GEO & AEO Playbook

GEO and AEO are the work of getting your company named, and linked, in the answers that ChatGPT, Perplexity, Claude and Google’s AI Overviews and AI Mode write. Most of it is search basics done thoroughly: let the right crawlers in, serve pages they can read, answer buyers’ questions with facts worth repeating, and get mentioned where buyers talk. Then measure it with repeated runs, because one answer tells you very little.

Buyers now ask these tools for shortlists. Google says AI Mode has passed 1 billion monthly users since it expanded worldwide^[1]. And when Google shows an AI summary, people click less: in Pew Research Center’s browsing data from 900 US adults in March 2025, they clicked a standard result on 8% of searches that showed a summary against 15% of those that did not, and clicked a link inside the summary 1% of the time^[2]. Being cited seems to soften the loss. Seer Interactive, an agency, found organic click-through of 0.70% for brands cited in an AI Overview against 0.52% for brands left out, a correlation it cautions may not be cause and effect^[3].

The names overlap. SEO earns a place in ranked results. AEO, answer engine optimization, earns the direct answer, from a featured snippet to a voice reply. GEO, generative engine optimization, earns a citation in an answer a model writes. Google treats optimizing for its AI features as SEO, because they rest on its core ranking systems^[4]. Ranking alone is not enough, though: in Ahrefs’ early-2026 study of 863,000 keywords, 38% of the pages cited in AI Overviews ranked in Google’s top 10 for the same query, though Ahrefs changed its method since an earlier run that found 76%^[5].

This field is full of confident numbers with little behind them. I say where the evidence is thin, and I keep the steps to what you can check yourself. For a longer review of the research, see What Is GEO? What the Evidence Actually Says About AI Citations.

Start by finding out what the engines say today, because every later step is judged against it. Write 15 to 20 questions a buyer would ask before they know your name: “best tool for…” questions, comparisons and alternatives, how-to problems your product solves, and what it costs. Leave your brand out. Asked by name, a model usually knows a company, which flatters you: in a January 2026 preprint, gpt-4o-mini recognized 99.4% of 112 Product Hunt startups by name but named them in only 3.32% of answers to questions about their category, and Perplexity in 8.29%^[6]. Ask about yourself by name separately, to check what engines get wrong.

Then ask each question on the engines your buyers use, several times each, because answers vary from run to run. A University of St. Gallen team that tracked ChatGPT, Gemini, Google AI Mode and Perplexity daily in early 2026 found that the sources cited for the same prompts overlapped by only about a third to two-fifths from one day to the next, and that estimates steadied only at seven or eight runs per prompt per day^[7]. A startup cannot run that much by hand. Five runs per question per engine, in one week, is my working minimum, and it is why you read the result as a range rather than a score.

For every answer, record whether your company is named, whether your site is linked as a source, and who else is named. Report each engine separately, and freeze the question list so later runs compare like with like. Set no target yet: I know of no honest benchmark for a first-month citation rate. The free AI Visibility Significance Test sizes the question set before you run it, and the AI Visibility Starter Kit has a prompt template and a run log.

An engine can only quote a page it can fetch, and each AI company runs separate crawlers for separate jobs (Figure 2). Search crawlers build the index an assistant searches; user fetchers load a page when a person asks about it; training crawlers collect pages for future models. OpenAI says sites that block OAI-SearchBot are not shown in ChatGPT’s search answers, while blocking GPTBot only opts them out of training^[8]. Anthropic splits its crawlers the same way into Claude-SearchBot, Claude-User and ClaudeBot^[9], and Perplexity uses PerplexityBot for search and Perplexity-User for a person’s requests^[10]. Google has no separate AI crawler: AI Overviews and AI Mode draw on pages Googlebot has indexed, and a page must be indexed and eligible to show with a snippet^[4].

Training crawlers are a separate decision. Blocking GPTBot or ClaudeBot keeps your pages out of future training, not out of today’s answers. Let them in if you want future models to know your category; block them if your content is what you sell.

