Generative Engine Optimization: A Publisher's Guide

Generative Engine Optimization (GEO) is the practice of structuring your content, entities and technical signals so that AI answer engines such as ChatGPT, Google AI Overviews and AI Mode, Gemini, Perplexity and Copilot cite, quote or recommend your pages inside a generated answer, rather than optimising to win a blue link. It is measured by your citation or mention share inside AI answers, not by classic rank position. For publishers, GEO matters because AI answers increasingly sit between a search and a click, and the research now shows that gap costs real traffic. This guide covers what GEO is, how it differs from SEO, how to make your pages quotable, how to control which AI bots reach your content, and how to measure whether any of it is working.

What is generative engine optimization?

Generative Engine Optimization is optimising to be the source an AI answer is built on. When a reader asks ChatGPT or Perplexity a question, or sees a Google AI Overview, the engine synthesises an answer and, usually, names a handful of sources. GEO is the work of becoming one of those named sources: making your content easy for the model to extract, quote and attribute.

The term comes from an academic paper, "GEO: Generative Engine Optimization" (Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024). The study tested which content changes raised a source's visibility inside generative-engine responses across a benchmark of real queries. It found that GEO methods could boost a source's visibility by up to about 40 percent, that the effective methods were content changes such as adding citations, quotations and statistics rather than keyword tactics, and that the efficacy of each method varied across query domains (Aggarwal et al., arXiv / KDD 2024). That finding is the mechanistic backbone of almost all credible GEO advice: make pages quotable, statistic-dense, well-sourced and structured so a model can lift a clean passage.

The other half of GEO is technical. AI engines only cite what they can reach and are allowed to use, and different engines reach your content through different crawlers with different rules. So GEO is partly editorial (be quotable) and partly infrastructural (be reachable and citable, on your terms).

How is GEO different from SEO?

SEO optimises for a ranked list of links. The unit of success is a position, and the goal is a click. GEO optimises for inclusion inside a synthesised answer. The unit of success is a citation or mention, and the click is optional and often absent.

Three differences matter most in practice.

• The target output is different. SEO wants position 1. GEO wants to be quoted. In Google Search Console, an entire AI Overview occupies a single position, usually position 1, and every link inside it shares that one position (Google Search Console Help). Ranking logic does not map cleanly onto a block that collapses many sources into one slot.
• The winning tactics are different. The GEO study found keyword-style optimisation transferred poorly to generative engines, whereas cited sources, direct quotations and statistics were the higher-impact levers, with effects varying by query domain (Aggarwal et al., KDD 2024). Writing for a crawler's keyword match and writing for a model's extraction are not the same craft.
• Access is controllable per engine. In SEO you largely deal with one Googlebot. In GEO you deal with a fleet of separately named bots, each with its own robots.txt token and its own purpose, which you can allow or block independently.

The overlap is real, though. Both reward genuine expertise, clear structure and trustworthy sourcing. GEO is not a replacement for SEO. It is a second optimisation surface layered on top of the same content, with its own measurement and its own access controls.

How do I control which AI bots read my content?

This is the part publishers most often get wrong, because AI engines reach your content through two distinct crawler classes that must be handled separately in robots.txt:

• Training and indexing crawlers that feed model training or an AI's own index: GPTBot, Google-Extended, ClaudeBot, CCBot and PerplexityBot.
• Live, user-triggered fetchers that pull a page when a person asks a question right now: ChatGPT-User, OAI-SearchBot, Claude-User and Perplexity-User.

The crucial point: blocking a training crawler does not necessarily remove you from an AI's live search citations, and blocking a search fetcher does not necessarily stop training. So a blanket block can cost you citations, and a blanket allow can feed your content into training you did not intend. You need to decide those two things separately.

Here is how the main players break down.

A few details worth knowing per vendor:

OpenAI runs GPTBot (training), OAI-SearchBot (ChatGPT search citations), ChatGPT-User (user-triggered fetch) and OAI-AdsBot (ad-safety validation), each independently controllable. OpenAI publishes per-bot IP-range JSON files, for example openai.com/gptbot.json and openai.com/searchbot.json, so you can verify a request is genuine (OpenAI crawler docs).

Anthropic runs ClaudeBot, Claude-User and Claude-SearchBot, all of which honour robots.txt and support Crawl-delay. Anthropic publishes verifiable crawler IP ranges and advises against IP-blocking, because it can stop the bot from reading your robots.txt, which is the very mechanism you use to opt out (Anthropic Help Center).

