Generative Engine Optimization (GEO): How AI Engines Choose What to Cite

Somewhere this morning a buyer described a problem to a chat window, an engine assembled an answer from four or five sources, and a shortlist got written. No ranking report recorded it. The mechanics of how those four or five sources got chosen are knowable, unevenly evidenced, and the subject of this guide. Separating what has been tested from what merely gets repeated is most of the work at Cite First AI.

What is generative engine optimization?

GEO is the discipline of structuring your content, evidence and brand signals so that generative engines retrieve your pages and quote them. A generative engine is any AI system that answers questions by synthesising web sources: ChatGPT, Google AI Mode, Perplexity, and the growing field of assistants built the same way.

GEO is an academic term, not a marketing one. Researchers at Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI coined it in a 2023 paper, “GEO: Generative Engine Optimization”, and their framing of the problem still holds: “content creators have little to no control over when and how their content is displayed,” Aggarwal and his co-authors wrote. You can’t control the engine. You can control whether your content has the properties engines demonstrably prefer.

The same practice goes by other names (AI search optimization, AI search engine optimization, LLMO) depending on who is selling it. The label matters less than the mechanics, and the terms overlap more than the people selling them admit; the AEO vs SEO split is where they actually diverge.

How do AI engines actually pick their sources?

Type a question into ChatGPT and one of two things happens before your page is ever in the running. Either the model answers from what it already knows, or it decides the question needs the live web and goes looking. Which one fires decides whether your content is eligible at all, and most advice about AI visibility only addresses what happens after the search. Three stages, each rewarding different work:

1. Trained knowledge. The model answers from what is encoded in its weights, with no web access.
   

2. Retrieval. For some questions, the system decides to search and pulls back candidates.

   
   

3. Selection. The model chooses which retrieved pages to quote and cite.

   
   

Stage 1: Trained knowledge, the answer before any search happens

Every engine starts with what its model already knows. Retrieval is conditional, and the vendors say so: Anthropic publishes a list of when Claude searches and when it doesn’t (anything “current, changing, or outside its training data” triggers a search; it “answers directly without searching when the request draws on stable knowledge”), and OpenAI documents the same choice. Search is a tool the model elects, not a step that always runs.

This stage handles the settled questions: what a category is, who the players are, how an approach generally works. On those, a buyer can get a complete answer with no page of yours in it, because nothing was fetched.

For a marketer, being one of the names the model learned during training is a real advantage on those early, orientation questions, and it is earned almost entirely off your website: mentions, reviews and coverage, accumulated long enough to make it into a training run. It also moves slowly in both directions. A model that shipped last year cannot know what you published last month, and a name that took years to learn will not be forgotten in a quarter.

Stage 2: Retrieval, getting into the candidate pool

The closer a question moves toward a decision, the likelier the engine is to search. Nobody has published a reliable rate for how often that happens, but the vendors’ own criteria tell you what qualifies: anything current, changing, or specific. Prices, vendors and comparisons all qualify, and so does any subject where the knowledge itself keeps moving: technology, regulation, medicine.

Product tier matters too. Quick, default models are tuned to answer fast and lean hardest on what they already know; the deep-research modes exist precisely to search widely, pulling in dozens of sources per session. Between the two, most of the questions that carry commercial weight pass through this gate.

When an engine searches, it doesn’t read the whole web. It runs several queries at once and pulls back a shortlist of candidates. Three things decide whether you make that shortlist:

• You’re in the index the engine searches. Google’s AI features search Google’s index. ChatGPT’s browsing has historically leaned on Bing plus OpenAI’s own crawling. Perplexity maintains its own index. If a crawler can’t reach you, nothing downstream matters.

• You rank for the sub-questions, not just the head term. Engines decompose one prompt into many narrower queries. A page that ranks #4 for a specific sub-question can beat a page that ranks #1 for the broad term.

  
  

• You’re reasonably fresh, though less than you’ve been told. Retrieval does skew newer: across 17 million cited URLs, Ahrefs measured AI assistants citing pages about 26% fresher than organic results. The absolute number cuts the other way, though: the average cited page is still just under three years old. Freshness helps at the margin. It is not a gate.

