In generative search, visibility is decided in three stages: whether your content makes it into the candidate set, whether it is selected during ranking, and whether it "survives" answer synthesis. Classic SEO mostly optimises for the first. GEO (generative engine optimisation) targets all three.
Below is a technical overview of how search works in ChatGPT, Perplexity, Copilot and Claude, how these systems choose their sources, and how you can measure visibility in their answers.
How a generative search engine works
The engine described in the study has three parts: query rewriting, a search interface and a generative model. In simplified form, the process looks like this:
- Intent classification. In ChatGPT logs, a classifier decides in ~28 ms whether to search at all: no search, simple search or complex search.
- Query fan-out. The question is turned into several short queries, e.g. "latest technology news" and "tech news today".
- Retrieval and ranking. The queries run through search APIs, and results are grouped by domain and filtered. For complex questions, the loop repeats until there is enough information.
- Synthesis. The model combines the best snippets, removes duplicates and returns an answer with citations.
Platforms differ, and that shapes your strategy
Each platform relies on a different index and cites differently, so no single strategy works for all of them:
| Platform | Search source | Preferred sources | Citation style |
|---|---|---|---|
| ChatGPT | Bing (reportedly also SerpApi) | Wikipedia for facts, Reddit for experience | Selective; often doesn't cite general knowledge |
| Copilot | Bing + Microsoft Graph | Media, official documentation, e-commerce | Inline citations |
| Perplexity | Multiple APIs, refreshed on demand | News, tech sources, Reddit | Cites almost every claim |
| Gemini | Google's main index | Mirrors classic search | – |
| Claude | In the authors' assessment, mostly pre-training | Durable, high-quality content | Cites when the query requires it |
The search provider itself matters too: in the authors' comparison, Brave returned fact-dense snippets, while Tavily returned narrative-style text. The same page can look different to the model.
What the model actually "sees" on your page
- Snippet-first. These systems mostly read snippets, titles and meta descriptions, not the full page. A meta description that answers the question directly works better than clickbait.
- Sections compete, not pages. A strong H2/H3 section can be cited even on a weak page, so each section should answer one question on its own.
- Accessibility. AI agents often don't execute JavaScript, so the essential content must be in plain HTML.
- Citability. Precise, self-contained sentences with numbers are preferred. Similar claims from several sources get merged, and the more concise one survives.
The study also cites industry data: semantic URLs get ~11.4% more citations, lists and comparisons make up more than a quarter of cited content, and the top 20 Google results largely dominate AI citations.
How to measure visibility in answers
Your position in Google no longer tells the whole story. The study uses two main metrics: Share of Model (SOM), how often a brand is mentioned in AI answers compared with competitors, and Position-Adjusted Word Count (PAWC), how much of the answer is attributed to a source, weighted by position. PAWC gives each sentence of the answer an exponentially decreasing weight.
This captures not only whether you are mentioned, but how early.
What the experiments showed
Ten scoring criteria. The authors built a tool that calculates an "AI Visibility Score" from 10 metrics (structure, citability, readability, entity density, information gain and more).
The metrics conflict. Converting text into bullet points raises the structure score but lowers coherence (Spearman ρ = −0.77); information gain and coherence are also opposed (ρ = −0.73). This is a real conflict between machine reading and human reading.
Automated rewriting has a small effect. The best of 9 prompt combinations improved a page's score by just 2 points (+4%), and the conservative variants actually lowered it.
Brand visibility. Across 20 identical B2B sales queries, PAWC share of voice was distributed as follows:
| Domain / brand | PAWC share | Rank |
|---|---|---|
| techradar.com | 0.15 | 1 |
| en.wikipedia.org | 0.06 | 2 |
| Salesforce | 0.04 | 3 |
| HubSpot | 0.02 | 7 |
| Klikt | 0.00 | Not mentioned |
Review sites and reference works outranked the vendors themselves, and the smaller brand did not appear in the answers at all.
Takeaways for practitioners
- Optimise sections, not pages, with the answer in the first sentence.
- Ensure accessibility: plain HTML, clear meta descriptions, semantic URLs.
- Don't abandon classic SEO: the candidate set is still built from search engines.
- Be where AI looks, and measure the answers: Reddit, LinkedIn, review sites; SOM and PAWC.
GEO doesn't replace SEO; it extends it. Classic optimisation still decides whether your page makes it into the candidate set. But beyond that, what matters is not your position but how easily your text can be understood, cited and retold. The winner is whoever says the most in the clearest words.