Perplexity runs PerplexityBot, which surfaces and links sites in results and does not crawl for training, and Perplexity-User, which visits on behalf of a live user. Perplexity states that because Perplexity-User fetches are user-initiated, that fetcher "generally ignores robots.txt rules" (Perplexity crawler docs). Cloudflare has separately reported observing Perplexity use undeclared, disguised crawling to evade no-crawl directives, so a robots.txt block on Perplexity is not a guarantee.

Google-Extended is a robots.txt control token announced in September 2023, not a separate crawler. It governs whether your content trains Gemini and grounds Gemini answers. Google states that blocking Google-Extended does not affect a site's inclusion in Google Search and is not used as a ranking signal (Google crawler docs). Google-CloudVertexBot covers site-owner-requested Vertex AI crawls, and GoogleOther handles one-off research crawls.

Common Crawl's CCBot (user-agent "CCBot/2.0") builds the open corpus that has fed nearly every major LLM. It respects robots.txt and runs on dedicated IP ranges with reverse DNS for verification (Common Crawl CCBot). One caveat: Common Crawl's archives run back many years, predating widespread AI blocking, so historical content may already sit in training sets regardless of what you block today.

A sensible default robots.txt posture

Most ad-funded publishers want the same thing: stay citable in live AI answers, but limit feeding raw content into model training. That maps to a clear split.

• To limit training, disallow the training tier: GPTBot, Google-Extended, ClaudeBot, CCBot and PerplexityBot in its indexing role. For each, use a User-agent: line naming the bot followed by Disallow: /.
• To stay citable, leave the search and user tiers allowed: OAI-SearchBot, Claude-SearchBot, ChatGPT-User, Claude-User and Perplexity-User. Blocking these is what quietly removes you from AI citations.

This is a genuine trade-off, not a free win, and reasonable publishers land in different places. The point of GEO is that you make the decision deliberately, per bot, rather than accepting whatever a blanket rule does to you.

How do I make my content easier for AI to cite?

The GEO study gives an evidence-based hierarchy: adding cited sources, direct quotations and statistics were the stronger levers for generative-engine visibility, and classic keyword optimisation was weak (Aggarwal et al., KDD 2024). Translate that into editorial practice.

• Make pages statistic-dense and sourced. A model synthesising an answer prefers a page that hands it a concrete, attributable number over one that gestures vaguely. Include real figures, name the source, and link it. This article does exactly that, on purpose.
• Write quotable passages. Short, self-contained sentences that state a fact cleanly are easier to lift than long, hedged paragraphs. Lead each section with a direct answer, then support it.
• Structure for extraction. Use clear question-shaped headings, tight paragraphs, and tables or lists where they genuinely help. A model, like a reader, extracts more reliably from a page with visible structure.
• Add citations and quotations. Quoting a named authority and citing sources both raise perceived trustworthiness and give the engine attributable material. These were among the highest-impact tactics in the study.
• Keep entities clear. State plainly what a thing is, who runs it, and when. Ambiguity is the enemy of citation.

Notably, the study found the efficacy of each method varied by query domain, so the right mix differs between, say, a health explainer and a product comparison. Test rather than assume. And nothing here excuses thin content dressed in statistics: the tactics amplify a genuinely useful page, they do not rescue a hollow one.

How do I measure whether GEO is working?

You cannot manage GEO from your rank tracker. Measurement splits across three surfaces, and each has a catch.

Citation and mention share. The headline GEO metric is how often AI answers cite or mention you for the queries you care about, ideally as a share against competitors. This is a different KPI from position, and you have to track it by sampling AI answers for your target questions over time rather than reading it off Search Console.

Google Search Console. Understand how AI features are counted before you draw conclusions. An AI Overview occupies one position, usually position 1, and every embedded link shares that position. AI Overview and AI Mode impressions and clicks are aggregated into the standard Performance report by default, with same-URL de-duplication, so two results from your site inside one AI feature count as one impression at property level. Google's separate "Generative AI features" report, a Beta launched around June 2026, reports impressions only, not clicks, rolls out gradually, has a minimum-traffic threshold and excludes Search Labs (Google Search Console Help). In short, GSC will tell you that you appeared in AI features, but it is coy about what that earned you.

GA4 and referral traffic. When an AI answer does send a click, you want to see it. GA4 added a native "AI Assistant" channel in May 2026 that auto-tags sessions from recognised AI domains such as ChatGPT, Gemini and Claude with the medium "ai-assistant". But it misses Perplexity and Copilot, and any session where the AI app strips the referrer header, which is common in mobile apps and some in-app clicks, falls back to Direct. So native detection substantially under-counts AI traffic. The reliable fix is a custom referral regex covering all AI hosts, placed above Referral in the channel ordering so it captures those sessions before the generic Referral rule does.

