data, examples, sources), and what is off topic. Do not write the article."

 

The list of "evidence to gather" is the most valuable part: it shows what you need to go and find in your own data before writing. A brief that asks for no evidence produces copy that looks like all the rest. Have this brief validated by someone who knows the subject before starting to write: a framing error corrected at this stage costs a few minutes, against a complete rewrite once the article has been produced.

 

Build an outline with no repetition

 

Then ask for the outline, with explicit structural rules:

 

  • "From this brief, propose an outline in H2 and H3. Each H2 answers a distinct question from the reader. No subject appears in two sections. For each section, state in one line what it should enable the reader to decide. Flag the sections that overlap with these existing pages: {pages}."

 

The last sentence guards against cannibalisation: the assistant does not know your site, so give it the list of related pages. The writing that follows, and its review, are described in our page on AI writing.

 

Prompts for tags and titles

 

Tags lend themselves well to the exercise: the task is short, the constraints are measurable, and there is much to gain from comparing several variants.

 

Titles and meta descriptions under constraint

 

Always state the lengths, the keyword and what the page really provides:

 

  • "Page: {three-line summary}. Keyword: {keyword}. Propose 8 title tags of 50 to 60 characters and 5 meta descriptions of 150 to 160 characters. The keyword appears in each, without always being placed first. Each proposal announces a different benefit. No superlatives, no promise of results. State the number of characters."

 

Check the lengths yourself: assistants often get their own counts wrong. Also set aside proposals that promise what the page does not contain: a tag that attracts the click on an unkept promise loses the visit as soon as it is won. The construction rules are detailed in our guides to the title tag and the meta description.

 

Section headings and FAQ questions

 

For subheadings, ask that they be understandable out of context:

 

  • "Here is the text of a page: {text}. Rewrite the H2s and H3s so that each announces what the section provides, without the outline. Then propose 6 FAQ questions worded as a searcher would ask them, with an answer of 40 to 60 words drawn only from the text supplied."

 

The constraint "drawn only from the text supplied" matters: without it, the FAQ introduces claims the page does not support. Read the rewritten headings as a list, on their own: if the sequence does not tell the reader what they will learn and in what order, the headings still describe the outline rather than the content.

 

Prompts for updating existing content

 

Updating a page is often more cost-effective than writing a new one. Here the assistant serves first as a reviewer, before serving as a writer.

 

Have it reviewed before rewriting

 

Start with a diagnosis:

 

  • "Here is a page: {text}. Target query: {query}. Reader: {persona}. Without rewriting, list: the reader's questions the page does not answer, the passages that repeat themselves, the unsourced claims, the information that is probably dated, and the sections whose first sentence does not answer their heading. Rank by order of importance."

 

That list is then yours: you decide what gets corrected, with your data. Do not let the assistant "update" figures: it would replace them with others, just as unverifiable. The same applies to links and names of sources: any reference that appears in the output and was not in your material must be treated as unverified until you have opened it yourself.

 

Rewrite one section at a time

 

Once the corrections have been decided, work section by section, supplying the new material:

 

  • "Rewrite this section: {section}. Keep the facts and figures exactly as they are. Add these elements: {material}. Start with a sentence that answers the heading directly. Keep the tone and vocabulary of the original. Length: 150 to 200 words."

 

The instruction "keep the facts exactly as they are" stops a rewording from distorting a piece of data. Reread the section by comparing it with the original, not in isolation. If the rewrite has changed a figure, moved a caveat or turned a condition into a statement, the output should be rejected: this is the most common flaw, and it goes unnoticed on a plain read-through.

 

Prompts for analysing SEO data

 

The assistant can read a table, provided you supply it and set limits on what it may conclude.

 

Make a Search Console export speak

 

Paste or attach the export, then ask for findings and hypotheses separately:

 

  • "Here is a Search Console export over two periods: {data}. Identify: the pages whose position is stable and whose click-through rate is falling; the queries in positions 4 to 15 with many impressions; the queries on which several pages appear. Present the findings first, with figures in support, then, separately, hypotheses to explain them. Do not present any hypothesis as a fact."

 

The separation between finding and hypothesis is essential: an assistant readily proposes a plausible cause without having the means to verify it. Always recalculate a sample of the figures cited before using them.

