Formats AI systems tend to cite
Six page formats, what each answers best, what proof each one needs and how to choose between them.
Last reviewed: 14 October 2026. Method authorship declared by Camillo Dantas. The detailed guides for each format are in Portuguese; this page gives the English summary.
Why talk about formats at all?
Different questions call for different answer shapes. "What is the difference between A and B?" calls for a table. "What does this term mean?" calls for a definition. "Is it worth it?" calls for an opinion grounded in experience. A fitting format puts the answer exactly where it is looked for, which helps readers and retrieval systems alike.
A warning comes with the advice. There is no official list of formats that AI systems prefer. What exists are editorial habits consistent with what search engines and assistants say they value: useful, clear content with its own perspective and proof. Google, for instance, advises creating "non-commodity" content with an original view or first-hand experience, and says no special markup is needed to appear in its AI features. Microsoft advises short sections, descriptive headings and direct language. The six formats below are the editorial translation of those principles.
How do the six formats compare?
| Format | Best answers | Proof required | Main risk |
|---|---|---|---|
| Comparison lists | "which one", "what is the difference" | open criteria, data sources, date | bias dressed up as neutrality |
| FAQ with direct answers | short, repeated questions | a source whenever there is a figure | invented questions, vague answers |
| First-person analyses | "what was it like", "does it work" | test method, evidence of real use | simulated experience |
| Opinion with a thesis | "what to think about" | argument, counterpoint, named author | opinion passed off as fact |
| Glossaries and definitions | "what is", "what does it mean" | definition source, related terms | copied or circular definitions |
| Original data and studies | "how many", "how often" | methodology, sample, date, limits | a small sample presented as the market |
Comparison lists
State the criteria before the results. Put data in a table with a source and a date. Say who each option suits and who it does not. Disclose any commercial tie to the items compared. A reader, human or machine, must be able to check how the order was reached. A ranking without open criteria is advertising.
Quick method: define the exact question with audience and context; choose four to eight measurable criteria and explain why each matters; collect each data point from the primary source and note the collection date; leave a cell empty rather than guess; write a "suits / does not suit" paragraph per item. The Portuguese guide Listas e comparativos has a filled-in example.
FAQ with direct answers
Collect questions people really ask, from support tickets, sales conversations and Search Console queries. Answer each in 40 to 80 words with the answer in the first sentence, the entity named in full and a source when a figure appears. Each answer must make sense on its own. One caution: Google announced that the FAQ rich result stopped appearing in Search on 7 May 2026, so the format no longer buys a visual feature there. It remains a clean way to organize objective doubts. The Portuguese guide FAQ com resposta direta shows a filled-in example and the matching schema.
First-person analyses
This format reports a test or use that actually happened: what was done, under what conditions, with what result and with which limits. It carries a described method, evidence such as dated screenshots or exports, and a statement of any commercial tie. Its value is verifiable experience. Invented or simulated experience destroys credibility and breaks the method's ethics. Google's guidance for its AI features favors content with first-hand experience or expertise, and this format is the editorial way to supply it. See Análises em primeira pessoa.
Opinion with a thesis
State the thesis in one sentence, back it with two to four arguments and evidence, answer the strongest counterpoint, and name the author and why they can speak on the subject. The mark of a good opinion piece is that readers can separate fact, interpretation and position. Opinion presented as fact loses credibility with readers and with systems that weigh sources. See Artigos de opinião com tese.
Glossaries and definitions
One page per term. Open with a definition of one or two sentences that works alone, then add an example, a myth-versus-fact line where it helps, related terms and the source of the definition. Third-party concepts (RAG, GEO, E-E-A-T, schema.org) are defined with credit to their origin. A method of your own is defined with a declaration of authorship. The schema.org type DefinedTerm, with inDefinedTermSet and termCode, can describe a term when the page shows the same definition. See Glossários e definições.
Original data and studies
Answer a specific question with data only the brand holds. Describe the sample and the method, show results in a table with n beside each figure, state the limits, and give a date and an owner. Publish the anonymized data file when you can. In the best-known academic study of generative engine optimization (Aggarwal et al., KDD 2024), adding statistics, quotations and cited sources raised visibility in the authors' benchmark, though the result applies to that experiment and not to every site. A vendor study should be labeled as such, with its sample size. See Dados e estudos próprios.
How do you choose a format?
- 01Take the customer's real question, worded the way they word it.
- 02Identify its nature: compare, define, decide, opine, measure or settle a small doubt.
- 03Pick the format in the table above that matches that nature.
- 04Check that you hold the proof the format requires. If not, choose another or produce the proof first.
- 05Combine when it makes sense. A comparison page can end with an FAQ, and an original study can feed a glossary entry.
Is there a mandatory format?
No. In some fields a well-built glossary outweighs ten opinion pieces. In others, original data with methodology opens more doors than a comparison. What matters is that each page has the structure described in the method and that the format serves the question, not a trend.
Where should you start?
If the brand gets repeated questions in customer service, start with the FAQ. If clients confuse certain terms, start with the glossary. If there is internal data nobody else has, treat the original study as the priority, because it is the hardest format to copy. Read what LLM Seeding is for the concept and use the checklist to score the page you end up with.
Sources
- Optimizing your website for generative AI features on Google Search, Google Search Central (2026-07-10)
- Optimizing Your Content for Inclusion in AI Search Answers, Microsoft Advertising (2025-10-08)
- Google Search Central updates, Google (2026-05-08)
- GEO: Generative Engine Optimization, Aggarwal et al., arXiv (KDD 2024) (2023-11-16)
- DefinedTerm, schema.org (2026-10-14)
Check the page you already have
Eighteen checks in six groups show what to fix first.