How to make content for ChatGPT, Perplexity and Google Gemini in 2025

AI traffic is any visit that starts through an AI assistant or through an AI answer in Google. The big difference with classic search traffic is the starting point. Classic traffic starts with a click on a link. AI traffic starts with an answer that already carries value, after which the visitor clicks on for more.

For you as a business owner the same thing counts in the end: contacts and customers, whatever the channel. For a marketer it is about entities, topical authority and structures that language models can read well. Happily, the shared basis is simple: plain language, clear headings, calm paragraphs and practical examples.

Worth remembering: AI models see no colours and no design. They do see text, headings, lists, code and schema data. And they pay attention to coherence, explanation, examples, definitions and step by step plans.

You do not have to write four versions of the same article. Each system simply looks at your page differently:

• Google sees your title, meta description, URL, headings and text.
• ChatGPT and Perplexity read whole passages of text, definitions, lists and step by step plans.
• Google Gemini leans on pages that are already well indexed and that carry schema data.

The preferences per model differ as well. Gemini likes structured content with bullet points, a human voice and concrete advice. ChatGPT values completeness, nuance, a logical build and examples. Perplexity is the most aggressive at reading websites, gives literal source references and goes for clarity: quotable blocks, strong headings and checklists.

A finished guide usually carries these building blocks: a clear title, an introduction, chapters, a practical example, frequently asked questions and a summary. Which of them you need depends on the question you are answering. Use one H1 per page, H2 for chapters and H3 for sub subjects. Let the length follow from the question. A guide that answers every sub question becomes long by itself, but a word count as a goal in its own right produces padding, and padding is exactly what a language model skips. Use your brand name, services and tools consistently, so that AI recognises you as an entity.

Write for people and AI at the same time

The human layer consists of short sentences, no difficult words, examples, clear steps and an honest tone. The AI layer consists of consistent terminology, headings phrased as questions, explicit lists and a logical order. Avoid abbreviations without explanation, unnatural keyword stacking and long slabs of text with no structure.

AI reads well: headings, definitions, step by step plans, schema data, lists and summaries. AI reads badly: script blocks, text inside images, blocks with no headings and pure marketing language.

Alongside the content itself you need a technical basis. It consists of four fixed parts:

• robots.txt with AI access: if you want to appear in answers, allow OAI-SearchBot and ChatGPT-User for ChatGPT and PerplexityBot and Perplexity-User for Perplexity. GPTBot is about training, not about search. Google-Extended never sends a request of its own and only governs whether Gemini may use your content. CCBot is the Common Crawl archive.
• llms.txt: a curated table of contents you offer. No large provider has committed to fetching it from external sites.
• Schema data: article schema and FAQ schema give your structure extra meaning.
• Internal links: they create thematic coherence between your pages.

AI crawlers discover your pages through links, through your sitemap and through users sharing a URL. In practice llms.txt is not part of that: in the measurement by Ahrefs from May 2026, 97 per cent of valid llms.txt files received no request at all, retrieved 25 August 2026. This technical layer builds on the base you lay when you make your website easier to find.

Not every subject performs equally well. AI is happy to use you as a source on practical problems, complex subjects, step by step plans, cases of your own and niche subjects. What performs badly is hyper general themes, subjects dominated by large brands, fast moving news and articles with no examples.

So put every idea through three filters: how much the question is asked, how practically it applies, and whether you have a unique angle. Then build authority with three to five pillar articles per theme, supporting articles around them and internal links connecting everything.

If you want AI to use your article as a top source, you need three parts:

• An explicit definition in the first 200 words, short and exactly phrased.
• A step by step plan or process that is logical and complete.
• A practical example that shows you are reliable.

Positive signals are long complete guides, clear definitions, examples, step by step plans, a consistent structure, schema data and internal links. Negative signals are vague short articles, marketing language, outdated content, missing examples and the absence of internal structure.

This is the order we keep to when an article is meant to serve as a source. Not every piece needs all ten; the heavier the subject, the more of them you want in place:

• A title that solves the problem.
• An introduction with context, three to five sentences.
• The definition of the subject.
• Why the subject is relevant in 2025.
• A worked out step by step plan.
• Examples and scenarios.
• Common mistakes.
• A checklist block.
• A short summary.
• An FAQ.

If you want to go deeper, work with the Triple Ranking technique in three layers: semantic completeness (definitions, steps, examples, scenarios and FAQ), AI extraction blocks (try it yourself sections, steps, bullet lists and summaries) and a contextual mini analysis with nuance, exceptions and expertise.

Five techniques for the deepest AI layer

• Semantic echo: repeat meanings, not literal words, at least twice.
• Naming exceptions: paradoxes, risks and context from the real world.
• Micro definitions: small definitions inside larger sections.
• Intent markers: sentences that state the purpose of a block.
• Human context layers: short remarks about experience or situation.

Finished writing? Then test your article with five tests in ChatGPT, Perplexity and Gemini:

• Rephrase test: have AI summarise your article in different words.
• Nuance test: ask for counter arguments and exceptions.
• Question variation test: connect your article to ten other phrasings of the question.
• Comparison test: have your article compared with a competitor's.
• Practice test: does the article really solve the reader's problem?

Google and AI assistants weigh partly different things. Google publishes no list of ranking factors, but it does publish this: it uses a mix of factors that identify content with strong E-E-A-T, and E-E-A-T itself is not a ranking factor, checked on 25 August 2026.

AI assistants look above all at definitions, step by step plans, examples, scenarios and nuance. The shared sweet spot: definitions, step by step plans, at least two examples, scenarios, common mistakes, scannable blocks, a mini summary and an FAQ. So do not see AI as a threat, but as an extra distribution channel.

