What Is Generative Engine Optimization? GEO, AEO and How AI Picks Who It Cites

We sell this work to real estate firms, so read us with that in mind. Every number below says where it came from and when, and which of our claims rest on our own studies and which on someone else’s. Some of it is unflattering to GEO, and some to us.

What is generative engine optimization?

Generative engine optimization is the practice of making your brand easy for an AI engine to find, quote and name when it writes an answer.

“GEO” stands for generative engine optimization, and a “generative engine” is any search system that writes an answer from sources it has retrieved instead of only listing links. The term comes from a November 2023 research paper by Aggarwal and colleagues, later accepted at KDD 2024. It reported visibility gains of up to 40% from tactics such as adding citations, quotations and statistics to a page. The tests used a 10,000-query benchmark the authors built, run on their own engine and on a 200-query sample from Perplexity, so the paper shows the idea works in a lab and says nothing about your market.

The practical difference from SEO is how few places there are to win. A results page shows ten links. In our 150-query real estate study the engines cited an average of 4.1 domains per answer on ChatGPT, 6.3 on Google AI Overviews and 7.7 on Perplexity. A business outside that handful has not ranked low, because there is no ranking to be low on.

For a real estate agent the reader is a buyer deciding who to call, and buyers do ask AI. Veterans United Home Loans, a lender, found that 45% of prospective homebuyers had used AI tools in their search, up from 37% the year before, in a survey of 859 people fielded on 1 to 17 June 2026. Surveys on this disagree widely, from 20% to 82% depending on who asked, and our real estate SEO statistics page lists each one with its sample.

What is answer engine optimization (AEO), and is it different?

Answer engine optimization is the older and narrower job: shaping one passage so an engine can lift it as the answer.

It began with featured snippets and voice replies, where a single extracted passage was the whole prize. This is how we separate the three terms, including how AEO differs from both SEO and GEO:

• SEO gets a page ranked in a list of links, and is measured by position and clicks.
• AEO gets a passage lifted as the answer, and is measured by whether your wording is the one quoted.
• GEO gets your brand named inside a written answer across several engines, and is measured by how many of the answers name you.

That split is our usage and not an industry standard, and vendors draw the lines in different places. For us AEO is the extraction craft inside GEO. Both reward a passage that answers its heading and still makes sense when lifted out of the page, so arguing about labels costs more than it returns. The full comparison, with a table, is in GEO vs SEO vs AEO.

How invisible are most businesses to AI? Measured in two markets

Most businesses are absent from the answers that matter to them, in both markets we measured.

The first market is Houston real estate. From 18 to 22 September 2026 we put 500 Houston questions to four AI surfaces and recorded every business named in the 2,000 answers. Only about 3,000 agents, teams and brokerages were named at all. We estimate that 94.0% of the Houston Association of Realtors’ 50,000-plus members were never named (95% interval 93.6 to 94.5). The estimate rests on a random sample of 394 extracted names that an AI reviewer labelled, with the labels published so anyone can dispute them, and it comes from a single run. The 500-query Houston study has the method and every name.

The second is online retail, and the shape was the same. From 18 to 25 June 2026 we asked ChatGPT, Perplexity, Gemini and Google AI Overviews one category question for each of 181 US ecommerce brands, all of them advertising on Meta. A total of 158 brands, 87%, were neither named nor cited by any engine. That sample was selected for having an ad budget, and it was ecommerce, so read 87% as a fact about those 181 brands. The 181-brand study says the same on its own limitations list.

The best and worst engines sit three percentage points apart, so no single engine was dragging the average down. Every brand was paying for ads at the time, which makes them working businesses that the engines still did not name.

How does generative engine optimization work?

Before an AI engine names you it does two things in order: it retrieves a pool of pages, then it selects which ones to quote. GEO is the work of being in the pool, and then being the easiest thing in it to quote.

The sequence has four steps, and each can fail separately:

1. Look up or remember. The engine decides whether the question needs a live search. Some answers come straight from training memory and carry no sources, and you cannot edit what a model has already learned, so only the live-search kind responds to GEO quickly.
2. Retrieve. The engine turns the question into one or more searches (Google’s documentation describes this as issuing multiple related searches across subtopics) and collects candidate pages. A page its crawler cannot fetch, or that is not indexed, never enters the pool.
3. Select. The model reads the candidates and picks the passages that answer the question cleanly. Nobody outside the vendors knows the exact weights, and the best published evidence, the KDD paper above, points to citations, quotations and statistics.
4. Attribute. The engine names or links some of the sources it used and leaves others unnamed. A page can shape an answer without ever appearing in it.

Which step failed decides what you fix, and the symptoms are different enough to tell apart:

We would like to tell you which step fails most often, and we cannot. Our studies record who got named, not why a page was left out, so most of the absence in them is unexplained. The one retrieval failure we have caught cleanly is on our own site, described below, and it turned out to hit a different crawler from the one that matters for search.

How AI decides which brands to recommend

Three inputs decide it, and you can move two of them.

Training memory fixes what the model knows without searching, and it cannot be edited retrospectively. Live retrieval decides what the model reads at the moment of the question, and your crawler access and page structure control that. Corroboration, meaning what independent sites say about you, decides whether a retrieved claim is trusted enough to repeat. It is earned on other people’s sites, which makes it the input you control least.

How much corroboration outweighs your own website is the part we would be careful about. In Houston, the one market where we measured it, the answers cited portals and referral networks far more than the firms’ own sites (the figures are in the real estate section below). We have no evidence that this holds outside real estate. Engine by engine the emphasis also shifts, so the same work pays differently in ChatGPT and in Google AI Overviews.

Why AI citations do not transfer between engines

Each engine retrieves from different sources and selects by different rules, so being cited by one tells you very little about the others.