 

Prepare a report

 

For a summary intended for senior management:

 

  • "From these validated findings: {findings}, write a one-page summary: what has changed, on what scope, what is established, what remains a hypothesis, the three decisions proposed. No technical term without an explanation."

 

SEO automation extends this principle when such analyses come round every month. Only entrust the summary with findings you have verified: a report written with confidence from a false figure is more dangerous than a raw table, because it is read without suspicion.

 

Prompts for GEO: making content citable

 

Generative engines reuse passages that are self-contained and verifiable. A prompt can help you check that a page lends itself to this:

 

  • "Here is a page: {text}. For each section, say whether the first sentence alone answers the question in the heading. Identify the passages that lose their meaning once isolated, the figures with no source or year, and the concepts used without a definition. For each flaw, propose a correction that adds no new information."

 

A second use consists of simulating a reader's questions:

 

  • "You are {persona}. You are trying to {objective}. Ask the ten questions you would put to an AI assistant on this subject, in the order they would occur to you, follow-ups included."

 

Then set that list against your page: each question without an answer is a section to write. Keep one reservation: these questions are a simulation, not a measurement of what your buyers really ask. The extractability criteria are detailed in our article on AI-optimised content.

 

Organising a team prompt library

 

A prompt that works deserves to be kept, versioned and shared. Without a common library, each writer reinvents their instructions, and quality varies from one person to another.

 

Four rules are enough to maintain it:

 

  • One prompt per task, named after its use, with its variables in curly brackets.
  • An example of the expected output attached to each prompt, to judge at a glance whether the result conforms.
  • An associated check: what must be verified before using the output, and who verifies it.
  • A common base placed at the top of every prompt: description of the company, tone, vocabulary, banned terms, safety instructions.

 

This base is what distinguishes brand output from generic copy. The spread of AI makes it necessary: 86% of enterprise SEO teams had integrated AI in 2025 (Squid Impact, 2025). When everyone has the same tools, the difference lies in the material and the rules each team feeds into them.

 

Review the library at a fixed interval. Models change, a prompt that used to give good results can deteriorate, and the brand's rules evolve. Note the date each prompt was last checked, as you would for a page. When a prompt is revised, run the old and the new version on the same input and compare the two outputs side by side: it is the only way to know whether the change improved the result or merely altered it.

 

FAQ on SEO prompts

 

What is an SEO prompt?

 

An SEO prompt is the instruction given to an AI assistant to carry out a search optimisation task: keyword research, brief, outline, tags, update or analysis. It specifies the context, the material to use, the task, the constraints and the expected output format.

 

What is the best ChatGPT prompt for SEO?

 

There is no universal prompt. The best prompt is the one that describes your reader, supplies your data, asks for a single task and imposes a checkable format. An instruction copied without your context produces the same copy as your competitors get.

 

Can ChatGPT do keyword research?

 

It can expand a topic, suggest questions and group queries by intent. It does not know search volumes or difficulty: those measurements come from your tools. Use it to generate and classify leads, then validate them with data.

 

How do you stop AI inventing figures?

 

Supply the data yourself and add two instructions: write "missing data" when a piece of information is not available, and add no figure that is absent from the material supplied. Then check every figure in the output: the instruction reduces the risk without removing it.

 

Should you write your prompts in English?

 

Write the instruction in the language of the result you expect, using your own trade vocabulary: a prompt in French works for producing content in French, and avoids translated turns of phrase. The precision of the context matters more than the language.

 

Is content written with a prompt penalised by Google?

 

No, Google does not penalise content because it was produced with AI: it judges its usefulness. The risk lies elsewhere: generic copy, with no evidence or experience of its own, has no reason to be preferred to the rest. Review and material make the difference.

 

Continue reading

 

Depending on what you need:

 

  • You want to cover a topic better before writing: if you are looking to identify the concepts, entities and gaps in a piece of content → AI semantic analysis.
  • You need to justify the use of AI to senior management: if you need benchmarks on adoption and observed gains → AI statistics.
  • You want to go beyond the hand-written prompt: if your instructions need to carry the brand's tone, personas and data without being copied out each time, Incremys personalised AI generates content from your brand DNA and a brief per persona, with validation before publication.

Discover other items

See all

Next-Gen GEO/SEO starts here

Complete the form so we can contact you.

The new generation of SEO
is on!

Thank you for your request, we will get back to you as soon as possible.

Oops! Something went wrong while submitting the form.