LLMs decide what to answer in two stages. First, they choose whether to answer from training memory or search the web; if they search, they split your question into sub-queries and retrieve pages. Then the model favors passages that answer directly, cite evidence, and match what other trusted sources say, and those become its citations.
What is GEO?
Generative engine optimization (GEO) is the work of making your content more likely to be used and cited inside AI-generated answers from tools like ChatGPT, Perplexity, Gemini, and Google’s AI Overviews.
The term comes from a 2023 research paper by a team at Princeton and IIT Delhi, later published at the KDD 2024 conference: GEO: Generative Engine Optimization. Google’s view is simpler. Its guide to generative AI features states that, from Google’s perspective, optimizing for AI search is still SEO.
Both are right. The foundations are SEO. What changes is the last step: instead of ranking a page, an LLM picks a few passages to build an answer from. To influence that, you need to know how the picking works.
The two ways an LLM answers a question
Every AI answer comes from one of two sources, and each requires a different strategy.
| Answering from memory | Answering with retrieval (search) | |
| Where facts come from | Patterns learned during training | Pages fetched live from a search index |
| Citations shown | Usually none | Yes, links to the pages used |
| How current | Only up to the model’s training cutoff | Current |
| What you can influence | How widely and consistently the web described you before the cutoff | Whether your pages are retrieved, and whether your passages get picked |
| Speed of change | Slow, months to a new model | Fast, days to weeks |
Retrieval is where you have the most influence, so the rest of this guide focuses on it. Google calls this grounding, or retrieval-augmented generation (RAG): its AI features retrieve pages using its core Search ranking systems, then generate a response with links to those pages.
How LLMs decide what to answer: the 5 stages
The exact rules differ by product and aren’t public. But Google’s documentation and the GEO research describe the same basic pipeline.
- Decide whether to search. Questions that need current, specific, or local facts usually trigger a search. Stable general knowledge may be answered from memory. Pew Research found Google showed an AI summary for 60% of searches starting with who, what, when, or why, and for 53% of searches of 10 words or more (Pew, July 2025).
- Fan out into sub-queries. The model rewrites one question into several related searches. Google’s own example: a question about a weed-filled lawn might fan out into searches about herbicides, chemical-free removal, and prevention. You can be retrieved for a sub-query even if you don’t rank for the original question.
- Retrieve candidate pages. Each sub-query pulls results from an index. Google’s AI features use the Google Search index, and a page must be indexed and eligible for a snippet to appear. ChatGPT search relies on OpenAI’s own crawler: per OpenAI’s crawler documentation, sites that block OAI-SearchBot won’t be shown in ChatGPT search answers.
- Pick the passages. The model reads the retrieved pages and chooses the parts that best answer each sub-question. The GEO research shows that passages with sourced statistics, quotations, and citations are used more. It’s also likely, though not documented by the platforms, that models favor claims several sources agree on.
- Write and attribute. The model writes the answer and links the sources it drew from. Being cited near the top of the answer, or throughout the text, is worth more than a passing mention.
Two things follow. You need to be retrievable (stages 1 to 3, mostly classic SEO and crawler access). Then you need to be the best passage (stage 4, mostly writing).
What the research says earns citations
The GEO paper tested nine content changes on 10,000 queries, using a test engine built on GPT-3.5 and the top five Google results, then repeated key tests on Perplexity. Here is what it found.
| Change to the page | Result in the study |
| Add quotations from credible sources | Among the top three: 30–40% more visibility on its main metric. Best performer on Perplexity (+22%) |
| Add statistics instead of vague claims | Among the top three; up to 37% gain on Perplexity’s subjective measure |
| Cite credible sources in the text | Among the top three; strongest when combined with other changes |
| Improve fluency and readability | 15–30% visibility gain |
| Make the tone more persuasive or authoritative | No significant improvement |
| Keyword stuffing | Little to no improvement; about 10% worse than doing nothing on Perplexity |
The most useful finding for small businesses: lower-ranked pages gained the most. Adding citations raised visibility by 115% for pages ranked fifth in Google, while the top-ranked page’s visibility fell by about 30% on average. Content quality can beat domain strength at the passage-picking stage.
