GEO in three sentences
No provider publishes an official GEO guide: what exists is technical documentation on how each AI search engine crawls and which sources it uses, and the rules follow from that. The academic reference is the paper GEO: Generative Engine Optimization (Aggarwal et al., Princeton, 2023), which found that adding statistics, expert quotations and sources improved visibility in generative answers by up to 40%, while keyword stuffing did nothing. Google, for its part, says its AI features need no special requirements beyond good SEO, which is also a clue.
How AI models choose sources
- Memory (training). What the web said about your brand up to the model's cut-off. It changes slowly, over months, and consensus matters: how many different sources associate your brand with what you do.
- Live search. To answer, the model rewrites the question into several sub-queries (query fan-out), reads the top results from its search engine and summarises what repeats. ChatGPT and Copilot rely mainly on Bing; Gemini and Google's AI Mode on Google; Perplexity on its own index.
- Synthesis. The model cites the page it reads, but recommends the brand that appears in the most independent sources.
In practice: ranking in Bing's top 10 for each sub-query can be worth more than first position on Google for the main query. And your website is only one source: the main lever is the consistency of your brand on third-party sites.
The method
1. Entity and consensus
- One name, always the same. If the website, domain, distributors and legal name differ, the model sees several weak entities instead of one strong one.
- The anchor sentence. A 20 to 25-word definition of your brand with concrete facts, repeated almost verbatim on the site, listings and profiles.
- Wikidata before Wikipedia. A well-sourced Wikidata item feeds the Knowledge Graph and many models.
- Data nobody else has. Original, verifiable, consistent figures: a model prefers to repeat a concrete number over "exceptional quality".
2. Where models read (and almost nobody works)
- Bing first. Bing Webmaster Tools, sitemap submitted and IndexNow enabled so new URLs reach Bing within hours. It feeds ChatGPT and Copilot.
- Don't block yourself. Check that OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot and Bingbot can get in, in robots.txt and in the CDN or firewall.
- Forums and Q&A. Reddit, sector forums and Quora are cited a lot. Useful answers from a real, identified profile work; spam gets deleted.
- Comparison sites, marketplaces and niche directories, read as neutral sources, plus your Google Business Profile.
- Awards with a URL. Recognition only counts if the organiser has a public page saying so.
3. Writing to be cited
- Answer the sub-queries in the first two or three sentences of a page.
- Self-contained sentences with the brand and the fact in the same sentence, because models clip fragments.
- Honest comparisons on your own site, fair to every option.
- Visible freshness: update dates and current-year data.
- Other languages: English often has less GEO competition.
4. Product channels
For ecommerce, the shopping carousels in Gemini and Google's AI Mode come from Google Merchant Center: a clean feed with GTIN, brand, reviews and shipping is essential. OpenAI runs its own merchant programme for ChatGPT product recommendations.
Measurement
The metric that matters is the recommendation rate: the share of answers in which your brand appears as a recommended option.
- A fixed panel of about twenty prompts, checked regularly in ChatGPT, Gemini, Perplexity, Copilot and Google's AI Mode in a clean session: do you appear, in what position, what is said, is it correct, which URLs are cited.
- Bing Webmaster Tools, which now offers an AI performance report with citations in Copilot and partners.
- Google Search Console, which includes AI Overviews and AI Mode in the web performance report.
- Google Analytics 4, with an "AI" channel group for visits from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com.
My clients see all of this in their private panel (panel.soydavid.es).
Realistic timelines
Model memory takes months to change; live search can move in four to eight weeks. The honest horizon is around six months, measured from day one.
FAQ
Do I need an llms.txt file or special schema to appear in AI?
Google says its AI features have no additional requirements: no llms.txt or special structured data. Structured data still helps your site be understood and llms.txt does no harm, but neither is the main lever. The lever is consistent presence in third-party sources and good indexing in Bing and Google.
Does GEO replace SEO?
No. Assistants' live search reads results from engines like Bing and Google, so without solid SEO it is hard to be cited. GEO adds entity work, external sources and measurement inside the assistants.
Can you guarantee ChatGPT will recommend me?
No, and nobody can. What I do is work on the levers known to influence it and measure every month whether your recommendation rate goes up.