GEO (generative engine optimization): the fundamentals of getting cited by AI

GEO, short for generative engine optimization, covers everything you do so that a brand, a site or a page gets picked up by the AI tools that answer questions directly: ChatGPT, Gemini, Perplexity, Copilot, Claude, or Google's AI Overviews and AI Mode. The goal is no longer just to appear in a list of links, but to be cited in the answer, named as a solution or used as a source.

The term comes from a research paper published in late 2023 by researchers at Princeton University and IIT Delhi, later presented at the KDD 2024 conference. You'll also come across AEO (answer engine optimization), LLMO or “SEO for AI”: the words change, the idea stays the same.

GEO doesn't start from scratch. AI tools that answer questions about news or products fetch pages from the web at the moment of the question, then summarize them: a site that's invisible to search engines is almost always invisible to them too. That's why people often talk about “GEO SEO”: classic SEO remains the foundation, and GEO adds a floor on top.

A ready-made answer at the top of the page reduces clicks to websites. In March 2025, the Pew Research Center tracked the Google searches of 900 U.S. adults: with an AI-generated summary, a traditional link was clicked in only 8% of visits, versus 15% without a summary, and links inside the summary were clicked in just 1% of cases. These summaries then appeared on 18% of searches, a share that has grown since with AI Mode.

So the visit is no longer the only moment a customer discovers a brand. They read the answer, remember two or three names, and come back later by typing one of them directly. Being absent from that answer means losing part of your brand awareness before the first visit even happens.

On its own, a language model only knows the world up to its cutoff date (see our AI glossary). To answer on a recent or specific topic, the assistant first runs searches, retrieves pages, then writes its answer from what it has read, citing some or all of its sources.

Google describes this mechanism in its documentation on AI features: AI Overviews and AI Mode may use a “query fan-out” technique, which splits the question into several searches on subtopics, then draws on a wider and more diverse set of pages than the classic results page. A page can therefore be cited for a question on which it wouldn't rank on page one.

Three steps, six workstreams. The three cards above sum up the approach; the six workstreams below detail it, from the most technical to the most editorial. The first three are simply SEO done right; the next three are more specific to GEO.

1. Let the right crawlers in. AI tools don't all use the same crawler, and each one is controlled separately in robots.txt. At OpenAI, for example, OAI-SearchBot powers ChatGPT search while GPTBot collects content for training: blocking the second doesn't remove a site from answers, blocking the first does, according to OpenAI's documentation.
2. Get the SEO basics right. Google says it in black and white: no special optimization is needed for its AI features. A page must be indexed and eligible to show with a snippet; crawlable internal links, text-based content, a good page experience and structured data that matches the visible text remain the rule.
3. Make every page easy to summarize. A title that announces the answer, a first sentence that gives it, subheadings phrased as questions, lists and tables for comparisons: what helps a reader in a hurry also helps a model that breaks the page into chunks.
4. Bring evidence. In the founding GEO study, the most effective methods were adding statistics, expert quotes and sources, with up to 40% more visibility in generated answers; keyword stuffing, on the other hand, did nothing or even hurt. Lower-ranked sites benefited the most. The study was, however, run on a generative engine built by the researchers, not on ChatGPT or Gemini: treat these gaps as orders of magnitude and test on your own questions.
5. Exist beyond your own site. Up-to-date business profiles, presence in your industry's comparisons and directories, customer reviews, press articles, expert contributions: AI tools cross-check these sources to decide which brands to recommend.
6. Keep information up to date. Prices, offers, hours and specs must be identical everywhere. An AI that finds two conflicting versions picks one of them, not necessarily the right one.

The llms.txt file. Proposed in 2024 to guide models toward a site's important pages, it isn't used by Google: Gary Illyes stated in July 2025 that Google doesn't support it, and John Mueller still calls it “speculative” in 2026, according to Search Engine Roundtable. Creating one does no harm, but don't expect anything from it.

“AI-specific” markup. Google states that no special file, tag or structured data is required for its AI features. Structured data remains useful for rich results, as long as it describes what the page actually shows.

Keyword stuffing and copy written for bots. Repeating the question in every possible form doesn't help a model cite you, and it hurts your readers.

GEO doesn't mean opening everything up. You can agree to be cited in answers while refusing to let your content be used to train models: these are two separate settings.

The separation has a limit. When an assistant answers without running a search, on a generic question like “which SEO agencies do you recommend?”, it draws on what it learned in training: a brand absent from that data, because it blocked training crawlers or is rarely mentioned on the web, won't appear. Blocking GPTBot is a legitimate choice, but it comes at a cost on this kind of answer.

A control shared across several providers is on its way: in late October 2026, Google announced that it would recognize a single robots.txt line, based on IETF work, to opt out of training, with support targeted for 2027. Microsoft has adopted the same syntax.

Measurement is improving fast but remains incomplete: answers change from one run to the next and not all platforms share their numbers. So you combine three sources.

• Search engine tools. Since 2026, Google Search Console has offered reports dedicated to generative AI (impressions by page, country and device, without clicks), opened to all sites in late August. Bing Webmaster Tools, for its part, measures a site's citation share in Copilot and Bing answers.
• Your analytics tool. Visits from chatgpt.com, perplexity.ai or gemini.google.com show up as referring sites; group them into an “AI” segment to track how they evolve and convert (ChatGPT also adds a utm_source=chatgpt.com parameter that makes them easier to spot). Clicks from AI Overviews and AI Mode, however, arrive as ordinary Google organic traffic and can't be isolated in your analytics tool: only Search Console separates them.
• A GEO tracking tool. Platforms like Peec AI, Otterly.AI or Profound ask your key questions to the assistants every day and record whether your brand is cited, in what position and against which competitors. Our comparison of GEO tools sorts them by use and by price.

Allow several months before judging. Technical changes take effect as soon as pages are recrawled, within days or weeks, but presence on third-party sites and brand awareness build slowly: track your key questions every month and read the trend rather than a single snapshot.

• robots.txt allows Googlebot, Bingbot and OAI-SearchBot; training crawlers are blocked or not, as you choose.
• Key pages are indexed in Google and Bing, with no unintended nosnippet tag.
• Every important page answers a specific question from its first lines.
• Figures, dates and claims are sourced and up to date.
• Your brand is present and consistent on the third-party sites that matter in your industry.
• Structured data accurately describes what the page shows.
• An “AI” segment tracks visits from assistants in your analytics tool.
• A list of 20 to 50 key questions is tracked every month, by hand or with a GEO tool.