Eleven rules for getting your article cited by ChatGPT, Perplexity, and AI Overviews: one question per URL, a 60-word answer at the top, and dated figures.
An article is cited when it answers a single question, answers it entirely within the first sixty words, and supports every figure with a visible source and date in the text. Everything else is structure: literal headings, self-contained paragraphs, tables for comparisons, and markers where data does not yet exist. No markup tricks.
This is not a theory on language models. It is the template we use to write every guide on this site, published as is. We use it ourselves before recommending it to you.
Everything verifiable in this article was checked against official documentation on 27-08-2026: Google Search Central, developers.openai.com, docs.perplexity.ai, and web.dev. Structural rules are stable; the names of crawlers and rich result features change. If you are reading this months later, check the three links before touching your robots.txt.
Before the rules: what a citation really buys
A citation in ChatGPT, Perplexity, or an AI Overview is not a ranking: it is your sentence reproduced with a link underneath. It is earned by being the easiest unit of text to extract without betraying the original meaning.
And here is the first disclaimer. None of these eleven rules guarantee a citation: they guarantee that your page is extractable. The rest is decided by a system that neither you nor we control.
It is best to start by dismantling the most expensive myth in the sector.
This is the central contradiction of the Spanish market right now. Half a dozen guides in Spanish sell llms.txt as a requirement for visibility. Google's documentation literally says that no new machine-readable file is needed. Both cannot be true.
The actual state: llms.txt is a proposal by Jeremy Howard, whose version 2 was published on September 3, 2024, and no search engine has adopted it as an official standard (llmstxt.org, accessed 27-08-2026). Adding it does no harm. Paying to add it as if it were mandatory, however, does.
The industry norm: The sector sells "AI optimization" files, schemas, and plugins because they are billable and verifiable in a demo.
Our criteria: We fix the text structure first, because it is the only thing official documentation cites as a condition: being indexed and being able to display a snippet.
01 · One question per URL
An article that answers three questions is not cited in any of them. The reason is mechanical: the system looks for a block of text that resolves a specific query, and a multi-topic text forces cutting at points where the meaning breaks.
Write the question in the H1 exactly as someone types it, with their words and their jurisdiction: "Do website prices include VAT?", not "Fiscal policy applied to the digital design sector."
02 · The answer at the top, and self-standing
Below the H1, include a block of between 40 and 60 words that answers the complete question. Self-contained: no "as we will see," no "in this article," and no pronouns pointing to a previous paragraph that does not exist.
The test is simple. Copy that block, paste it into a blank document, and read it. If it needs the rest of the article to be understood, it is not an answer: it is a promise of an answer.
03 · Every H2 is a question, not a clever title
Headings are the index used to navigate your page. "Errors everyone makes" is not a heading: it is a headline. "How much does it cost to redo a poorly built MVP?" is one.
Operational rule: if an H2 cannot end with a question mark or a figure, rewrite it.
04 · Self-contained paragraphs of 40 to 80 words
A paragraph is the extraction unit. It must be understood when ripped out of its context, which prohibits starting with "this," "the previous," or "as we mentioned." Name the subject every time, even if it repeats.
Three or four sentences. The first states, the following qualify. Thesis first, nuance later: never the reverse.
05 · Figure, source, and date in the same sentence
A figure without a date is an expired figure waiting for its turn. Write it all together, in the body, not in a footnote that gets lost during extraction: "The general VAT applicable to web design services in Spain is 21% (Law 37/1992, art. 90)."
Note what that block does: it gives the number, states how it is measured, and says when it was checked. All three pieces of data or none.
06 · If the data does not exist, use a placeholder; never filler
This is the rule that separates us from almost everything published in Spanish on this topic. When an article needs its own figure that we have not yet measured, we write the gap with instructions and publish anyway:
[]
A placeholder costs credibility for a week. A fake figure costs it forever, and it only takes one person trying to trace the source once.
We publish a gap with measurement instructions, not a number that masks the gap.
