Generative engine optimization, in 7 steps.
When someone asks an AI assistant a question, it searches the web, reads a few pages and writes one answer. Generative engine optimization (GEO) is the work of making your pages the ones it reads, trusts and cites. Here is what the research found, and what to do.
Published 13 October 2026 · Updated 14 October 2026 · By Visibian
What changed.
A search engine returns a list of links. An AI assistant returns one answer, with a few brand names and sources inside it. Whether your name is in that answer depends less on where your page sits in a list of results. It depends more on whether the assistant can reach the page, understand it and quote it with confidence.
The term comes from a 2024 study, “GEO: Generative Engine Optimization”, by researchers at Princeton University and IIT Delhi, published at the KDD conference (arXiv 2311.09735). They tested changes to web pages against a simulated AI search engine. Adding quotations, statistics and cited sources raised a page’s visibility in the answer by about 30 to 40% on their main measure, and by up to 37% in a smaller test on Perplexity. Keyword stuffing did not help.
Read those numbers as a direction, not a promise. The study used 2023 models (GPT-3.5 for most tests), and the results varied by topic. When every page in a test was improved at once, the page that had been first in search lost between about 20 and 30% of its visibility. Your own results are the ones that count.
1. Let the AI crawlers in.
Nothing else matters if an assistant cannot fetch your pages. Each AI company sends its own crawlers, the programs that fetch pages. Several send separate ones for search, for answering a person in the moment, and for collecting training data, so blocking the training crawler does not have to block the others.
What each company says about its own crawlers, as of 14 October 2026:
- OpenAI: OAI-SearchBot puts pages in ChatGPT search, ChatGPT-User fetches a page when a person asks for it, and GPTBot collects training data.
- Anthropic: Claude-SearchBot improves Claude’s search results, Claude-User fetches a page when a person asks Claude, and ClaudeBot collects training data.
- Perplexity: PerplexityBot builds Perplexity’s search results and is not used to train foundation models; Perplexity-User fetches pages for a person’s question.
- Google: AI Overviews and AI Mode are part of Google Search, which Googlebot crawls. Google-Extended is a separate robots.txt setting for Gemini training and grounding, and Google says it “does not impact a site’s inclusion in Google Search”.
- Microsoft: Copilot sends a search query to the Bing search service and writes its answer from the results, so Bing’s crawler, Bingbot, is the one that matters.
Check 2 places: your robots.txt file, and any firewall or security service in front of your site, which can turn a crawler away even when robots.txt lets it in. Then check that your main text is in the page itself. Many crawlers fetch a page without running its scripts, so text that appears only after scripts run may never reach them.
The free AI Readiness Check shows what ChatGPT, Claude, Gemini and Perplexity can read on one page of your site, in about 20 seconds.
2. Put the answer first.
Assistants lift short passages out of pages. A passage is easy to quote when it answers a question on its own, in plain words, near the top.
- Open the page, and each main section, with a direct answer of 2 or 3 sentences.
- Use headings that ask the question people ask, such as “What does it cost?”, and answer it in the first sentence under the heading.
- Keep paragraphs short, with one idea each, so any one of them makes sense if it is quoted alone.
- Use lists, tables and numbered steps where the content is a list, a comparison or a process.
These habits help human readers too. As of October 2026, public research does not say how much each one moves an AI answer on its own.
3. Back every claim.
The strongest results in the 2024 study came from adding evidence: statistics, quotations and cited sources. A vague claim is easy to skip. A specific, dated, sourced one is easy to repeat.
- Replace general claims with measured ones, giving the number, the date and where it comes from.
- Name your sources in the text, with the year and a link to the original.
- Quote named experts, your own people included, with their role.
- Publish your own data, from your research or your customers’ results. It gives assistants something no other page has.
4. Make it clear who you are.
An assistant has to work out which company a page is about, and whether to trust it. Mixed messages, such as an old company name in one place and a new one in another, make both harder.
- Describe your company the same way everywhere: your site, your profiles, directories and listings.
- Give articles a named author, with a page that says who they are.
- Earn mentions on the sites assistants already read in your field: review sites, industry publications, forums and reference pages. Your own site is only part of what they read.
In Visibian Brand Intelligence, Source Domains shows which sites the answers in your field draw on, and Hallucination Watch shows where an assistant describes you wrongly.
5. Label your pages with structured data.
Structured data is a block of code, usually JSON-LD, that tells machines what a page is: an article, a product, a company, a set of questions. As of October 2026 there is little public evidence that it gets a page cited on its own. Its value is that it states your facts plainly, in a form machines read without guessing.
- Organization and WebSite, for the site as a whole
- Article, with the author, datePublished and dateModified
- FAQPage, only where the page shows the questions and answers
- Product or SoftwareApplication, where that is what the page describes
- BreadcrumbList, for where the page sits on the site
One rule above the rest: the structured data must match what the page shows. A label that describes something a reader cannot see is ignored at best, and misleading at worst. Visibian’s Page Code writes this block for a page, on the Business and agency plans.
6. Answer the questions people ask.
A short questions-and-answers section, in the words people use with assistants, gives them passages that are easy to lift. Keep each answer short and complete, keep the list to the questions that matter, and update the answers when the facts change.
7. Keep it current, and measure it.
Some assistants fetch pages at the moment someone asks (the user crawlers in step 1), so an update can reach an answer quickly. Update your most important pages when something changes, and change the “last updated” date only when the page itself changes.
Then measure. AI answers vary from one try to the next, so one question asked once tells you little. Ask the same questions on a schedule, across the assistants your buyers use, and compare periods with their margins of error before you call anything a change.
Sources.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, “GEO: Generative Engine Optimization”, KDD 2024.
- OpenAI, “Overview of OpenAI crawlers”, read 14 October 2026.
- Anthropic, “Does Anthropic crawl data from the web, and how can site owners block the crawler?”, read 14 October 2026.
- Perplexity, “Perplexity crawlers”, read 14 October 2026.
- Google, “Google’s common crawlers”, last updated 14 July 2026.
- Microsoft, “Data, privacy, and security for web search in Microsoft Copilot and Microsoft Copilot Chat”, read 14 October 2026.
See where you stand.
Visibian Brand Intelligence asks your buyers’ questions to 7 AI assistants, every day or every week. It shows how often you are named, why answers leave you out, and whether a change you made moved the numbers. The Visibian Academy teaches the discipline behind all 7 steps, with 8 Build Guides for the work itself.