The LLM Seeding method
Six stages that turn what a brand knows about itself into information AI systems can find, understand and cite.
Last reviewed: 14 October 2026. Method authorship declared by Camillo Dantas.
How is the method organized?
LLM Seeding starts from a simple observation. AI answer systems use what they can retrieve and verify. If the correct information about a brand is scattered, ambiguous or unproven, the answer inherits those flaws. The method attacks the three flaws in the order they show up. The full definition sits on the parent page, and the stages are here.
Three principles hold everything together: canonical sentences, proof and verifiability, and clarity for models. Each stage below applies one or more of them, and each ends with a test that lets you say "done" or "not done" without arguing about it.
Stage 1: which questions and which entities have to be right?
Start with what your customers actually ask, not with what you would like to answer. List 20 to 50 real questions in natural language, taken from support, sales, Search Console, comments and related searches. Group them by intent: definition, comparison, price, location, trust, problem.
Then list the entities that appear in those answers: the brand, its products or services, the responsible people, the places and the technical terms. For each entity, note the official name, spelling variants, the address of the page that defines it and the facts that cannot be wrong.
Search systems with AI often split a complex question into sub-queries. Google calls this technique "query fan-out". The mechanism is a market concept, and Camillo's coverage method for it is documented at queryfan-out.com, in Portuguese. For LLM Seeding it is enough to know that one question becomes several, and each one needs a block that answers it.
Success test: a spreadsheet with grouped questions, entities with official names and, for every critical fact, the page where it will be defined. If two pages define the same fact differently, the stage is not finished.
Stage 2: how do you write the canonical answer?
The canonical answer is the official wording of a fact. It runs 40 to 80 words in a short-answer block and opens with the subject named in full. It answers before it gives context, it survives being read without the rest of the page, and it carries a review date.
A good test: copy only that block and paste it into a chat. Would someone who has never seen your brand understand who you are, what you do, where and for whom? If yes, the block is canonical. If it needs the previous paragraph to make sense, it is not.
Avoid pronouns that depend on context ("he", "the company", "our product"). Avoid superlatives without proof. Use the same sentence, word for word, wherever the fact appears: home page, profile, press kit. Here repetition serves consistency. It should not turn into keyword repetition.
Success test: every critical question has a direct-answer block on a home page, and the same fact is identical on every surface the brand controls.
Stage 3: what proves each claim?
Without proof, the canonical sentence is only the brand talking about itself. Proof can be a document (registration, certification, public contract), an official third-party page, data with method and date, or an original study with its methodology laid out. For each critical claim, record what the proof is, where it lives, when it was last checked and by whom.
Numbers need a source, a date and context. Figures older than twelve months carry their year in the text. If a number cannot be verified, the rule is simple: describe the idea without it. The author's own statements go in as statements, not as audited fact.
Authorship is proof too. Pages with an identifiable author, a publication date, an update date and a profile page explaining the author's experience give readers and machines a way to weigh the source. Google speaks of "non-commodity" content: material with a unique, first-hand or expert perspective that does not simply repeat what is everywhere.
Success test: no critical claim is left without a reference, and every reference opens, works and says what the page says it says.
Stage 4: how do you structure the page so robots can read it?
Structure is the technical part of the method, and much of it overlaps with good SEO. Use headings that describe the subject, one topic per section, tables for comparable data and lists for enumerable items. Keep the HTML semantic and the main content present in the HTML the server delivers, without requiring an interaction to appear.
Structured data in JSON-LD belongs here when it describes what the page already shows: organization, person, article, product. Google states that no special markup is required for its AI features and that markup has to match the visible text. Microsoft, in guidance published in October 2025, recommends schema as an aid to interpretation and says there is no "secret sauce" that guarantees selection. The Portuguese page Schema e entidade shows what to declare and what to leave out.
Check access too. Look at robots.txt for the search and on-demand crawlers, the HTTP status, the canonical URL, response time and what a crawler without JavaScript can read. The Portuguese guide robots.txt para robôs de IA lists each company's crawlers and rule examples checked against their documentation.
