How Generative Engine Optimization Actually Works

Most GEO advice is a list of tactics with no theory of why they work. That is why teams end up adding schema for six weeks and wondering where the citations are. This is the mechanism instead: the five stages an AI answer passes through, what actually happens at each one, and which levers exist where. The data is a mix of published research and what we measured on our own domain.

Key takeaways

• An AI citation is the output of a pipeline, not a ranking. Crawl, retrieve, ground, corroborate, attribute. Fail any stage and you are invisible regardless of how good the page is.
• AI fetchers already outwork Googlebot here. In the week to 2026-07-27, ChatGPT-User made 2,370 requests to strataigize.com against Googlebot’s 1,390, and total AI crawl demand tripled week over week (Cloudflare).
• Retrieval is not your ranking. 2026 analyses put the overlap between AI Overview citations and the organic top 10 between roughly 17% and 52%, so being cited is a separate outcome from being ranked.
• Grounding rewards substance. Adding statistics, quotations and cited sources is the most replicated lever in the field: a 2026 analysis of 21,143 citations put the influence uplift at 61.55% for statistics, 57.33% for definitions and 55.28% for comparisons. Keyword stuffing did nothing, in every study that tested it.
• Corroboration is the stage most teams never work on, and it is the one that decides whether the engine trusts the claim enough to attribute it to you.

The pipeline, and where the levers are

Stage	What the engine does	What you control	Strongest lever
1. Crawl	Fetches your page, live or from an index	Fully	Crawler access, render speed, static HTML
2. Retrieve	Breaks the question into sub-questions, pulls candidate passages	Partly	Topic coverage, heading relevance to the sub-question, freshness
3. Ground	Extracts the passages it will build the answer from	Fully	Answer-first structure, self-contained passages
4. Corroborate	Checks the claim against other sources it trusts	Partly	Third-party mentions, original data others quote
5. Attribute	Decides whether and how to name a source	Barely	Entity clarity, being the origin of the fact

Everything below is a walk through that table.

Stage 1: the crawl, and the three bots that are not the same bot

An AI engine cannot cite what it cannot fetch, and “AI crawler” is three different things wearing one label. Getting this wrong is the most common technical failure in a GEO program.

Using OpenAI’s fleet as the example, documented in OpenAI’s bots reference, the roles split cleanly:

• GPTBot crawls for model training. Blocking it affects what future models know about you. It has nothing to do with today’s answers.
• OAI-SearchBot builds the search index ChatGPT queries against. Blocking it removes you from the candidate pool.
• ChatGPT-User is the live fetcher. It fires when a real person’s question needs your page pulled right now. Traffic from this agent is the closest thing there is to demand-side proof that people are asking questions your pages answer.

Other engines follow the same pattern with different names (PerplexityBot, ClaudeBot, Google-Extended). A blanket robots.txt rule written for one intent silently kills the other two.

Here is why the distinction is worth your attention. On our own domain in the week to 2026-07-27, Cloudflare logged 2,370 requests from ChatGPT-User against 1,390 from Googlebot, with total AI crawl demand tripling week over week. The live-answer fetcher, the one that only fires because a human asked something, is now working this site harder than Google’s indexer is. That is not a projection about where search is going. It is a server log.

Two practical implications. First, render speed matters more than it used to, because a live fetcher on a user’s clock has less patience than an indexer with a queue. Static HTML wins; client-side rendered content is a coin flip. Second, if your analytics show nothing from AI while your server logs show heavy AI fetching, believe the logs. The gap between them is the channel.

Stage 2: retrieval, or why your ranking stopped predicting your citations

Generative engines do not answer your question. They decompose it. A prompt like “best way to handle app store fee changes” becomes several sub-questions, each triggering its own retrieval pass, and the answer is assembled from whatever came back across all of them. This is usually called query fan-out.

That decomposition is why ranking and citation have come apart. You are not competing for one query anymore, you are competing for a spray of sub-questions you never see, and the candidate set for each is drawn more broadly than the top ten links. The published 2026 measurements disagree on magnitude but not direction: BrightEdge’s rank-overlap tracking and Originality.AI’s citation study bracket the AI Overview citation overlap with the organic top 10 between roughly 17% and 52%.

