In one sentence: Generative engine optimization (GEO) is the work of making your business easy for an AI to retrieve, verify and cite, so that when the AI generates an answer about your trade in your town, your name is in it.

The AEO page explained what an answer engine is and why the answer replaced the list. This page opens the hood. How does the machine actually decide whose name to write down?

What is a “generative engine”?

A generative engine is an AI system that writes its reply, word by word, rather than retrieving a pre-written one.

That distinction matters. A search engine stores pages and returns them. A generative engine (ChatGPT, Google AI Mode, AI Overviews, Perplexity, Gemini, Claude) composes a new paragraph every time. The paragraph did not exist before you asked.

The engine has two parts you need to picture. There is the language model, which is the part that knows how to write and reason. And there is the retrieval step, where the engine goes and fetches current information to write from. When someone asks “who should I call for a slab leak near me?”, the model does not know your business from memory. It goes and looks.

Google describes AI Mode as “breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf”.[1] That is retrieval. One question becomes several searches, run at once, and the results are handed to the model to write from.

So how does it build the answer, step by step?

Let me walk it the way it actually happens, using a local plumber as the example.

Step one: it reads the question and breaks it up. “My water heater is leaking and I’m in Carmichael, who’s good and is it worth fixing?” becomes something like: causes of water heater leaks; repair versus replace; plumbers near Carmichael; recent reviews of those plumbers.

Step two: it retrieves. For each sub-question it pulls sources. For the local part, the sources are overwhelmingly listings and websites. In 1.9 million citations across local AI answers, Google Business Profile was 28.63% of citations and Yelp 9.53%. Businesses’ own websites were 42% of all citations and 93% of the unique domains cited.[2] ChatGPT used a business’s own website 58% of the time in an earlier BrightLocal analysis.[3]

Step three: it reconciles. The model now holds a stack of retrieved facts about maybe fifteen plumbers. Some are consistent: the profile says Carmichael, the site says Carmichael, the reviews mention Carmichael. Some conflict: the profile says one phone number, the site says another, a directory says the business closed. The model weights the consistent ones higher because consistency is its main evidence that a fact is true.

Step four: it generates. It writes a paragraph naming, on average, 4.1 businesses if it is ChatGPT, 3.5 if it is Google AI Mode, and 2.5 if it is an AI Overview.[2] It attaches citations to some or all of what it wrote.

Step five: the person checks. This step is outside the machine but it decides whether you get the call. 88% of people who use AI for local recommendations say they check the legitimacy or source of what the AI told them.[4] They open your Google profile. They read the reviews. If those agree with what the AI said, you get the call.

GEO is the work of being strong at steps two, three and five.

What is a citation, and why does it matter?

A citation is the link or source reference the engine attaches to a claim in its answer. It is the engine’s way of saying “I got this from here.”

Citations are the currency of a generative engine, for two reasons.

First, they are how the engine decides what it is allowed to say. An engine will name a plumber it can point to a source for. It is reluctant to name one it cannot.

Second, they are concentrated. In an analysis of over one million citations, 10% to 11% of URLs received 37% of all citations, and 64% of URLs were never cited by Perplexity at all.[5] Google’s AI Overviews cited 15.2 links per answer on average in February 2026, up from 6.8 in November 2024, but only 19% of those sources overlapped with the organic top ten.[6]

Put those together and you get the shape of the game: more citations are being handed out, they go to a different set of pages than the ranked list did, and a small share of pages collects most of them. The pages that collect them are the ones the engine found easy to retrieve and easy to trust.

What does “entity trust” mean?

An entity is a thing the engine can recognize as one thing. Not a page, not a keyword: a business.

When the engine sees “Mike’s Plumbing” on a Google profile, “Mike’s Plumbing Inc.” on a website, “Mikes Plumbing Roseville” on Yelp and a phone number that matches on all three, it can form a single entity: this is one business, here is where it is, here is what it does, here is what people say about it.

When the engine sees three different names, two phone numbers, an address that is a P.O. box on one listing and a street on another, and a website that never states the service area, it cannot confidently form an entity. It does not know if that is one business or three. It moves on to a plumber it can be sure about.

