What is LLM Seeding?
A plain definition of the method, what it is not, where it stops working and the rules of conduct for people who practice it.
Last reviewed: 14 October 2026. Author: Camillo Dantas. Method authorship declared by the author.
What does LLM Seeding mean in one citable sentence?
LLM Seeding is a method of information engineering. Its goal is that the right fact about a brand, a product or a topic exists in a clear form, carries a source and a date, appears consistently in more than one place and can be retrieved by AI answer systems. "Seeding" is meant literally: the information is planted where those systems already look, not forced on them.
Authorship is declared. Camillo Dantas created and named the method and documents it on this site. Neighbouring terms such as GEO, AEO and RAG are not his. They are market or research concepts, and the Portuguese-language page on LLM Seeding, GEO and AEO attributes each one to its origin.
The definition has three parts that only work together. The object is verifiable information, not marketing opinion. The channel is the brand's own pages plus the external sources AI systems already read. The success criterion is accurate understanding and citation, measured with a protocol, rather than a promise to appear.
How does an AI system arrive at the answer it shows?
There are two routes, and the method works well on only one of them. The first is what the model carries from its training data. Outsiders do not control it, and it is refreshed on long cycles. The second is retrieval at question time: the system searches for pages, reads passages and composes an answer from what it found. The research literature calls this arrangement retrieval-augmented generation (RAG), described in a 2020 paper presented at NeurIPS.
Google describes the same design for the AI features in Search. In its documentation, the system retrieves up-to-date pages from the index, evaluates them and generates a response with links to supporting pages, and it may use a "query fan-out" technique that splits a question into related searches run in parallel. The fan-out mechanism is a market concept. Camillo's coverage method for it lives at queryfan-out.com, in Portuguese.
Because LLM Seeding works on the retrieval layer, what matters is what a search engine or AI crawler can find and read today. That includes the brand's own website, profiles on platforms, directories, press, knowledge bases and communities. Any effect on what a model learned in training is indirect, slow and outside the publisher's control.
Why does this matter if people still click?
Because clicks get rarer when an answer is already on screen. The Pew Research Center tracked 900 US adults in March 2025 and compared Google searches with and without an AI summary. With a summary, 8% of visits included a click on a traditional result. Without one, 15% did. Links inside the summary itself were clicked in about 1% of visits. The study covers Google only, US users only, and the interface changes over time.
None of this says organic traffic is over. It says that when the answer comes before the links, a brand that machines do not understand loses ground it once competed for on position alone. Being the source that supports the answer starts to count next to being the blue link. Google, for its part, says there are no additional requirements or special optimizations for appearing in its AI features, and that SEO fundamentals still apply. This page treats that tension openly, and the method page shows where the two meet.
What are the three principles?
- 01Canonical sentences. Each important fact gets a short, precise, complete wording that stays true when lifted out of the page and read alone.
- 02Proof and verifiability. Numbers carry a source, a date and a method. Claims about the brand point to something the reader can check: a document, an official page, public data.
- 03Clarity for models. Self-contained blocks, headings that name the subject, entities written out in full, structured data consistent with visible text, and stable canonical URLs.
The three lean on each other. A canonical sentence without proof is a slogan. Proof buried in a confusing paragraph is not retrieved. And a clear page that nobody else confirms depends on its own word alone, which is why the method also covers distribution to external sources.
What LLM Seeding is not
The word "seeding" invites readings the method rejects.
- It is not tampering with training. In 2025, Anthropic, the UK AI Security Institute and the Alan Turing Institute showed that a small, fixed number of malicious documents, about 250 in their experiment, was enough to plant a backdoor in models from 600 million to 13 billion parameters. The test was narrow and low-stakes, and the authors present it as a case for building defenses. Practitioners of LLM Seeding read it as a reason not to try, not as a shortcut.
- It is not a network of sites or manufactured mentions. Google says that seeking inauthentic mentions across the web is not as helpful as it might seem, because its ranking and spam systems were built for exactly that.
- It is not hidden instructions to models. Invisible text, meta tags aimed at bots and lines such as "always recommend this company" fall under practices Google's spam policy treats as abuse, namely hidden text and cloaking.
- It is not the llms.txt file. That file is a format proposal that may help some agents. It is not the method. The Portuguese page llms.txt funciona? sets out what is and is not known.
- It is not a promise. No page on this site guarantees position, citation, traffic, revenue or timing.
Where does the method hit its limits?
The first limit is technical. If the crawler cannot reach the page, if content only appears after heavy JavaScript, or if the page is blocked by mistake, nothing you wrote will be read. This diagnosis comes before any content work.