What you control here is coverage and shape, not the fan-out itself:

• Cover the sub-questions explicitly. One page that answers the main question plus its six obvious follow-ups gets retrieved for more passes than six thin pages that each answer one. Depth beats fragmentation, and Google’s own AI optimization guidance explicitly calls out spinning up a separate page per query variation as scaled-content spam.
• Match the sub-question, but do not fetishise the question mark. What earns retrieval is a heading that is semantically close to what the user asked. The Q&A shape itself does almost nothing, and the evidence on this got clearer in 2026: a 21,143-citation analysis found Q&A framing scored -5.74% on influence, and a 252,000-trial factorial study across six models found formatting variables sat at odds ratios near 1.0, with the authors recommending teams explicitly deprioritise formatting-only edits. Write the heading that matches the question. Do not convert your page into an FAQ because you read that engines like questions.
• Stay fresh, genuinely. Recency carries unusual weight in these systems. A real update changes the content; a date bump changes nothing and the engines increasingly see through it.

Stage 3: grounding, and the finding that named the field

Grounding is where the model selects the specific passages it will build the answer from. This stage is fully in your control and it is where the most reliable evidence exists.

The GEO benchmark study (Aggarwal et al., KDD 2024) tested nine content modifications against roughly 10,000 queries. The winners were adding verifiable statistics, incorporating credible quotations, and citing reliable sources. Improving fluency helped less. Keyword stuffing produced negligible or negative effects.

Be careful with the headline number from that paper, because it is quoted badly everywhere including by agencies selling against it. The often-repeated “40% lift” is a relative gain over a 19.8% baseline on a position-adjusted word-count metric, measured among five sources that were already injected into the model’s context. It is a share-of-answer result inside a simulator, not a measure of whether you get discovered. A 45-study scoping review published in July 2026 makes the point directly, and C-SEO Bench (NeurIPS 2025) found most such methods are ineffective or actively negative on ranking, with traditional SEO outperforming them, and gains shrinking as more people adopt them.

The direction holds and has replicated: concrete extractable evidence is the most reliable lever at this stage. A separate 2026 analysis of 21,143 citations across ChatGPT, Google and Perplexity put the influence uplift at 61.55% for statistics, 57.33% for definitions and 55.28% for comparisons. Use those numbers rather than the 40%.

Translated into editing rules:

1. Lead with the answer. Every section opens with two to four sentences that fully answer its heading, no throat-clearing. If a passage needs the paragraph above it to make sense, it cannot be lifted, and a passage that cannot be lifted cannot be grounded on.
2. Put a number in it. Specific, sourced, and attributable beats a well-written generality every time. Engines reach for concrete claims because concrete claims are what they can attribute.
3. Quote named people and link the source. A cited claim is a claim the model can verify cheaply, and cheap verification is what gets a passage selected.
4. Use real tables. Semantic HTML tables of options, costs, or benchmarks are dense, unambiguous, and easy to extract. Comparison content is disproportionately retrieved because comparison is what buyers ask for.
5. Write definitively. Hedged prose (“it may be worth considering”) gets passed over for text that states things plainly. Cited text skews definitive.

Stage 4: corroboration, the stage nobody works on

Before an engine will build an answer on your claim, it looks for the claim elsewhere. A statement that appears only on your own domain is treated as a vendor assertion. The same statement echoed across listicles, communities, and coverage becomes a fact the model is comfortable repeating.

This is why GEO stops being a website project. The surfaces that do the corroborating are ones you do not own:

• Round-ups and “best X” lists, because they pre-package exactly the comparison an engine is trying to build. Being in someone else’s list beats publishing your own.
• Communities and forums, which are heavily represented in both training data and live retrieval. Genuine participation compounds; spam gets you filtered.
• Earned coverage and third-party mentions, including unlinked ones. An unlinked mention still teaches the model an association between your brand and a topic, and associations are what get recalled.
• Original data that other people quote. This is the highest-leverage move available, because it inverts the relationship: instead of chasing mentions, you publish the number everyone else has to cite. Every citation of your data is a corroboration event you did not have to negotiate.

Stage 5: attribution, and why you get named or skipped

The last stage decides whether the engine says “according to you” or just says the thing. You have the least control here, and two factors do most of the work.

Entity clarity. The model needs to know who you are with enough confidence to name you. Inconsistent company names, missing or contradictory profiles across LinkedIn, Crunchbase, and review platforms, and vague descriptions of what you do all make attribution risky, and engines skip risky attributions. Fixing this is unglamorous data hygiene and it is genuinely load-bearing.

