How does generative engine optimization work? It raises your odds at five stages an AI engine runs before it cites anyone: deciding whether to search, fanning the question into sub-queries, retrieving pages its crawlers can reach, selecting quotable passages, and synthesizing an answer with attribution. You control access, passages and proof, not the model.
That is the whole mechanism in one paragraph. The rest of this guide walks each stage with the primary-source documentation behind it and the numbers we measured ourselves on 2026-09-26: an AI visibility check across 10 of our own product sites, and a calibration of our GEO scorer on 45 pages (23 cited by ChatGPT or Google AI Overviews, 22 ranking on Google but not cited) across 12 queries in 12 niches. If you want the definitions first, read GEO vs SEO and SEO vs AEO vs GEO. This piece is about how the machine actually works.
What are the 5 stages of generative engine optimization?
GEO is five separate filters, and a page has to survive all of them in order. Most guides list tactics without saying which stage a tactic acts on, which is why teams add schema or an llms.txt file and see nothing move. Map every change you make to one of these stages first.
| Stage | What the engine does | What you can influence | How to check it |
|---|---|---|---|
| 1. Retrieval trigger | Decides whether to search the web or answer from training data | Almost nothing directly; queries needing fresh or specific facts trigger search more | Run the prompt and see whether sources appear |
| 2. Query fan-out | Splits the question into several related sub-queries | Cover the sub-questions a real expert would answer on one page | Bing Webmaster Tools grounding queries; ChatGPT activity panel |
| 3. Retrieval | Pulls candidate pages from a search index its crawlers built | Crawler access, indexing, ranking in Google and Bing | robots.txt test, Search Console URL Inspection, Bing Webmaster Tools |
| 4. Passage selection | Picks the chunks that best answer each sub-query | Answer capsules, specific numbers, quotes, structure | Our scorer calibration below; your own prompt tests |
| 5. Synthesis and attribution | Writes the answer and decides which sources to show | Distinct, checkable facts; consistent brand description | Search Console generative AI report, Bing AI Performance, GA4 referrals |
The reason this framing matters: <mark class="km-highlight" style="--hl:#FEF08A;background:#FEF08A">a failure at an early stage makes every later optimization worthless</mark>. A beautifully written answer capsule on a page that OAI-SearchBot cannot fetch will never be cited by ChatGPT search.
Stage 1: Does the AI engine search the web at all?
No, not always, and when it does not search, no page can be cited. For many prompts an assistant answers from what it learned in training and shows zero sources. GEO only acts on the prompts where the engine decides to retrieve.
We measured this directly. On 2026-09-26 we took 10 informational queries where our own pages already rank in Google's top 10 or 11 and ran each through ChatGPT. <mark class="km-highlight" style="--hl:#BBF7D0;background:#BBF7D0">ChatGPT browsed the live web for only 4 of the 10 queries</mark>. For the other 6 it answered from training data with no citations at all, regardless of who ranked where. In the same test, Google's AI Overview appeared on all 10 results pages, because Google's AI features are built on top of live Search retrieval.
What this means in practice:
- Queries that need current or specific facts tend to retrieve more. Prices, versions, dates, statistics, "best X for Y" and comparison prompts are the ones most likely to push an engine to search. Stable definitional questions often do not.
- You cannot force the trigger. The lever you do control is being the obvious source when retrieval happens, and being present in training data over time by being widely mentioned.
- Test before you optimize. If your target prompt never shows sources in ChatGPT, rewriting the page for ChatGPT citation is effort spent on a stage that never runs. Google AI Overviews, which always retrieve, may be the better target for that query.
Stage 2: How does query fan-out work, and should you write for it?
Query fan-out is the engine splitting one question into several related searches, then answering from the combined results. Google documents it by name in its guide to optimizing for generative AI features, describing it as "a set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query."
Google's own example: for "how to fix a lawn that's full of weeds", the fan-out queries "might include 'best herbicides for lawns', 'remove weeds without chemicals', and 'how to prevent weeds in lawn'." Each sub-query retrieves its own results. A page that answers two of those sub-questions well has two chances to be pulled in, which is why a thorough page often beats a page that only matches the head query.
There is a trap here. The same Google document warns that creating separate content for every possible variation, "for example, by focusing on other queries that people have asked, or fan-out queries", done "primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy." Write one genuinely complete page that answers the sub-questions an expert would expect, not 30 thin pages for 30 sub-queries.
How to see real fan-out queries for your own site:
- Bing and Copilot (documented). Bing Webmaster Tools added an AI Performance report in February 2026. Its "grounding queries" view, in Bing's words, "shows the key phrases the AI used when retrieving content that was referenced in AI-generated answers." That is fan-out data for Copilot, tied to your own cited URLs. Bing added intent and topic grouping plus citation share in June 2026.