Entity trust is the degree to which the engine is confident that you are a real, specific, currently operating business doing the things you claim, where you claim. It is built from agreement across sources. It is the single most important idea in GEO, and it is also the least glamorous, because most of it is making sure your name, address, phone, hours and services say the same thing everywhere the engine looks.

There is a technical layer here too, called structured data or schema, which is a way of labeling facts on your website in a format machines read directly. Only 3.97% of mobile web pages carry the “LocalBusiness” kind of this labeling.[7] It helps the engine read you. But one Ahrefs analysis of 1,885 pages found that adding this markup alone produced no measurable change in AI citations.[8] The label is not the trust. The agreement across sources is the trust.

Why does one plumber get named and another not?

Here is the honest answer, drawn from the numbers above and from what the engines say about themselves.

The named plumber is retrievable. Their Google profile is complete, their website has a page per service with the town named, and they are on Yelp and the other directories the engine pulls from. When the engine issues its five sub-queries, this plumber shows up in three of them.

The named plumber is consistent. The same name, phone, address and services appear everywhere. The engine forms one confident entity.

The named plumber is corroborated. Reviews mention the specific service (“replaced our water heater same day”) and the specific place (“in Carmichael”). The engine can match the review text to the question. Reviews are the engine’s evidence that the claims on the website are true. Google as a review source is used by 71% of consumers, and 47% say they would not use a business with fewer than 20 reviews.[9]

The named plumber is nearby. Between 52.8% and 59.3% of businesses the engines picked were within 5 kilometers of the searcher, and 71.7% to 78.2% within 10 kilometers.[2] ChatGPT casts a wider net than Google does, but proximity still dominates.

The plumber who is not named is usually missing one of the four, and most often it is consistency or corroboration. The work itself might be superb. The engine cannot see the work. It can only see what is written about the work.

There is one more thing to know: the engine is not deterministic. The same prompt, repeated, returned the same set of businesses only 20% to 33% of the time.[2] A plumber who is strong on all four factors is named more often, not every time. GEO raises your frequency. It does not buy a permanent seat.

What this is not

GEO is not about the AI model’s training. You cannot get “into” ChatGPT’s memory. The engine retrieves current sources at question time. You influence what it retrieves, not what it remembers.

GEO is not link building in the old sense. The 2010 SEO tactic of collecting links from anywhere is not what generative engines reward. They reward sources that corroborate specific facts about a specific entity.

GEO is not a text file or a tag. A file called llms.txt was proposed as a way to talk to AI crawlers; in an Ahrefs check of 137,210 domains, 97% of those files received zero requests.[8] Schema markup helps readability but did not, on its own, move citations in the same analysis.[8]

GEO is not spam. Generating hundreds of thin pages to be retrieved more often produces more inconsistent signals, which lowers entity trust.

GEO is not separate from your reviews. In a generative engine, reviews are evidence, not decoration.

Where it is going

Forecasts, labelled as forecasts

Forecasts, labeled as forecasts.

Forecast one: retrieval will lean harder on structured, verified sources. Google Business Profile is already 28.63% of local citations.[2] As engines add booking and transactions, they will need facts they can act on (hours, services, availability), and those live in profiles and structured listings more than in prose.

Forecast two: the engines will read more of your site, not less. Citations per AI Overview more than doubled in fifteen months. If that continues, deeper service pages, not just home pages, will be what gets retrieved.

Forecast three: consistency will become measurable and sold as a score. Because the same prompt yields different names most of the time, the useful measure is frequency across many prompts and platforms. Expect that to be the standard way businesses are told where they stand.

Forecast four: visitors who arrive from an AI answer will keep being worth more than visitors from a list. Semrush measured AI-search visitors at 4.4 times the conversion value of organic visitors.[10] The mechanism is simple: they arrived pre-recommended. I would expect that gap to narrow as AI traffic grows, but not to close.

What I will not forecast is a single winning platform. ChatGPT, AI Mode and AI Overviews each named a tracked business about a third of the time, and each reaches a very large audience of its own.[11][12][13] The work that makes you retrievable and trusted by one makes you retrievable and trusted by all of them, because they read the same sources.