07 · Table where there is comparison, prose where there is reasoning
A bulleted list with long sentences is not a list: it is chopped-up prose. Reserve the table for things with two or more comparable axes and prose for things with cause and effect.
| Situation | Correct Format | Why |
|---|---|---|
| Three options with price and deadline | Table | Comparable axes, cross-reading |
| A decision with conditions | Prose | The cause does not fit in a cell |
| Checks before signing | Checklist | It is executed, not read |
| A figure with context | Data block | It is cited without dragging the paragraph |
08 · Land in the jurisdiction
An article about web prices that does not name VAT, the surcharge of equivalence, intellectual property rights, or the self-employed status is applicable to any country and therefore not cited in any. Geographic specificity is the cheapest advantage that exists in Spanish.
Name the rule with its reference: Law 37/1992 for VAT, Royal Legislative Decree 1/1996 for intellectual property, Law 11/2023 for accessibility. You write for Spain, not for "the Spanish-speaking market."
09 · Publish the contradiction between sources
When two serious sources contradict each other, most pages that rank today choose one and keep silent about the other. We publish both and explain which one we prefer and why.
Real example, from this very article: Google states that no new markup is needed; much of the sector recommends llms.txt. We have written it above with both sources and their dates. That comparison is the most citable part of the page, precisely because no one else offers it resolved.
10 · FAQ at the end, with literal questions
Write the questions as they are typed into a search engine, using clumsy grammar if necessary. Two or three-sentence answer: the first affirmative and concrete, the second a disclaimer or a guarantee.
And do so without expecting a rich result.
The FAQPage markup no longer gives you stars on Google. The FAQ section is still worth it for another reason: it is the easiest form of text to cite that exists. You write the FAQ for the reader and the extractor, not for the rich result.
11 · Visible date, dated revision, stable URL
Put the publication and last revision date on the page, not just in the metadata. When you update, say what changed. And do not move the URL: every redirect dilutes what a system had already associated with that address.
We revise each guide when a cited source changes, not on a fixed calendar.
| Item | Figure | Note |
|---|---|---|
| Research and source verification | 40 | — |
| Structure and headings | 15 | — |
| Drafting | 30 | — |
| Reviewing figures and dates | 15 | — |
— = data being measured; we will publish it once we have our own sample.
What is not structure but invalidates it
You can write the best article in the cluster and never be cited because a crawler does not enter. Three technical checks, with their exact names.
The agents. To appear in ChatGPT answers, you must allow OAI-SearchBot; GPTBot is for training, and ChatGPT-User responds to actions initiated by a person, in which case—according to OpenAI's own documentation—"robots.txt rules may not apply" (developers.openai.com, accessed 27-08-2026). In Perplexity, PerplexityBot is for searching and Perplexity-User for human requests, which "generally ignores robots.txt rules" (docs.perplexity.ai, accessed 27-08-2026).
Google-Extended. Controls if your content is used to train Gemini and for grounding in Vertex AI. Google specifies that it does not affect inclusion in its search engine nor is it a ranking signal (Google crawler documentation, accessed 27-08-2026). Blocking it is a licensing decision, not a visibility one.
Snippet controls. nosnippet, data-nosnippet, and max-snippet limit what can be shown. If your page cannot show a snippet, it cannot appear as a supporting link in Google's AI features. A max-snippet inherited from an old template leaves you out without anyone warning you.
- The H1 contains the question with the words people use
- The answer block makes sense when pulled from the page
- No paragraph starts with "this," "the previous," or "as we mentioned"
- Every figure has a source and date in the same sentence
- No data gap has been filled with an example number
- There is at least one comparison that no other page in the language offers resolved
- The revision date is visible on the page, not just in the metadata
robots.txtallowsOAI-SearchBotandPerplexityBotif you want to appear in their answers- There are no
nosnippetormax-snippettags inherited from the template - The URL is the final one: there will be no redirect in six months
How we apply this in a project
We do not sell single articles. When we set up a content system for a product, editorial structure enters the same movements as the brand and the web.