Two practical notes. First, referrer-stripping means a chunk of AI-driven traffic will always hide in Direct, so treat your AI number as a floor, not a ceiling. Second, because the surfaces disagree, the useful discipline is to watch the trend on each rather than chase a single perfect figure. If citation share is rising, GSC AI impressions are growing and your AI-referral channel is climbing, GEO is working, even if no one number is exact.

If you want a clean read of your existing GA4 with AI referrals, AI crawler activity and AI citation trends broken out without adding any tracking script, you can try Ramprt free on the live demo. It reads your GA4 read-only and shows what AI is doing to your traffic.

Does GEO help or hurt my normal search traffic?

Two separate questions live inside this: does the GEO technical work hurt SEO, and does the rise of AI answers hurt clicks regardless of what you do?

The technical work is largely safe for SEO. Google states that blocking Google-Extended does not affect a site's inclusion in Google Search and is not used as a ranking signal (Google crawler docs). So you can make training-tier decisions without touching your organic rankings. And the editorial levers of GEO, being well-sourced, statistic-backed and clearly structured, are exactly what strong SEO content already does. There is little tension there.

The rise of AI answers is a real headwind, and it is measured. This is the part publishers should not soften. In the July 2025 Pew Research study of 900 US adults across 68,879 unique searches, users who saw an AI summary clicked a traditional search-result link on only 8 percent of visits, versus 15 percent when no AI summary appeared, and clicked a link inside the AI summary itself on just 1 percent of visits. Browsing sessions ended after 26 percent of pages with an AI summary, versus 16 percent of pages with only traditional results, and about 18 percent of searches produced an AI summary (Pew Research Center). Ahrefs, analysing 300,000 keywords, found the presence of an AI Overview correlated with a 58 percent lower average click-through rate for the top-ranking page, and that position-1 CTR for AI-Overview keywords fell from 0.073 in December 2023 to 0.016 in December 2025 (Ahrefs).

So GEO does not cause that loss, but it also cannot fully reverse it. What GEO does is change your position within the new reality. When an AI summary cites three or more sources, which Pew found in 88 percent of cases, and the median summary runs just 67 words, you want to be one of the cited three rather than one of the uncited many. Being quotable is how you salvage visibility, and some referral traffic, from a results page that now clicks through far less often.

The honest summary: the technical GEO decisions are neutral-to-positive for SEO, the editorial ones are positive, and the broader AI-answer shift is a structural pressure on clicks that GEO helps you survive rather than defeat. The right response is to measure net clicks and citation share together, optimise to be cited, and stop treating an unclicked impression as a win.

Frequently asked questions

Is GEO the same as answer engine optimization or AI SEO?

They describe the same idea under different names: optimising to be cited or recommended inside AI-generated answers rather than to rank a link. "Generative Engine Optimization" is the term from the founding academic paper (Aggarwal et al., KDD 2024), so it is the most precise label.

If I block GPTBot, do I disappear from ChatGPT's answers?

Not necessarily. GPTBot is OpenAI's training crawler. ChatGPT's live search citations come through OAI-SearchBot and ChatGPT-User, which are controlled separately. Blocking GPTBot limits training use but can leave you citable in ChatGPT search, provided you leave the search and user fetchers allowed.

Can I stay out of AI training but still be cited in AI answers?

Largely, yes, and that split is the whole point of per-bot control. Disallow the training tier (GPTBot, Google-Extended, ClaudeBot, CCBot, PerplexityBot's indexing role) while allowing the search and user tier (OAI-SearchBot, Claude-SearchBot, ChatGPT-User, Claude-User, Perplexity-User). Two caveats: Perplexity-User "generally ignores robots.txt", and old content may already sit in Common Crawl archives that predate your rules.

Why does GA4 under-report my AI traffic?

GA4's native AI Assistant channel (May 2026) only recognises some AI domains, notably ChatGPT, Gemini and Claude, and misses Perplexity and Copilot. Worse, many AI apps strip the referrer, so those sessions land in Direct. A custom referral regex covering all AI hosts, ordered above Referral, recovers most of the gap.

Does blocking Google-Extended hurt my Google rankings?

No. Google states that blocking Google-Extended does not affect a site's inclusion in Google Search and is not used as a ranking signal. It only governs whether your content trains and grounds Gemini.

Related reading: see what AI Overviews are doing to publisher traffic and how to track AI referral traffic in GA4.