From 8 to 12 September 2026 we ran 150 real estate queries across five engines and recorded 747 individual measurements. Those answers cited 1,297 distinct domains. Of those, 1,077 were cited by exactly one engine, which is 83%, and three domains were cited by all five. We left Grok out because only 19 of the 150 queries completed on it.

Cited domains

1,297

distinct domains cited across the 747 measurements from five engines in our real estate study.

Photo: 652234, Pixabay

Gemini’s row is a floor. It returned a citable source list on only 55 of its 150 answers. Google AI Overviews leaned hardest on YouTube, which it cited in 55 answers through 91 separate links. An engine this different from its neighbours makes one blended AI visibility score misleading, because it hides the engine you are actually losing in. The 150-query real estate study publishes the full matrix. It was one run in one week, and answers move between runs.

How to appear in AI search results: the access floor

Check that each engine’s own crawlers can fetch your page, because a host or CDN can refuse them and nothing in your own configuration will show it.

We found this on our own site. On 14 September 2026, 8 of 8 cache-miss requests carrying OpenAI’s training crawler identity, GPTBot, were refused with HTTP 429. Seven other identities, including OpenAI’s search crawler and an ordinary browser, passed on 56 of 56 requests from the same machine in the same minutes. Our host confirmed in writing that the rule is deliberate, applies across its platform, is not configurable per domain, and will not get an exception on shared hosting.

That sounds like a retrieval disaster and it is a smaller problem than that. OpenAI runs three crawlers with three jobs, and the rule blocked one of them:

• GPTBot builds the training corpus. It was refused.
• OAI-SearchBot builds the index behind ChatGPT search. It got through, and fetched 91 pages on our site between 14 and 21 September 2026.
• ChatGPT-User fetches a page live when a question triggers browsing. It got through, and fetched 130 pages in the same window.

OpenAI’s documentation says each setting is independent, so blocking GPTBot does not remove a site from ChatGPT search. What we lost is a place in what the model knows without searching. The GPTBot case study has the raw runs. The lesson is to find out which crawler is refused before you act on it, and our guide on whether AI crawlers can reach your site walks through the checks.

What generative engine optimization looks like in practice

A real programme runs four jobs in a fixed order, because each is wasted without the one before it.

1. Access. Test the search and live-fetch crawlers of each engine against your pages and fix CDN or robots.txt rules. This takes days.
2. Answer-first pages. Open every section with a direct answer that survives being lifted out of the page, and put the number or date in it. This takes weeks.
3. Something original. Publish a figure other people can cite, with its method and date. The studies on this page are our version of that.
4. Outside mentions. Get described, consistently, on sites the engines already read. This takes months, and in real estate it did the most visible work in our data.

Measurement runs under all four, per engine and repeated. One reading tells you nothing because answers vary between runs, and a signed-in profile with memory switched on flatters you. We run this sequence for real estate firms as an AI SEO agency, but the order matters more than who carries it out.

How does generative engine optimization work for a real estate agent or brokerage?

For an agent or brokerage, GEO means being the name an AI engine gives when a buyer or seller asks who to hire in a specific area.

In the Houston study, an estimated 82.6% of the agents, teams and brokerages named appeared in just one of the 40 areas we tested, so the work is won one neighbourhood and one intent at a time. The answers lean on a thin band of intermediaries. Of 7,503 citations, 29.9% went to property portals (Zillow 9.1%, HAR 8.6%, Realtor.com 6.3%) and 13.7% to agent referral networks, while a firm’s own domain was cited in 8.4% to 10.4% of the answers that named it.

Read that carefully, because the study does. It counts which domains an answer cited. An engine can read a site during retrieval without citing it, so the own-website figure is a floor on that site’s influence. What it does remove is the assumption that more posts on your own blog is the lever. Crawler access and answer-first neighbourhood pages still come first, and then a complete, current profile on the portals and referral networks the engines cite does more visible work than another article.

Measurement changes too, because an estimated 83.3% of the agents named appeared on only one of the four surfaces tested. Our GEO for real estate programme runs the sequence area by area and surface by surface, and the guide to AI search for real estate agents covers the agent side in more depth.

What generative engine optimization can and cannot do

It can remove obstacles and improve your odds, and it cannot guarantee a mention or give a young site authority it has not earned.

The honest answer to “is this just SEO with a new name?” is partly yes. For Google’s AI features, Google’s own documentation says there are no additional requirements and no special optimizations needed, and it says indexing and serving are not guaranteed. The parts that are not SEO are measuring answers engine by engine, working out which crawler an engine uses, and earning mentions in the places an engine reads. An SEO report cannot tell you whether ChatGPT names you.

Evidence and vendor claims: what to trust

Trust a claim in proportion to who measured it, how, and who profits from the result.

What our own Search Console shows

This page is a fair test of what we just said, and it has not passed.

In the three months of Search Console data we read on 7 November 2026, Google showed this page 8,875 times and nobody clicked, at an average position of 61. For the exact phrase “generative engine optimization” it appeared 4,375 times at an average position of 69, which is around page seven. The UK spelling did better: “generative engine optimisation” appeared 1,515 times at 47.6, and “how does generative engine optimization work” appeared 137 times at 19.7.

We think the explanation is authority. The page answers the question about as clearly as we know how and still sits where almost nobody scrolls, because a head term this contested is decided by who links to you, and we are a young site with a thin link profile. A rewrite does not fix that, and we would distrust anyone who says it does. It is the same reason corroboration is the slow input in the sections above.

If you would rather see your own numbers than read ours, the free AI visibility audit takes your website address and an email, and within 48 hours sends back 200 AI answers about your market: 50 questions on each of ChatGPT, Perplexity, Gemini and Copilot, every answer screenshotted.

If this page ranks better next quarter, we expect links to be the reason.