| # | Phase | What we do | Deliverables |
|---|---|---|---|
| 01 | Listen | We start by asking what your client actually writes, not what keyword has volume. | Interviews · Real support and sales questions |
| 02 | Define | We agree on what winning means: citations, forms, or calls, with the metric chosen before writing. | Success criteria · Cluster · One question per URL |
| 03 | Map | We draw the map of pillars and satellites and decide what proprietary data will need to be measured. | Link map · List of data markers |
| 04 | Design | We shape the blocks: answer, data, contrast, checks. | Block system · Typographic rhythm |
| 05 | Build | It is written, each figure is verified against its source, and missing data is marked. | Drafting · Dated verification |
| 06 | Launch and follow | We publish, check crawling, and revise when a cited source changes. | Handoff · 30 days of support |
And one more disclaimer, because it is necessary. A well-built content system takes months to yield citations. If what you need is visits next week, this is not for you, and we will tell you that on the first call, not on the invoice.
Writing to be cited is not writing differently: it is writing with the discipline of having to support every sentence in front of someone who is going to check it. That discipline is noticed in the business before it is noticed in the traffic.
Does this way of working fit you? Tell us what you have in front of you in a 30-minute call at no cost.
Frequently asked questions
Do I need llms.txt for ChatGPT to cite me?
There is no official evidence that it is necessary: Google explicitly states that no new machine-readable files are needed, and no search engine has adopted the standard. Adding it costs little and does no harm; if someone invoices you for it as a requirement, ask them for the source.
Is the FAQPage markup useful if it no longer provides rich results?
It is useful as text, not as visual decoration: a question followed by a short answer is the easiest unit to cite. Keep the question section even if you remove the markup, because the value is in the form of the text.
How many words should an article have to be cited?
Length is not the criterion: a paragraph is cited, not an entire article. Our satellite guides hover around 2,400 words because that is how long it takes to exhaust a question, not because a threshold exists.
What if I do not have my own data to publish?
Publish the article with the markers in place and measure later: a declared gap is honest, a fake figure is a ticking time bomb. Start with what you already track, such as real project deadlines or accepted budgets.
Can I block training without losing visibility?
Yes in the case of Google: their documentation says Google-Extended does not affect inclusion in the search engine or ranking. With OpenAI, the separation is equivalent, GPTBot for training and OAI-SearchBot for search, but decide with the documentation in front of you, not with what you read in a forum.
Keep reading
- How to get your website to appear on Google and be cited by ChatGPT, Perplexity, and AI Overviews
- What is AEO and how does it differ from SEO and GEO
- Report: which questions ChatGPT and Perplexity cite us for
- Redesigning your website without losing traffic or ranking
- Web performance for businesses: what are Core Web Vitals and when does it pay to optimize
Transparency note. Some of our own data in this guide is still being measured on real projects, with a single method. We will publish it with its sample size and date range once the sample is large enough; until then, every figure you see carries its source and date of consultation.
Verified sources
- https://developers.google.com/search/docs/appearance/ai-features — Google states there are no additional requirements or special schema.org for appearing in AI Overviews or AI Mode, and that no new machine-readable files are needed; also the nosnippet, data-nosnippet, max-snippet controls and the condition of being indexed and able to display a snippet
- https://developers.google.com/search/docs/appearance/structured-data/faqpage — restriction from August 2023 to authorized government and health websites and removed from documentation on 15-06-2026
- https://developers.openai.com/api/docs/bots — exact names and function of OAI-SearchBot, GPTBot, and ChatGPT-User, and the warning that robots.txt may not apply to user-initiated requests
- https://docs.perplexity.ai/guides/bots — PerplexityBot (search, not training) and Perplexity-User, which generally ignores robots.txt; IP ranges published in JSON
- https://developers.google.com/search/docs/crawling-indexing/google-common-crawlers — Google-Extended controls Gemini training and grounding in Vertex AI and does not affect inclusion or ranking in Google Search
- https://llmstxt.org/ — proposal by Jeremy Howard, version 2 published on 03-09-2024, without official adoption by any search engine
- https://web.dev/articles/vitals — LCP 2.5s, INP 200ms, CLS 0.1 thresholds measured at the 75th percentile
- https://kom.pe/como-escribir-contenido-que-la-ia-cite-como-fuente/ — competition in Spanish reviewed: recommends FAQPage as a priority and does not land on Spanish regulations or mark data gaps; used only to know what the pages that already rank do not say
Want to see this thinking applied to a real project?