- ChatGPT (undocumented). When a reasoning model searches, the activity panel beside the answer lists the searches it ran. Practitioners also inspect the conversation request in the browser's developer tools (Network tab) for the search queries the model generated. None of this is a documented interface, field names change, and it shows one conversation at a time, so treat it as a spot check, not a dataset.
- Google (not exposed). Google does not show fan-out queries. Its generative AI performance report in Search Console, rolled out to all sites on August 31, 2026, shows impressions from AI Overviews and AI Mode by page, device and country.
Copy this into a doc and fill it in for your target query before you rewrite anything:
Head query: ______________________________
Sub-queries seen (Bing grounding queries / ChatGPT activity panel):
1. ______________________ Answered on my page? yes / no Section: ______
2. ______________________ Answered on my page? yes / no Section: ______
3. ______________________ Answered on my page? yes / no Section: ______
Sub-questions an expert would expect that I do not answer yet:
- ______________________
Stage 3: Which index does each AI engine retrieve from?
Each engine retrieves from an index built by its own crawler or a search partner, so your page must be reachable by the right bot and indexed in the right engine. Retrieval is the stage most sites fail without knowing it, because it is invisible in Google Analytics and looks fine in Search Console.
Google is explicit that its AI features sit on normal Search. Its documentation describes retrieval-augmented generation as a technique "used to improve the quality, accuracy, and freshness of AI responses by relying on our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index." The other engines publish their own crawlers:
| Engine | Search / citation crawler | Training crawler (separate opt-out) | Source |
|---|---|---|---|
| ChatGPT search | OAI-SearchBot | GPTBot | OpenAI crawler docs |
| Perplexity | PerplexityBot | none; PerplexityBot is "not used to crawl content for AI foundation models" | Perplexity crawler docs |
| Claude | Claude-SearchBot | ClaudeBot | Anthropic crawler help |
| Google AI Overviews, AI Mode | Googlebot (normal Search) | Google-Extended token for Gemini | Google common crawlers |
| Microsoft Copilot | Bingbot (Bing index) | n/a | Bing Webmaster Tools |
OpenAI states that "sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers" and that robots.txt changes can take about 24 hours to apply. Perplexity gives the same 24-hour window. Two details most guides miss:
- Google-Extended is not free to block. Google's docs say it controls Gemini training and also grounding in Gemini Apps and Vertex AI, while it "does not impact a site's inclusion in Google Search." Blocking it keeps you in AI Overviews but can remove you from grounded Gemini app answers.
- Search Console now has an opt-out switch. Since August 31, 2026, the Search generative AI control lets a property exclude itself from AI Overviews and AI Mode. Inclusion is the default, but check nobody on your team flipped it.
What we found on our own sites. On 2026-09-26 we requested one of our own product sites with four user agents: GPTBot, OAI-SearchBot, ClaudeBot and a normal desktop Chrome string. All four got HTTP 403 from a CDN bot-protection rule, and even /robots.txt returned the block page. Google's AI Overview still cited that site on 2 of the 10 queries we tested, while ChatGPT cited it on none, which is what you would expect if Googlebot is allowlisted and the AI crawlers are not. Nothing in robots.txt or on the page caused it. The firewall did.
A robots.txt that keeps you citable in every answer engine while opting out of model training (adjust the Disallow lines to your own policy):
User-agent: OAI-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: Bingbot
Allow: /
User-agent: Googlebot
Allow: /
User-agent: GPTBot
Disallow: /
User-agent: ClaudeBot
Disallow: /
If you would rather click than type, our robots.txt generator builds and validates the same file. Then test the firewall, not just the file, from a terminal:
for ua in "OAI-SearchBot/1.4" "PerplexityBot/1.0" "Claude-SearchBot" "GPTBot/1.4"; do
echo "$ua: $(curl -s -o /dev/null -w '%{http_code}' -A "Mozilla/5.0 (compatible; $ua)" https://yoursite.com/your-page)"
done
Anything other than 200 for a search crawler is a stage 3 failure. Being indexed and ranked still matters on top of access: Ahrefs found that 38% of pages cited in AI Overviews also rank in the top 10 for the same query, so classic ranking remains the biggest single input for Google's engine.
Stage 4: How does an AI engine choose which passage to quote?
The engine scores retrieved passages for how directly and specifically they answer each sub-query, and it quotes the self-contained ones. It rarely lifts a whole page. It lifts a paragraph, a table row or a list that makes sense on its own.