  
  

Stage 3: Selection, getting quoted from the pool

Retrieval gets you considered; this stage decides who gets credit. The model reads what it fetched, builds its answer from it, and names only some of its sources. ChatGPT, for instance, cites roughly half of what it retrieves.

Trust runs high in this stage. In a May 2026 test of five frontier models, when a fetched page disagreed with the model’s own trained knowledge, the models sided with the page more than 97% of the time. The model believes what it reads. Your job is to be worth quoting when it does.

And what gets quoted is the passage that answers cleanly: a complete claim, something checkable behind it, and wording that stands on its own out of context. A paragraph that says “as mentioned above, it depends” is invisible. The graded evidence below covers the rest, including the tested tactics that turned out not to move this stage at all.

How does ChatGPT choose its citations?

ChatGPT searches when a prompt needs current information, pulls candidates through its own crawlers and search partnerships, then cites the handful of sources whose passages directly support its answer. It retrieves about 16.6 URLs per prompt and cites roughly half of them, Ahrefs measured across 1.4 million prompts.

What happens to the other half is the most interesting number in that study. Reddit makes up 67.8% of the URLs ChatGPT reads and then doesn’t cite. The naive reading is that Reddit doesn’t matter to this engine. The better reading is that Reddit is where ChatGPT does its homework. Those threads sit in the model’s context while it decides what to recommend and whom to cite. And as the selection research above shows, a model overwhelmingly uses what’s in front of it. The citation may land on a brand’s website, but the shortlist was often shaped in a forum thread the answer never names. If buyers discuss your category on Reddit, that conversation is steering ChatGPT’s answers even when no Reddit link appears. Influence without attribution.

How do Google AI Overviews and AI Mode pick sources?

Google answers by “fanning out”: a Gemini model breaks your question into many sub-queries, runs them against Google’s index simultaneously, and synthesises the results into one cited answer. Sources get chosen per sub-question, which is why a page that thoroughly answers one narrow question can beat a broader, better-ranked competitor.

This isn’t SEO-community speculation; it’s Google’s own description. “AI Mode is basically doing a dozen searches for you in the time it takes to do one,” explains Dounia Berrada, Senior Engineering Director on Google Search. The strategic consequence: the unit of competition stops being the keyword and becomes the question cluster. A CMO asking one question about approach quietly triggers a dozen sub-queries (cost, timeline, in-house versus agency, measurement, risk), and every one is a separate chance to be retrieved.

Two Google-specific notes. Its AI features inherit Google’s ranking systems, so ordinary SEO health remains the price of entry. And the domain mix has shifted toward video: YouTube is now the most-cited domain in AI Overviews.

How does Perplexity choose what to cite?

Perplexity runs its own crawler and index, registers crawler-setting changes within 24 hours per its own documentation, and cites inline for nearly every claim: roughly three citation slots for every one ChatGPT offers, refilled faster. Users prize it for fact-checking, and the reputation is earned, but notice what earns it. Every claim carries a source you can click, which is a verifiability habit, not a recency one. It cites the oldest content of any assistant, averaging just over three years, because for checking facts, established beats new. The engine’s speed lives in its index, not its taste: it will notice a new page in days, then go on citing three-year-old sources beside it. For anyone doing GEO work, that combination makes it the first scoreboard worth checking.

One real change to know about: Perplexity is no longer the Reddit engine. Reddit fell from 25% of its citations in February 2026 to 7% by April, after Reddit sued Perplexity over scraping. Source preferences move for legal reasons as well as algorithmic ones.

What about Claude, Grok and Gemini?

Claude’s citations track Brave Search closely, so a Brave visibility check is worth ten minutes and almost nobody does one. Grok is the outlier: its documentation confirms two retrieval pools, its own web search plus live search over public posts on X, making it the one engine where an active X presence is a direct citation input. Gemini grounds through Google Search, so your AI Overviews work carries over almost unchanged.

None of them changes the three stages. Build for the shared mechanics (crawl access, sub-question coverage, checkable claims, off-site corroboration) and every engine inherits the work.

Sources and dates for every figure in this table are in the graded evidence below.