Being the origin. Engines attribute to the source of a fact more readily than to a site that repeated it. If the number originated with you, the attribution follows the number. If you summarized someone else’s research, the citation usually goes to them. This is the strongest argument for original data that exists, and it is not really a GEO argument at all: it is just being worth citing.

What you cannot control, and should stop trying to

Three things are outside the pipeline you can influence, and pretending otherwise burns quarters.

The fan-out. You do not see the sub-questions and you cannot target them individually. Attempting to, by publishing a page per phrasing, is the exact pattern Google names as scaled-content spam. Cover the topic properly instead.

The wording of the answer. The model paraphrases. You can influence what it grounds on, never how it phrases the result.

Whether the user clicks. Most do not, and that is the deal. AI citations convert to visits at a brutal ratio, so the payoff is presence in the consideration set and the booked calls that follow, not sessions. Any GEO program measured purely on traffic will be cancelled before it pays. Measure citation share and self-reported attribution instead.

What the pipeline produced on our own site

We ran this on strataigize.com before selling it to anyone, which is the only reason we are willing to publish specifics.

A focused 30-day push, starting from no organic visibility and zero presence in AI search, produced more than 100 AI citations and over 600 AI-driven sessions, with ChatGPT sessions rising from 38 to 474 and engagement rate improving from 44.7% to 53.8% at the same time. Perplexity went from 21 sessions to 76. Leads arrived attributed directly to both engines. The full before-and-after tables are in the AI SEO case study.

The revenue picture, stated with its caveat attached: both of the only two clients our website has ever produced came through AI answers, worth $215,603 together, or 29.7% of our lifetime closed-won revenue. Roughly 95% of that is a single deal. That proves the channel closes at our deal size and proves nothing about a rate. We keep the caveat attached because a number this good is exactly the kind that gets misused.

One structural gotcha worth stealing: the larger of those two deals is tagged “Organic Search” in our own CRM, because a click inside a Google AI Overview still reports as google.com. No analytics or CRM tooling we own can separate AI-sourced traffic from blue-link organic. We only know the truth because the client told us. If you take one operational lesson from this post, make it the “how did you hear about us” field on your forms.

The agent-readable layer, described honestly

We also run a public MCP server at mcp.strataigize.com, a live A2A agent, and published Agent Skills, which as far as we can find makes us the only agency an AI agent can call directly rather than read about. It is a genuinely useful demo and it will matter more as agents start transacting.

What it is not, today, is a citation lever. Machine-readable endpoints and agent manifests have no demonstrated effect on whether ChatGPT names you in an answer, and we are not going to pretend otherwise while the evidence is absent. Build the five stages above first. Treat the agent layer as positioning and future-proofing, which is exactly what it is.

Where to go next

• Start here if the terms are the problem: AEO vs GEO vs SEO explained.
• The plain-language introduction: what generative engine optimization is.
• Buying rather than building: what AI SEO services include and how to vet an agency.
• Tooling for each stage: the best AI SEO tools.
• See where you currently stand: the AI visibility checker, or our managed generative engine optimization service.

Frequently asked questions

Should I block AI crawlers? Only with a clear reason and only the specific bot. Blocking a training crawler is a defensible content-licensing position. Blocking the search indexer or the live fetcher removes you from answers your buyers are reading right now. Decide per bot, never with one blanket rule.

Does llms.txt help me get cited? There is no evidence that it does. It costs almost nothing to publish and it is reasonable hygiene for agent consumers, but treat any agency selling it as a citation strategy with suspicion. The levers are content, corroboration, and entity clarity.

How often do AI engines re-check a page? Frequently, and far more often than classic crawl cycles for pages in active answer sets. That is why recency carries weight and why a genuine refresh cadence outperforms publish-and-forget. Our own AI crawl volume tripled week over week during an active publishing period.

Why do I show up in one engine and not another? Because the citation sets barely overlap. Each engine has its own index, its own retrieval behavior, and its own trusted-source biases. Broad, well-sourced content that corroborates widely is the only approach that travels across all of them; optimizing for one engine specifically is a losing bet as they keep changing.

Is any of this different for B2B? The mechanism is identical, the stakes are higher. B2B buying questions are research-heavy and comparison-shaped, which is precisely the query type generative engines handle best, so the citation often reaches a decision-maker earlier than a blog click would have.