The best controlled evidence is the Princeton and Georgia Tech GEO paper (KDD 2024), which tested rewrites on a benchmark of 10,000 queries and concluded that "GEO can boost visibility by up to 40% in generative engine responses." Adding statistics and adding quotations were the strongest single tactics, and keyword stuffing did not help.
Our own calibration points the same way. We scored 45 live pages for 12 queries, then checked which ones ChatGPT or Google AI Overviews actually cited. AUC below is the chance a random cited page scores higher than a random non-cited one on that signal (0.5 is a coin flip):
| Signal on the page | AUC, cited vs not cited | Read |
|---|---|---|
| Answer capsule in the first 300 words | 0.74 | Strongest single signal |
| Specific, concrete claims (names, numbers, versions) | 0.72 | Strong |
| Scannable structure (H2s, lists, tables) | 0.69 | Moderate |
| Attributed quotations | 0.67 | Moderate |
| First-hand evidence or original data | 0.67 | Moderate |
| Outbound citations to sources | 0.51 | Flat in this sample |
| Visible freshness date | 0.52 | Flat in this sample |
| Total score | 0.74 | Cited page beat a same-query non-cited page in 67% of 39 pairs |
Two honest caveats: 45 pages is a small sample, and the capsule question was chosen after testing on these same pages, so 0.74 is optimistic. Treat it as direction, not a law. It still tells you where to spend an afternoon: the opening answer and the specificity of your claims.
What does not act on this stage, according to the evidence:
- Schema markup. Ahrefs tracked 1,885 pages that added schema against 4,000 matched controls and found no major citation uplift on any platform. Google says there is "no special schema.org markup you need to add." Keep it for rich results, as covered in our structured data guide, not for GEO.
- llms.txt. An SE Ranking analysis of roughly 300,000 domains found no relationship between having the file and being cited. Our llms.txt explainer covers when it is still worth the ten minutes.
- Chunking. Google's guide says "there's no requirement to break your content into tiny pieces for AI to better understand it." Self-contained sections help; artificially short fragments do not.
The answer capsule template we use on every page, 40 to 60 words, placed directly under the H1 or the relevant H2:
[Topic] is/works by [direct answer in one clause]. [The one mechanism or number
that makes it true, with its source]. [The practical implication for the reader,
including the main exception or limit].
Stage 5: How do AI engines decide who gets the citation?
Attribution happens after synthesis: the engine writes a combined answer from several sources and then links the ones that support specific claims, so being retrieved is not the same as being credited. Distinct, checkable facts are what earn a link, because a generic sentence that five sources share can be credited to any of them.
Three pieces of evidence show how uneven this stage is:
- Engines rarely agree. Kevin Indig's Consensus Gap analysis of a 20,000-prompt sample found only 2.37% of cited URLs appeared in ChatGPT, Perplexity and Google AI Overviews for the same prompt, and 91% appeared in only one engine. "AI visibility" is three separate scoreboards.
- Google's engine and ChatGPT pull from different pools. Ahrefs reports that only 12% of sources cited by AI search platforms match Google's top results, and that AI platforms cite content that is 25.7% fresher than traditional organic results.
- Even within one portfolio, credit moves. In our 2026-09-26 test, ChatGPT cited one of our sites for a CCAT scoring query, but not the site that actually ranks on Google for it. It chose a different property of ours that covers the same fact. Google's AI Overview cited neither for that query.
What you control at this stage is narrow but real: publish facts that are yours (original measurements, dated tests, named examples), describe your brand the same way everywhere, and earn mentions on independent sites the engines trust. Semrush's study of more than 100 million citations across 230,000 prompts found Reddit, Wikipedia and LinkedIn among the most-cited domains, which is why a consistent presence off your own site feeds this stage.
Which generative engine optimization statistics hold up?
Only a handful of GEO statistics come from controlled or large, disclosed samples; most numbers in circulation are vendor claims. Use this table when a stat shows up in a pitch deck or a client report.
| Statistic | Source | Sample | Evidence strength |
|---|---|---|---|
| GEO rewrites boost visibility by up to 40% | Aggarwal et al., KDD 2024 | 10,000-query benchmark | Strong, controlled, but a lab pipeline |
| 38% of AI Overview citations also rank top 10 | Ahrefs | Large SERP sample | Strong, correlational |
| Schema adds no major citation uplift | Ahrefs | 1,885 pages vs 4,000 controls | Strong, negative result |
| 2.37% of cited URLs appear in all 3 engines | Growth Memo | 20,000 prompts | Moderate |
| llms.txt shows no citation effect | SE Ranking via SEJ | About 300,000 domains | Moderate, correlational |
| ChatGPT browsed for 4 of 10 queries | Our measurement, 2026-09-26 | 10 queries | Small sample, first-hand |
| Answer capsule AUC 0.74 | Our scorer calibration | 45 pages, 12 queries | Small sample, first-hand |
Conversion multipliers for AI referral traffic vary wildly by who reports them (Ahrefs reports its own AI visitors convert 23x better than organic ones on the page linked above), so treat any single multiplier as that company's result, not a benchmark.