What actually increases your odds of being cited? The evidence, graded

The public evidence is not the whole picture. There is real money in knowing what moves AI citations, and the practitioners getting results in this field keep their playbooks private. So do we: we don’t reveal the details of our methods publicly, and we don’t need to, because they show up in the results they deliver. What we can do is grade what the published studies actually found, because most of what circulates as GEO advice has never been tested at all, and some of the most heavily promoted tactics have been tested and failed.

Each grade means one thing. Strong means several independent studies point the same way. Moderate means the evidence is good but thinner. Mixed means credible studies disagree. Contradicted means the best-controlled test says it doesn’t work. Nothing published before 2025 counts on its own, every source is linked in the table, and where a row reports a correlation, remember what that is: a 0-to-1 measure of how tightly two rankings track each other, never proof of cause.

Three observations from this table shape everything we recommend.

First, the lever is concreteness, not evidence-shaped decoration. The conflicting studies stop conflicting when you notice what each one manipulated. The null results came from inserting statistics into existing prose. The large positive results came from testing whether a page states checkable specifics: a price, a specification, a figure. So the instruction is not “add three statistics per section.” It is “state the specifics a buyer would need, and make each one checkable,” which is what a statistic is for.

Second, the biggest effects aren’t on your website at all. Position in the retrieved set beat every rewrite. Being talked about across the web predicts AI visibility far better than your domain’s authority score does. The two loudest examples are YouTube and Reddit: YouTube mentions are one of the strongest visibility signals measured (across ChatGPT and Google’s AI surfaces), and Reddit threads shape ChatGPT’s shortlist even when no Reddit link appears in the answer. And the trained knowledge that answers when no search fires is built from years of coverage across the open web. All three stages point at the same conclusion from different angles: what the rest of the internet says about you moves AI visibility more than what your own site says about you. That is why a serious program includes digital PR, video and community presence, not just on-page work.

Third, the popular technical hacks now have evidence against them, not merely absent. Schema markup went backwards in the best-controlled test ever run on it. llms.txt goes unread by the systems it was supposedly built for. Both are cheap and harmless, and fine to implement for other reasons, but anyone selling either as a visibility lever is selling against the evidence.

Does llms.txt do anything?

No, and this is the clearest gap between what the field promotes and what anyone has measured.

llms.txt is a proposed standard: a file at your site’s root that hands AI crawlers a tidy summary of your content. It is currently the most enthusiastically recommended technical tactic in AI search, and it is close to inert. Of 137,210 domains surveyed, 28% publish a valid file, and 97% of those files received not a single request in the month measured. The largest category of requests that did arrive came from SEO audit tools checking whether the file existed.

Google’s documentation is blunter than we would dare to be: Search “doesn’t use them,” and publishing one “will neither harm nor help” your visibility. Jeremy Howard, who wrote the specification, says the same about his own standard: it was designed to help coding assistants navigate documentation, not to influence search.

Publish one if you like. It costs an afternoon and breaks nothing. Just don’t let anyone bill it as a visibility strategy, and be suspicious of any proposal that leads with it.

What gets content retrieved but never cited?

Most content that fails at GEO fails at selection: the page enters the pool but loses the quote to a competitor. Nothing in your analytics tells you it happened. Auditing pages against the citation evidence, the same failures account for nearly all of it:

1. The answer arrives too late. A question-shaped heading followed by three paragraphs of context before the answer. The model extracts passages, finds no direct claim near the heading, and moves on to a candidate that answers immediately.

   
   

2. Claims with nothing checkable attached. “Industry leaders agree that response time matters” gives an engine nothing defensible to quote. “Leads answered within five minutes converted at twice the rate, in a named 2026 study” does.

3. No corroboration anywhere else. Engines cross-check. A claim that exists only on your domain is a lonely claim, and lonely claims lose to corroborated ones.

   
   

4. Visible staleness. A 2023 date stamp, statistics from two model generations ago, screenshots of retired interfaces. A stale timestamp is the cheapest way to lose a citation you’d otherwise have won.

5. Context-dependent passages. Paragraphs built on “as we discussed above” collapse when extracted alone. Every important paragraph should name its subject and stand as a complete, correct statement, because a passage, not your page, is what gets lifted.

   
   

One discipline fixes all five: write every section as if it will be read by itself, judged by itself, and quoted by itself.

How do you optimize content for AI citations?