One question always follows this one: if you use AI to write the article, does Google hold that against you? The short answer is that Google judges the value of the page and not the tool it was made with, and we set out what that does mean for your text in how Google reads AI content.

The framework above lists common mistakes as step seven, and then most guides leave you to guess which ones. These are the six we run into most often when we open somebody else's article, in the order in which they cost the most.

1. A definition that is really a run up

“In today's fast moving digital landscape, more and more entrepreneurs are wondering what X actually is.” That is not a definition, that is throat clearing. A model looking for a passage it can quote finds nothing here that stands on its own. Write the sentence a reader could copy into a message to a colleague: X is a Y that does Z, used when W. Then explain.

2. A step by step plan made of nouns

“Keyword research. Content planning. Optimisation. Monitoring.” Four headings, no instruction. A step is something a reader can carry out this afternoon and afterwards knows whether it worked. If your step cannot fail, it is not a step, it is a category.

3. Twelve thin pages instead of one complete one

Splitting one subject over a dozen short pages was a habit from the era when every page needed its own keyword. It works against you here, because each of the twelve is now too shallow to be useful on its own and none of them carries the full answer. The reverse mistake exists too: dumping four unrelated subjects onto one page because it makes the word count look impressive. The rule is simple. One question, one page, answered completely.

4. Claims with no date and no source

“Recent research shows that most consumers now start their search in an AI assistant.” Which research, by whom, measured where, and when? A sentence like that costs you more than it earns, because a reader who checks and finds nothing stops trusting the rest of the page as well. Either you name the source with a link and the date you checked it, or you leave the number out and say what you do know.

5. Writing the article that already exists

If a model can already answer the question from ten comparable pages, it has no reason to reach for an eleventh. The part that makes you worth quoting is the part nobody else can write: your own measurement, your own case, the exception you ran into last month. That is also the part that takes the longest, which is exactly why most people skip it.

6. Blocking the readers you are writing for

It happens more often than you would think: an article written specifically to be quoted by assistants, on a site whose robots.txt turns those assistants away. Or a page where the substance sits inside a script that only runs in a browser. Check the technical layer from DOC.K.04 before you conclude that your content is the problem.

Publishing is the halfway point. An article that is meant to serve as a source has to stay correct, and correctness has a shelf life. Three habits keep it alive without turning into a second job.

Put a date on the perishable parts

Sentences do not all age at the same speed. A definition holds for years. A sentence about what a particular assistant does today may be wrong next quarter. Mark the perishable sentences as you write them, with the year in the sentence itself or with a check date after the source. That way a future update is a search job of ten minutes instead of a full reread.

Update rather than republish

The instinct to delete an old article and post a fresh one is expensive. The old URL carries whatever links and history it has built, and a new one starts at nothing. Rewrite in place, keep the address, and say in the text what changed and when. If a claim turns out to have been wrong, correct it visibly rather than quietly. A visible correction is the cheapest credibility you will ever buy.

Measure what can actually be measured

This is where honesty matters, because the market is full of dashboards that promise more than anyone can deliver. No assistant hands you a full report of how often it quoted you. What does exist: on 3 June 2026 Google announced Search Generative AI performance reports in Search Console, showing impressions, pages, countries, devices and dates for generative AI features in Search and Discover. In that same announcement Google states the reports are rolling out to a subset of websites first, so not seeing them does not mean you are invisible. Checked on 25 August 2026.

Beyond that you have your ordinary Search Console figures, your referral traffic, and the oldest measurement there is: asking every new enquiry how they found you and writing the answer down. For the wider picture of where you turn up, our page on AI visibility sets out what can be observed and what cannot.

With this framework, from title to FAQ, you can start today. If you are starting from nothing at all, first read how to build your own website. More free guides like this one are in our knowledge base. Would you rather we set the full AI layer up for you, from llms.txt to content structure and schema data? Then look at our service AI and automation, and the rates are listed openly on our pricing page.

A fair warning about what we do and do not promise. We can build the structure, write the definitions, set up the technical layer and keep the article correct over time. Nobody can promise you a mention in ChatGPT or a place in an AI Overview, and a supplier who does is selling you a guess.

So we do not.

If that is the conversation you want to have, half an hour is usually enough to work out whether this is worth doing for your subject at all. Bring one article you already have; we will go through it against the ten blocks of DOC.K.07 and you will see for yourself where the gaps are. You can also simply write to us through the contact page.

Do I have to write four versions of the same article?

No. You write one article; the systems just look at it differently. Google reads your title, meta description, URL, headings and text. Language models lift out separate passages they can quote without the rest. An article that serves both is one with a clear structure and with passages that hold up on their own.

What makes an article usable as a source?

Three things that have to be there together: an explicit definition in the first two hundred words, short and exactly phrased; a step by step plan or process that runs logically; and material that is not elsewhere, such as an example of your own or a measurement of your own. Without that third point you are a summary of what already exists, and a model does not need you for that.

Which subjects work best here?

Practical problems, complex subjects, step by step plans, cases of your own and niche subjects. What works badly is hyper general text on subjects that already have a thousand comparable pieces: there a model picks the source it already knows. The gain sits in the subject where you can say something nobody else can.

How do I test whether my article works?

Put it to the assistants yourself. Have your article summarised in different words and see whether the core survives. Ask for counter arguments and see whether the piece can take them. Ask which part the model would quote. What comes out is not a score, but it does show immediately which part of your text can be used on its own and which part only works as a whole.