Success test: the page reads fine without JavaScript, the schema is valid and consistent with the text, and the crawlers you want reading it are not blocked.
Stage 5: where else does the information need to exist?
No system trusts a brand's word about itself alone. Stage five takes the correct information to the external sources machines already consult and people would read: official profiles on platforms, industry directories, Wikidata when the entity meets its notability criteria, press, technical publications, question-and-answer communities, repositories and partner documentation.
The most important rule of conduct applies here: publish only what is true, and take part only where participation is legitimate. Fake profiles, bought reviews, answers disguised as ordinary users and networks of sites built to cite one another are out, on ethical grounds and for practical risk. Google says that seeking inauthentic mentions is not as helpful as it might seem. The Portuguese page Fontes externas lists the source types, what each requires and what to avoid.
Success test: critical facts match, letter for letter where possible, on every external source you can control, and at least one independent source confirms them.
Stage 6: how do you measure and review?
Measurement closes the loop. Take the fixed question set from stage 1, choose the platforms, collect at regular intervals and keep each answer, the cited sources, the date and the model version when it is shown. Judge three things: does the brand appear, is the description correct, and are the cited sources the ones you prepared or others?
Use the data the platforms provide as well. Google reports a generative AI performance report in Search Console. Bing Webmaster Tools offers an AI performance report with citations in Copilot and in Bing's AI summaries. Referral traffic from assistants shows up in analytics under its source domains, with the caveat that part of those visits arrives without an identifiable referrer.
Camillo's method for measuring presence in answers is Citation Share, documented in Portuguese. Review the cycle every quarter or whenever the page, the offer or a platform policy changes.
Success test: a historical series with the same question list, archived answers and a recorded decision for each relevant change.
How long does each stage take?
There is no universal timeline, and anyone who promises a date for appearing in AI answers is promising something outside their control. For internal planning, an effort estimate is reasonable for a small or mid-sized brand. Stage 1 typically takes a week of one person's time. Stages 2 to 4 depend on the number of key pages. Stage 5 depends on how many external sources already exist. Stage 6 never ends. A 30-day execution plan exists in Portuguese: Como começar em 30 dias.
What is the most common mistake?
Skipping stage 3. Many teams rewrite a page with direct answers and then cannot prove any of them. Another frequent mistake is treating stage 5 as volume to produce instead of consistency to ensure. Both appear in the Portuguese page Erros comuns.
Frequently asked questions about the method
Is the method a fixed sequence, or can I work out of order?
The order reflects dependency: you cannot prove what you have not defined, and you cannot distribute an answer that does not exist yet. On sites that already have good content it is common to start with the stage 4 audit and loop back. What does not work is skipping the baseline measurement, because without it you cannot tell what changed.
How many questions should I map in the first round?
Between 20 and 50 questions is enough to start. What matters is that they are real, grouped by intent and kept fixed over time. A very long list slows collection and dilutes analysis.
Does the method require structured data?
It is not required for Google to use a page in its AI features, according to Google's documentation. Consistent structured data helps describe entities precisely and may help other systems, such as Bing, but it never replaces clear, verifiable text.
Can I apply the method on my own?
Yes. The stages and tests were written so a small team can carry out and verify the work. If you prefer guidance, Camillo offers consulting and strategic analysis on his personal site, in Portuguese.
Where to go next
Look at which formats suit your questions in formats AI systems tend to cite, then score your current state with the checklist. If you are new to the idea, start with what LLM Seeding is.
Sources
- Optimizing your website for generative AI features on Google Search, Google Search Central (2026-07-10)
- AI in Search: Going beyond information to intelligence, Google (2025-05-20)
- Optimizing Your Content for Inclusion in AI Search Answers, Microsoft Advertising (2025-10-08)
- Bing Webmaster Tools AI Performance Report, Adobe Experience League (2026-10-02)
- GEO: Generative Engine Optimization, Aggarwal et al., arXiv (KDD 2024) (2023-11-16)
Score your starting point
Eighteen checks in six groups show which stage to tackle first.