How do you test whether the pipeline picks your page?
Run the same prompt in each engine, record whether it searched, what it cited and whether you were named, then confirm with first-party reports. Paste this into ChatGPT, Perplexity and Claude with web search on, once per target query:
Answer this question as you normally would, using web search: "[target query]".
After the answer, list every source URL you used, in order of how much you relied
on it. Then tell me whether [yourdomain.com] appeared in your search results,
and if it did, which passage you would quote from it.
Log the result in a sheet with columns: date, engine, query, searched (yes/no), cited URLs, our URL cited (yes/no), brand named (yes/no). Answers vary run to run, so repeat each prompt three times and count. Then cross-check with data the engines publish about you: the Search Console generative AI report for Google, Bing Webmaster Tools AI Performance for Copilot, and referral sessions from chatgpt.com or perplexity.ai in GA4 acquisition reports. For engine-specific playbooks, see SEO for ChatGPT, Perplexity SEO and how to rank in Google AI Overviews. To score a draft against the stage 4 signals before you publish, paste it into our SEO and GEO analyzer.
Pull the data behind these checks into Claude or ChatGPT
Stages 3 and 5 are answered by first-party data you already own: Search Console, Bing Webmaster Tools and GA4. The friction is that they live in three dashboards. Murkuz is a read-only connector that brings them into the assistant you already use.
You create a Murkuz account, add your site, and connect your own Google Search Console, Google Analytics and Bing Webmaster accounts inside the Murkuz app. Then you add Murkuz as a connector in Claude or ChatGPT (any MCP client works) and ask in plain language:
- "Is /blog/your-page indexed, and which canonical did Google choose?" (stage 3, through URL inspection)
- "Which of my pages rank on Bing but barely on Google, or the reverse?" (stage 3, both indexes side by side, which matters because Copilot runs on Bing)
- "Which landing pages get referral sessions from AI assistants, and do those sessions engage?" (stage 5, from your GA4 traffic sources)
- "Which pages sit just off page one for their main query?" (the ranking input that feeds retrieval)
The assistant answers from your real rows, not from a guess about your site. You can see every question it can answer on the Murkuz MCP tools page, how the connection works on the features page, and the AI search angle on AI search readiness. Plans are on the pricing page. Try it at murkuz.com.
FAQ
What is generative engine optimization (GEO)?
Generative engine optimization is the practice of making your pages easy for AI answer engines such as ChatGPT, Perplexity, Claude, Copilot and Google AI Overviews to retrieve, quote and cite. It adds passage-level and access-level work on top of SEO rather than replacing it, because every engine retrieves from a search index first.
How does generative engine optimization work in simple terms?
An AI engine decides whether to search, splits the question into sub-queries, retrieves pages its crawlers can reach, picks the clearest passages and cites the sources behind specific claims. GEO improves your odds at each of those steps: open access, rank in Google and Bing, answer directly, and publish specific facts others do not have.
Is GEO different from SEO?
Partly. Google says optimizing for its generative AI features "is optimizing for the search experience, and thus still SEO." The differences are crawler access for non-Google bots, passage-level writing and per-engine measurement. Our GEO vs SEO guide covers the split in detail.
How long does GEO take to show results?
Crawler access changes apply in about 24 hours for OpenAI and Perplexity, per their documentation. Content changes show up as fast as the engine re-crawls and re-ranks the page, usually days to weeks. Off-site mentions take months. Measure per engine, because a win in one rarely carries to the others.
Can you see ChatGPT's fan-out queries?
Partly. Reasoning models list their searches in the activity panel, and the browser's developer tools show the conversation request, but neither is documented or stable. For a documented view of the grounding queries behind your own citations, use the AI Performance report in Bing Webmaster Tools, which covers Copilot.
Does schema markup or llms.txt help GEO?
Not measurably for citations. Ahrefs found no major uplift across 1,885 pages that added schema, and a 300,000-domain analysis found no effect from llms.txt. Google says neither is required for AI Overviews or AI Mode. Keep schema for rich results and skip llms.txt unless a specific tool you use reads it.