Four principles carry most of the result. The order matters more than any individual tactic, and the thresholds (how many sub-questions, how often to refresh, which mentions to chase) are where the actual work lives.

Be reachable. Allow the crawlers, and verify indexation in Bing as well as Google, since ChatGPT’s search layer has leaned on Bing. This is the cheapest failure to fix and the most common one we find.

Answer the question the engine is actually asking. Engines decompose a prompt into narrower sub-questions and choose sources per sub-question. Content organised as direct answers to specific questions maps onto that. A long essay with clever headings does not.

Make every claim checkable. State the specifics a buyer needs, from a named source with a date, so any passage lifted out of context still stands up. A sentence you cannot support is better deleted than decorated.

Earn the off-site layer. Mentions in industry press, comparison articles, video and relevant communities are the strongest signal in the correlational data. Over years, they are what write you into the model’s trained knowledge, where the most durable answers live.

You can start the first three this week. The fourth is the compounding engine, and it cannot be rushed, which is why sequencing is the part of this that clients actually pay for.

What does a real GEO program look like?

A credible program runs three workstreams in parallel: technical access, checkable content built around question clusters, and off-site authority. Each maps to a stage: access feeds retrieval, content wins selection, and off-site coverage builds the trained knowledge that answers before any search fires. One-off “GEO audits” that produce a schema to-do list are theater; the evidence points at content and corroboration.

That three-workstream shape is the skeleton of our Citation Cascade Method™: five stages, Diagnose through Compound. The stages are public. The sequencing and thresholds inside them are the part we keep.

In 25 years of marketing I’ve learned to distrust anyone who guarantees a result, and AI search gives a guarantee even less to stand on. The field moves so fast that nobody can say what the work will require three months from now. You can’t see what your competitors are doing, either, and if one of them suddenly pours resources into AEO, you’ll have to match it. Underneath all of that, selection is probabilistic and answers vary run to run, so nobody can promise a specific citation. What the evidence supports is systematically raising the probability of citation across thousands of relevant prompts. That’s the honest product.

The teams winning citations right now aren’t the ones chasing loopholes. They’re the ones treating every page as a source an engine can defend quoting.

Want to know your current AI citation share?

Cite First AI will run your highest-value prompts across the major engines, show you exactly who’s being cited instead of you, and map the shortest path to changing that. It’s the same process that took InCorp from AI-invisible to consistently cited in their category (see the InCorp results).

The citation-share mini audit is free, and it’s the diagnostic stage of the Citation Cascade Method™. Book it below, and if it turns up a gap worth closing, we’ll scope the work on the same call.

Book your free mini audit →

FAQ: How AI engines choose citations

How long does generative engine optimization take to work?

Expect first movement in weeks on Perplexity, whose documentation puts crawler-setting changes at 24 hours and which picks up new pages on active sites within days, and one to two quarters for meaningful citation share in ChatGPT and Google AI Overviews. On-page fixes show fastest; off-site mentions compound over months.

Do backlinks still matter for AI search?

Yes, but less than mentions, for AI visibility specifically. Ahrefs measured 75,000 brands and found how often a brand is mentioned across the web predicts its AI visibility about two and a half times more closely than its Domain Rating does. Backlinks remain a major factor in the classic rankings that still gate retrieval, so this is a question of which stage you’re working on, not whether links matter.

Does schema markup help you get cited by AI ?

The best available test says no, and slightly worse than no. Ahrefs matched 1,885 pages that added structured data against 4,000 controls and measured AI Overview citations falling 4.6%. We still implement Article, FAQPage and Person schema as a cheap comprehension aid, but we never sell it as a citation lever, because the evidence doesn’t support that claim.

How do I know AI engines if are citing my company?

Sample the engines directly: run your highest-value buyer prompts across ChatGPT, AI Overviews, Perplexity and Gemini monthly, and track citation and mention share. One-off manual checks mislead, because answers vary by user, session and phrasing. Purpose-built tracking platforms, or a structured audit, are the credible route. Server logs add the retrieval view: watch for OAI-SearchBot, PerplexityBot and Claude-SearchBot.

Eric Sharp is the founder of Cite First AI, an answer engine optimization agency. He has spent 25+ years in marketing and now works exclusively on AI search visibility for national brands.