What is GEO? A guide to being cited in AI answers

Search is no longer one thing. Ask the same question of Google, ChatGPT and Perplexity and you meet three different surfaces: ten blue links, one written answer, and a handful of sources under it. As that third surface grows, a new question comes with it: when the engine writes the answer, whose pages does it cite, and how does it choose? That question is what GEO is about.

What is GEO?

GEO stands for Generative Engine Optimization. A generative engine is any system that answers a question with prose instead of a list of links: ChatGPT, Gemini, Perplexity, Microsoft Copilot, and the AI Overviews Google places above its own results. These systems assemble an answer from several sources and usually show those sources underneath. GEO is the work of getting your brand into that set.

The shift is subtle but real. In classic search the goal is a click. In a generative answer the user often never clicks: they read, a name stays with them, and later they search for that name or simply pick up the phone. So the return on GEO is not only traffic — it is being named, in the right context, with the right facts.

Why is this being discussed now?

Two things happened at once. Chat interfaces became an ordinary way to ask everyday questions — about prices, about which company to use, about how something is done. And search engines began putting written answers above their own results. What a user once picked out of ten links now arrives as a finished paragraph.

For a business that means something specific: before a customer finds you, they may meet a sentence written about you — and you did not write it. Making that sentence accurate is a matter of making accurate information visible. That is the whole job of GEO: handing the engine correct, consistent, quotable material.

A warning belongs here too. Most of the numbers circulating about this field have unclear provenance. Do not make decisions on industry statistics without your own measurement. There are no industry figures in this article: the one measured result comes from our own panel and carries its date, and the durations come from our own service definitions.

SEO, AEO and GEO: what is the difference?

The three acronyms look alike but target different surfaces, and they stack rather than compete: each one is built on the one before it.

• SEO aims to move your page up the list of results. Measured in positions and clicks.
• AEO (Answer Engine Optimization) aims to be the direct answer — the featured snippet, the voice assistant reply, the AI Overview. Measured by whether the answer is made of your sentence.
• GEO aims to be cited inside the answer a generative engine writes. Measured in mentions and citations.

All three stand on the same ground: a crawlable site, clear writing and verifiable facts. That is why selling GEO as a replacement for SEO is wrong. A site with a broken technical base is not quoted properly by any engine.

How does an engine decide whom to cite?

Source selection is not a single ranking list. Three patterns are observable, and all three involve things a brand can actually control.

1. The same information in several independent places

Seeing a brand on its own website is not enough for an engine to recommend it. Directories, association pages, sector lists, news and client sites — the same facts appearing in different places is what makes those facts trustworthy. Our own measurement showed two versions of this. Asked in English which agency to hire in İzmir, the engine cited agency directories: DesignRush, Semrush Agencies, The Manifest. In the Turkish social-media question it cited platforms like Armut and Edvido. But in the most crowded answer of all — a table of twelve agencies — every firm was cited from its own website. There are two doors into an answer today: a directory record, and your own page.

2. Information that matches everywhere

The second pattern is consistency. When the business name, address, phone and service description are written identically in every source, the engine merges them into one entity; when they differ slightly, it can read them as two businesses. This is the NAP rule of local SEO, carried into generative engines.

3. Content written to be quoted

The third is form. An engine lifts a single paragraph out of a page and drops it into its answer. Out of context, that paragraph still has to be true and complete: restate the question, answer in the first sentence, then explain. A sentence buried mid-article that opens with "as described above" cannot be quoted at all.

Structured identity: becoming one entity to a machine

The technical half of GEO is largely about identity. An engine has to answer "who is this brand, where is it, what does it do" without hesitating, and what makes that possible is the structured data embedded in your pages: the business, its services, its founder and its pages, joined into a graph.

On our own site that graph looks like this: an Organization node for the brand, a ProfessionalService node carrying the address and the map listing, a Person node for the founder, and a WebPage node for each page — all tied together by identifiers. Social profiles and the Google Business Profile listing are declared alongside them, which is a way of saying: at these addresses, this is the same brand.

A mistake illustrates why this matters. In our October measurement we asked Perplexity for our contact details. It returned the right address and phone number — and then asked the user a question, because another business in Germany carries almost the same name. A name collision is the most visible test of structured identity: the more complete the profiles, the directory records and the declared identities, the less an engine hesitates.

How to write answer-first content

Answer-first content means every section takes a question as its heading and answers it in the first sentence. In practice a section is built like this:

1. 1.The heading is a question someone actually types: "How long does a website take in İzmir?"
2. 2.The first sentence answers it; conditions and exceptions come afterwards.
3. 3.The paragraph is true and complete when read alone — no "as above", no "see below".
4. 4.Length is roughly 80–150 words: short enough to fit in an answer, long enough to say something.
5. 5.Any figure, duration or price carries its source or its condition. An invented number, once quoted, becomes a permanent error.

These blocks should not sit inside an accordion. Text behind a closed panel is not read, is usually not quoted, and looks like content hidden from the reader. Keep it in plain sight; the FAQ can still live separately.

What llms.txt does and does not do

llms.txt is a plain-text file at the root of a site that summarises the brand: what you do, where you are, which pages matter. The idea is to do for language models what robots.txt does for crawlers.

Can they read your site at all?

This is the step most often skipped and the cheapest to fix. However well your content is written, it counts for nothing if the crawler cannot reach it. Check:

• Does robots.txt block the AI crawlers — OAI-SearchBot, PerplexityBot, ChatGPT-User, Claude-User, Google-Extended?
• Does a firewall or bot protection drop those crawlers into a challenge? The server logs answer this.
• Is the content rendered on the server, or only after JavaScript runs in a browser? The second means an empty page for some crawlers.
• Is the main text in the first HTML response? Reading the page source is a sufficient test.

Which kinds of content actually get quoted?

Engines quote the content that helps someone decide, not the content that praises the brand. In our own measurement the one clear instance of this was a pricing section; the four kinds below are what we expect from that instance plus our own work and common practice — not a distribution counted from the panel. The record holds the type of the page that was cited; from the November measurement on it will hold the type of the content too.

• