Generative Engine Optimization (GEO): Getting Cited by AI

Where the term GEO comes from

The phrase was popularised by a 2023 academic paper titled "GEO: Generative Engine Optimization", written by researchers from Princeton University, IIT Delhi and other institutions. They defined a generative engine as a system that retrieves sources and uses a language model to write a single synthesised answer, and they tested how changes to source text affected how visible that source was in generated answers.

The research was an early experiment on a benchmark, not a description of how any commercial product ranks sources. Since then, GEO has become a marketing label with many competing definitions. Use it as a convenient name for a set of practices, and be sceptical of anyone who claims to know the exact formula.

How generative engines find and cite sources

Most AI search products follow a pattern often called retrieval-augmented generation. The system turns the user's question into one or more searches, retrieves candidate pages or passages from an index, passes the most relevant passages to a language model, and the model writes an answer with links to some of the sources it used.

1. Retrieval: your page must be in the index the product uses and allowed for its crawler. If it cannot be retrieved, nothing else matters.
2. Selection: the system chooses passages that match the question and its sub-questions, which is where query fan-out comes in.
3. Synthesis: the model combines facts from several passages, so a clear, self-contained statement is easier to use than a vague one.
4. Citation: the product links some sources. Providers have not published exactly how they choose which sources to show.

Because each provider uses a different index and model, the same question can produce different sources in ChatGPT, Perplexity and Google. The AI search overview lists which index each product relies on.

Examples of extractable, well-sourced content

Compare two ways of writing the same paragraph on a page about GST registration.

Hard to extract

"There are lots of things to think about when it comes to registration, and every business is different, so it really depends on your situation and you should check with someone."

Easy to extract

"In India, a business supplying goods must register for GST once its annual turnover crosses the threshold set for its state category. The current thresholds are listed on the GST portal; check them before you file, because they have changed in the past."

The second version names the subject, states a specific fact, points to the authoritative source and can be quoted on its own without the surrounding paragraphs. That is the core of writing for generative engines: each passage should be accurate and understandable in isolation. The content SEO guide covers the wider editorial process.

Step by step: applying GEO to an existing page

1. Confirm the page is indexed in Google and Bing and not blocked for search crawlers such as OAI-SearchBot or PerplexityBot.
2. Write a direct, two- or three-sentence answer to the page's main question near the top.
3. Break the body into sections with descriptive headings that match real sub-questions.
4. Replace vague claims with specific facts, and cite primary sources such as official portals, standards or documentation.
5. Add named authorship and a visible updated date, and make sure your organisation is clearly identified, as explained in the entity SEO guide.
6. Look for places where independent sites discuss your topic, such as industry publications, forums and review sites, and contribute genuinely useful information there.

Earning mentions and entity clarity

Generative answers often draw on several sources, including sites other than yours. If trusted publications, forums or comparison articles describe your product accurately, those descriptions can shape what the AI says about you even when your own page is not cited. This is the same work as ethical link and mention building: useful research, expert commentary and real participation in communities. Posting promotional comments or fake reviews breaks platform rules and can damage trust.

Checking citations and the limits of measurement

The most reliable method available is a manual prompt audit: a fixed list of real questions, run on a schedule in each product, with the cited sources recorded in a spreadsheet. Use a logged-out or fresh session where you can, and note the date, product and location.

Running each prompt more than once, on different days, shows how stable the citations are. A source that appears every time is a stronger finding than one that appeared once.

Be honest about the limits. Generated answers vary between runs, users, locations and model versions. A handful of prompts is a sample, not a measurement of total visibility, and third-party "AI visibility" tools also rely on sampled prompts. Combine audits with referral data in analytics and avoid reporting precise share figures you cannot reproduce.

Practical exercise: a 20-prompt citation audit

Pick one topic your site covers. Write 20 questions a real user might ask, mixing definitions, comparisons and how-to questions; keyword research data and People Also Ask boxes are good sources. Run each in ChatGPT search, Perplexity and Google, and record every cited domain.

You are done when you have a spreadsheet showing, for each prompt and product, whether your site was cited, which domains were cited instead, and one improvement you will make to a specific page as a result. Prompt auditing is covered in SkillNewton's AI SEO automation course.

Common mistakes

• Treating findings from the original GEO research as rules that commercial AI products are known to follow.
• Stuffing pages with statistics or quotations purely to look citable, without checking they are accurate and relevant.
• Blocking AI search crawlers in robots.txt while trying to be cited by those products.
• Reporting precise "AI share of voice" numbers from a handful of prompts that give different results on every run.
• Posting promotional comments in communities to influence AI answers, which breaks platform rules.

Frequently asked questions