Then check the four places a block can hide. First robots.txt. Second, your CDN or firewall: Cloudflare, which says it handles traffic for about 20% of the web, has started new domains with AI crawlers blocked unless the owner allows them since July 1, 2025^[11], and in July 2026 it announced new defaults from September 15, 2026 that allow search crawlers but block training and agent use on pages that carry ads^[12]. Third, your server logs: a 403 response to a crawler’s user agent means something in front of your site turned it away. Fourth, Search Console, which now has a setting that removes a site from AI Overviews, AI Mode and Discover’s AI features; it is on by default, so check nobody has switched it off^[13]. OpenAI and Perplexity say a robots.txt change takes up to about 24 hours to reach their systems.

Letting crawlers in is wasted if they find an empty page. Many crawlers read only the HTML your server sends and do not run JavaScript. In Vercel’s December 2024 analysis of traffic on its network, none of the major AI crawlers rendered JavaScript, OpenAI’s and Anthropic’s included, while Gemini, through Googlebot, and Applebot did^[16]. That is a single analysis and nearly two years old, and I have found no public update, so test your own pages rather than rely on it.

A single-page app built in React, Vue or a similar framework without server-side rendering sends a near-empty shell and builds the page in the browser. The test takes a minute: fetch the page without a browser, with curl -s https://yourdomain.com/pricing, and look for your headline, prices and main answer in what comes back. If they are missing, render on the server or prerender to static HTML, starting with the pages buyers ask about: pricing, product, comparisons and the guides in Step 6. Keep facts out of tabs and accordions that load on click, and out of logged-in areas.

Then make sure the pages are in the indexes the engines search. Google’s AI features draw on its Search index, so submit a sitemap in Search Console and fix what its reports flag. Do the same in Bing Webmaster Tools: Bing’s index feeds Microsoft Copilot and Bing’s AI summaries, and since February 2026 its AI Performance report shows how often your pages are cited there^[17]. Turn on IndexNow, which tells participating search engines, Bing among them, when a page is added, changed or removed. OpenAI does not say which outside search providers ChatGPT uses, so for ChatGPT the requirement you control is OAI-SearchBot’s access in Step 2. Keep pages fast as well: Google’s Core Web Vitals call a page good when it shows its main content within 2.5 seconds, responds to input within 200 milliseconds and shifts its layout by 0.1 or less, for three in four visits^[18].

Structured data helps engines read who you are and what you sell, but it is not a switch for AI answers. Google says structured data is not required for its generative AI features, that there is no special schema.org markup to add for them, and that it is still worth using as part of ordinary SEO^[4]. So mark up what is on the page: Organization on the homepage, with your name, logo, address and sameAs links to your official profiles; Article with a named author on every article; and Product or SoftwareApplication with an Offer where you publish prices. Google’s rule is simple: do not mark up content that readers cannot see on the page^[19]. Check each type with the Rich Results Test and the Schema Markup Validator.

Two types the old advice leans on no longer show anything in Google. How-to rich results were removed in September 2023, and FAQ rich results, by then limited to government and health sites, stopped appearing on May 7, 2026^[20]. Keep a FAQ section wherever buyers have real questions, because people and engines read the text; the markup on its own earns nothing in Google.

llms.txt is a 2024 proposal by Jeremy Howard for a Markdown file at the root of a site that lists its key pages for language models^[21]. It is not a standard, and no major AI search engine has said it reads one. Google says Google Search ignores these files, so making one neither helps nor harms you there. Almost nothing else reads them either: in Ahrefs’ May 2026 request logs, 97% of about 38,000 llms.txt files got no requests at all, and search crawlers from OpenAI, Perplexity and Anthropic made about 1% of the requests the rest received (Figure 4; Ahrefs, 2026^[22]). The named AI tools that did fetch them were led by GPTBot and Claude Code, a coding agent, so the file is most useful for developer documentation. Otherwise it costs ten minutes and does no harm: decide, write the decision down and move on.

When an engine answers, it lifts a passage, not a page, so each section should stand on its own. The pattern I use: a heading phrased as the question a buyer asks, the direct answer in the first sentence, then the evidence (a number, a date, a named source with a link), then what to do. That is good practice for readers and for extraction. No study I trust has measured it as a lift on its own.