Treat these as direction, not formulas. The study ran in 2023 on older models, and today’s engines retrieve more sources. The authors also found results varied by topic, so test on your own pages.
Which sources AI answers cite today
Pew’s analysis of 68,879 Google searches from March 2025 shows who wins the citations now. Wikipedia, YouTube, and Reddit were the most-cited sources in AI summaries, together accounting for 15% of all sources. Government (.gov) sites made up 6% of AI summary sources, against 2% of regular results.
The pattern is clear: reference-style pages, video explanations, and real discussions get picked, and so do official sources. For a business, that means your own site isn’t the only lever. A clear YouTube explainer, accurate directory and review profiles, and genuine participation in forums your buyers read all feed the same pipeline.
How to get your content chosen
This is the playbook we’d follow for a service business, in order of impact.
- Let the retrieval bots in. Allow Googlebot, Bingbot, and AI search crawlers such as OAI-SearchBot in robots.txt. OpenAI treats GPTBot (training) separately from OAI-SearchBot (search), so you can opt out of training without disappearing from ChatGPT answers.
- Answer first, then explain. Put a one- or two-sentence answer directly under each question-style heading. That’s the passage a model can lift without rewriting.
- Add evidence a model can quote. Replace “we’re fast” with a number you can stand behind and its source. Quote named experts, including your own team, with their role. Link the studies and rules you rely on.
- Cover the fan-out questions. For each service, think through the follow-ups: cost, timeline, permits, what affects price, when not to hire you. Answer them on the main page instead of spinning up dozens of thin pages, which Google warns can breach its scaled content abuse policy.
- Make your facts match everywhere. Services, service area, hours, and pricing language should be consistent across your site, Google Business Profile, directories, and review sites. Conflicting facts give a model no reason to trust you. This is the same work as building E-E-A-T trust signals.
- Bring something only you know. Google’s guide says unique, non-commodity content will likely matter more than any other tactic. Local job data, real project photos, and first-hand comparisons are things a model can’t get anywhere else.
If you’ve worked through optimizing for Google AI Overviews and AI Mode, steps 1, 2, and 5 will look familiar. Steps 3, 4, and 6 are where the GEO research adds the most.
GEO myths you can skip
- Keyword stuffing for AI. It didn’t help in the GEO study and hurt on Perplexity.
- Sounding more authoritative. A confident tone alone made no significant difference. Evidence did.
- llms.txt for Google. Google says its Search and AI features ignore llms.txt and similar AI text files.
- Chopping pages into tiny chunks. Google says there’s no need; its systems find the relevant part of a longer page.
- Buying mentions. Google warns that chasing inauthentic mentions doesn’t help in its AI features.
- Guaranteed AI citations. No one can promise these. The pipeline is probabilistic, and answers change between runs.
How Agreed Technologies approaches GEO
At Agreed Technologies, our SEO work centers on answer engine optimization, so GEO is part of how we plan and write content, not a separate add-on. We use Semrush, Ahrefs, Google Search Console, and GA4 to see which pages get retrieved and cited, and we keep reporting tied to leads rather than vanity numbers.
In practice, that means rewriting key service pages answer-first, adding sourced numbers and named expert input, and cleaning up inconsistent business facts across the web. [NEEDS CLIENT INPUT: confirm these are standard deliverables, and add one client example with permission] For the wider shift behind this, see why AI-driven SEO matters in 2026 and our AI marketing guide for small businesses.
Next step: call +1 (425) 216-6843, email contact@agreedtechnologies.com, or use our contact page. We’re headquartered in Newark, DE, with a team in Noida, India, and work with businesses across the US.
Want to know whether LLMs are picking your pages or your competitors’? Book a strategy call,ll and we’ll show you where you stand.