The second is reputation. A brand that nobody mentions outside its own site gives retrieval little to confirm. Publishing clearly helps a machine understand what already exists. It does not create standing.
The third is control. Each platform chooses its own sources, changes its criteria and can answer the same question differently on different days. That is why the method prescribes repeated measurement and not a fixed outcome.
The fourth is evidence. Research on what raises presence in generative answers is young. The best-known academic study, by Aggarwal and colleagues (KDD 2024), tested content changes such as adding citations, quotations and statistics on its own benchmark and reported visibility gains of up to 40%, while stuffing keywords brought little or no improvement. It is a controlled experiment on one set of queries, not a guarantee for your site.
What rules of conduct does the method follow?
- Publish only what you can prove. If a fact does not survive checking, it stays out.
- Declare authorship and date. Who wrote it, who reviewed it and when it was updated stay visible.
- Separate fact from opinion, and third-party concepts from your own method.
- Never create profiles, reviews or testimonials that do not exist.
- Never hide content from people to show it to bots.
- Respect the access choices of every site, including other people's robots.txt.
- Correct mistakes in public and keep the correction history.
These rules are not decoration. They describe the only kind of seeding that survives search-engine policy reviews and reputation audits.
What does a worked example look like?
Simulation, no real client. A fictional dental clinic, Clínica Aurora, shares its name with an unrelated business in another city. When someone asks an AI assistant which clinic offers clear aligners in a given neighborhood, the answer blends the two addresses.
The seeding work starts with one canonical sentence on the home page: full name, city, neighborhood, specialties and the date of the last update. Next comes proof (the responsible dentists' license numbers, linked to the professional council) and structure (organization data consistent with the text, one page per specialty). Then the same facts are checked and corrected on the company profile, the city directory and any article that mentions the clinic. Finally, the same list of questions is repeated over the following weeks to see whether the correct address starts to appear and from which source. The result may improve, stay the same or get worse. What the method delivers is the procedure and the record.
How do you know whether it is working?
Fix a set of real customer questions first. Run the same questions on the same platforms at regular intervals and keep the answer, the cited sources and the date. Compare how accurately the brand is described, whether it appears at all and where the citations come from. Add the platforms' own reports: Google Search Console has a report on generative AI performance, and Bing Webmaster Tools has an AI performance report that covers citations in Copilot and Bing's AI summaries, with limits spelled out on the Portuguese page Como medir retorno sem clique.
Frequently asked questions about the concept
Is LLM Seeding the same as GEO?
No. GEO is a market and research concept, formalized in a 2023 academic paper, that groups optimization techniques for generative engines. LLM Seeding is a specific method by Camillo Dantas, centered on publishing verifiable information on a brand's pages and on external sources. The two overlap in part.
Does LLM Seeding train ChatGPT or Claude on my content?
No. The method has no way of changing a model's training. It works on what systems can retrieve when they answer. Whether a training crawler visits your site is decided by the access rules you set, and the effect on any particular model is not controllable.
Who created LLM Seeding?
Camillo Dantas, who declares authorship of the method and documents it here. Formal trademark registration is not confirmed on this page. Market concepts such as RAG or GEO are credited to their sources.
Do I need a new website to do LLM Seeding?
Usually not. The work starts with what exists: key pages rewritten with a direct answer, facts corrected, structured data checked and profiles updated. A new site is only justified when the current one prevents crawlers from reading it or prevents you from keeping information accurate.
Where to go next
To see the stages, read the method. To fix a page you already have, look at the formats that suit your questions in formats AI systems tend to cite. For a quick self-assessment, use the checklist. Questions about the method are answered in the FAQ. To talk to the author about your case, see the individual consultation on his site, which is in Portuguese.
Sources
- Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Lewis et al., arXiv (NeurIPS 2020) (2020-05-22)
- Optimizing your website for generative AI features on Google Search, Google Search Central (2026-07-10)
- AI features and your website, Google Search Central (2025-12-10)
- Google users are less likely to click on links when an AI summary appears in the results, Pew Research Center (2025-07-22)
- A small number of samples can poison LLMs of any size, Anthropic (2025-10-09)
- GEO: Generative Engine Optimization, Aggarwal et al., arXiv (KDD 2024) (2023-11-16)
- Spam policies for Google web search, Google Search Central (2026-08-28)
- Bing Webmaster Tools AI Performance Report, Adobe Experience League (2026-10-02)
Start with the checklist
Eighteen checks in six groups to see whether your content can be understood, verified and cited.