The research on rewriting content for AI is thinner than the advice built on it. The 2024 GEO paper from Princeton and IIT Delhi found that adding quotations gave a source about 41% more of a GPT-3.5 answer, and adding statistics about 31%, but only for pages already placed in the model’s context; it did not test whether rewriting helps a page get found^[23]. A February 2026 preprint from Yonsei University rebuilt the whole pipeline over 171,003 documents and found that rewriting body text for citation made pages less visible at every stage, while rewriting only the title, meta description, headings and structured data helped them get retrieved (Figure 5; Kim and colleagues, 2026^[24]). So put the effort into what a retrieval system reads first: a title and headings that match the question, a description that states the answer, and a first sentence that gives it.

Where a claim needs a number, give the number, its source and its year. A vague claim gives an engine nothing to quote, and an invented one is worse than none. Keyword stuffing does not help either: in the GEO paper it cut a source’s share of the answer by about 8%.

An engine has to cite someone for a fact, and the surest way to be that someone is to publish a fact nobody else has: data from your product, a survey of your customers, a benchmark with its method written down. No study ranks content formats by how often they are cited, so the order I use is judgment, not measurement: original data first, then the definitions and comparisons buyers ask for, then how-to guides. Opinion pieces seldom give an engine anything to cite.

Build the rest around it: one pillar page that answers the category question in full, and five to seven supporting pages (a how-to, a comparison, a pricing explainer, a checklist, common mistakes, a glossary), each linking to the pillar and to the data, and both linking back. A single page is easy to miss; a linked set on one topic gives engines several ways in. Put a named author on each page and refresh the data on a schedule. Microsoft’s advice for its own AI answers is that regular updates help AI systems cite the current version of your content^[17].

Engines name companies that other sources name. Profound, a company that sells AI visibility tracking, counted about 680 million citations from August 2024 to June 2025: Wikipedia was ChatGPT’s most-cited source, Reddit led on Perplexity and on Google’s AI Overviews, and YouTube was close behind on both (Figure 7; Profound, 2025^[25]). It is one vendor’s data for one period, and engines change what they cite, but the pattern is clear enough: answers lean on communities, reference sites and video, not only on company websites. The preprint in Step 1 points the same way: on Perplexity, the number of other sites linking to a startup was the strongest predictor of being named, and the paper’s content score predicted nothing^[6].

So be present where your buyers already talk. Answer questions in the communities they use, as yourself and saying who you work for; most communities ban promotion, and a removed post helps nobody. Ask customers for reviews on the review sites in your category. Pitch guest articles and podcast appearances, and publish the transcripts on your own site. Grok can search posts on X as well as the web^[26], so if your buyers are on X, post there too.

Then make it easy to tell that all those mentions are one company. Write one sentence that says what you do and for whom, and use it word for word on your site, in your Organization markup, on LinkedIn, on review and directory profiles and in your press kit, with the same company name, founder names and category words everywhere.

Measure in three layers, because none is enough alone. First, the answers: re-run the frozen questions from Step 1 on the same engines with the same number of runs, and compare shares, not single answers. Second, the engines’ own reports: since August 31, 2026, Search Console’s generative AI performance report has shown how often links to your pages appear in AI Overviews and AI Mode, counting impressions, not clicks^[27], and Bing’s AI Performance report counts citations in Copilot and Bing’s AI summaries. Third, your logs and analytics: crawler requests appear only in server or CDN logs, because crawlers do not run analytics tags, and visits from assistants appear in GA4, where ChatGPT adds utm_source=chatgpt.com to the links it sends^[28]. Branded search is worth a glance too, but as context: I know of no study that ties it to AI citations.

Be honest about noise. Repeated runs of the same question are not independent, and engines change week to week, so a small rise can be chance. The AI Visibility Significance Test tells you whether a change is bigger than your sample could produce by chance. After the first 60 days (Figure 8), keep it running in an hour a month: re-run the questions, read the three layers, check the logs for new 403s, and choose one page to refresh and one to write.

What changes by stage?

What mistakes should I avoid?

• Blocking answer crawlers by accident. Check robots.txt, the CDN, the logs and Search Console (Step 2).
• Letting crawlers in to an empty JavaScript shell. Fetch each key page without a browser (Step 3).
• Treating llms.txt or FAQ markup as a lever. Neither moves Google; spend the time on access and content (Step 4).
• Rewriting body copy to sound quotable. Fix titles, headings and first sentences, and add sourced facts (Step 5).
• Reading one run as a result. Repeat each question and compare shares (Steps 1 and 8).
• Describing your company differently everywhere. Use one sentence, word for word (Step 7).