Generative Engine Optimization (GEO), Explained Honestly

A generative engine writes an answer instead of returning a list of links. ChatGPT search, Perplexity, Claude with web search, Microsoft Copilot, Google AI Overviews and Google AI Mode all work this way. Generative engine optimization is the work of influencing whether those engines use your content and name your brand.

The goal overlaps almost entirely with answer engine optimization. The labels differ, the target is the same. One naming trap: in local SEO, "geo" means geographic targeting, which is a different discipline. If that is what you searched for, read GEO SEO: generative engines or geographic search.

In practice, GEO means two different things:

• The narrow research claim. Rewriting a document can change how often a model cites it, once that document is already in front of the model.
• What the market sells. Techniques that supposedly make AI engines find, choose and recommend you.

The evidence supports the first meaning. It does not support the second. Keep that split in mind when you read any GEO pitch.

The term comes from Aggarwal et al., "GEO: Generative Engine Optimization" (arXiv 2311.09735, published at KDD 2024). It is a careful paper. It is also widely misquoted.

Here is what the authors actually did:

• They built their own two-step answer engine and supplied the source documents directly in the model's context.
• Their one "in the wild" Perplexity test uploaded source text as files, because Perplexity did not let users specify source URLs.
• Their main metric, Position-Adjusted Word Count, counts words from a source that appear in the answer. That includes words the optimization itself added, which makes the measure partly circular.
• They reported a 30 to 40% band across three methods. The itemised figures that circulate in GEO marketing are not in the paper.
• Keyword stuffing made things worse: the paper scored it at 17.8 against a 19.5 baseline.

What this means: "GEO gets you up to 40% more visibility" is a lab result about content the engine already had. It says nothing about whether a live engine finds your page. Our digest of the original GEO paper walks through the method in more detail.

An AI answer passes through four stages. Most GEO confusion comes from mixing them up.

Nearly every published GEO effect size sits in the generation stage. Those studies assume retrieval already succeeded. A rewrite that lifts citations among five pre-loaded documents tells you nothing if the engine never fetched your page.

Retrieval is also discretionary. As of October 2026, ChatGPT decides per query whether to search, Google AI Overviews often do not trigger at all, and web search in Claude.ai is a user toggle. Any answer you observe may come from live retrieval or from the model's frozen memory. See lab GEO studies vs live AI engines for what carries over and what does not.

In July 2026, a critical survey reviewed 45 GEO studies published between 2023 and 2026 (arXiv 2607.14035). Its central conclusion: "already-retrieved content can causally alter its citation or use, but no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior."

Two other studies point the same way:

• C-SEO Bench (NeurIPS 2025) found that most conversational SEO methods were "largely ineffective" and frequently hurt a document's ranking in the model's context.
• A 252,000-trial factorial study (arXiv 2605.25517) found that formatting-only edits had little impact. Topical relevance and position in the candidate list drove citation most.

Read these with their limits. The survey is single-author and not peer-reviewed. The factorial study was written by a company that ships a GEO checklist and used a synthetic corpus. C-SEO Bench's "traditional SEO" arm simply moved the target document to first place in the context, which no publisher can do. Agreement between these sources is corroboration, not replication.

The practical reading is still clear. Topical fit matters. Formatting tricks and citation-bait rewrites do not have evidence behind them. Our digest of the critical survey covers the full study list.

In your control

• Rendering. Content that only appears after JavaScript runs may be invisible to most AI crawlers.
• Per-bot robots.txt rules. Each engine uses its own crawler names, and the rules do not transfer between them.
• Network reachability. A firewall or CDN can block an AI crawler that robots.txt allows. From 15 September 2026, Cloudflare blocks Training and Agent crawler categories by default on ad-displaying pages for new domains and existing free-tier customers, so access can change without you editing anything.
• Suppression controls. noindex, nosnippet, max-snippet:0 and data-nosnippet keep content out of AI answers, including when applied by accident.
• Plain facts on fetchable pages. Stating a price, a number or a one-line definition gives an engine a sentence it can quote.
• Presence on sites engines already cite. When engines cite third-party pages instead of yours, being named there is associated with appearing in the answer. It is an association, not a guarantee.

Not in your control

• What a model learned before its training cutoff.
• How the engine rewrites the user's question into its own search queries.
• Which retrieved page the model chooses to cite.
• Brand size, which shapes citation patterns across every engine studied.

The access levers are subtractive. They remove blocks. Nothing on this list makes a page more likely to be chosen once it is reachable.

Work from the first gate outward. A fix at a later stage is wasted if an earlier gate is closed.

1. Check rendering. Compare your raw HTML with the rendered page. Vercel found that ChatGPT and Claude crawlers fetch JavaScript files (11.50% and 23.84% of their requests) but do not execute them. Googlebot does render, so this gap mostly affects non-Google engines. Put prices, product names and key claims in the initial HTML.
2. Check robots.txt per bot. OpenAI's crawlers are independent: GPTBot governs training, OAI-SearchBot governs ChatGPT search. A site can allow one and block the other:

User-agent: OAI-SearchBot
Allow: /

User-agent: GPTBot
Disallow: /

Check PerplexityBot, Claude-SearchBot and Meta-WebIndexer the same way. 3. Check the firewall. Request a page as each AI crawler and look for 403 responses or challenge pages. The AI Crawler Checker does this per crawler. 4. Remove accidental suppression. Search templates for noindex, nosnippet, max-snippet:0 and data-nosnippet. In Google Search Console, confirm your property has not opted out of generative AI features, directly or through a parent property. 5. Check index status engine by engine. Use Google Search Console and Bing Webmaster Tools separately. In OtterlyAI's 2026 experiment, Google deindexed two AI-generated sites and its AI surfaces stopped citing them, while Microsoft Copilot, which grounds on Bing, kept citing them. 6. State your facts plainly and consistently. Make pricing, founding details and product names match across every page. 7. Then measure. A free GEO audit checks access first, then accuracy. The free checker on this site shows whether AI engines name you for up to three buyer questions, as a one-time snapshot.

Most widely sold GEO advice is contradicted by the engines' own documentation or by controlled tests.

• Adding schema to get cited. Google says structured data isn't required for generative AI search. Ahrefs' matched-control test found AI Overview citations fell 4.6% after pages added schema. FAQ rich results were deprecated in May 2026. Existing markup does no harm, so there is no need to remove it. It is just not a lever.
• Publishing llms.txt. Google says Search ignores it. No other engine documents reading it.
• Chasing an E-E-A-T score. Google says E-E-A-T isn't a specific ranking factor. There is no score to improve.
• Building domain authority. Domain Authority is a Moz metric. No engine computes it.
• Chunking or writing in an "AI-friendly" style. Google rejects both and says there is no ideal page length.
• Allowing GPTBot so ChatGPT can find you. GPTBot governs training only. ChatGPT search uses OAI-SearchBot.
• Tracking an "AI rank". Google's generative AI report exposes impressions only, with no position data to check a rank against.
• Pinging IndexNow for AI citations. The IndexNow spec never mentions AI, and a successful ping only confirms receipt.
• Hidden text or prompt injection. Research shows manipulated content can sway some AI search agents once it is retrieved. It is illegitimate, it varies by model and safety update, and it puts your brand at reputational and policy risk. Do not use it.

For a side-by-side view of the evidence, see on-page GEO rewrites vs access fixes.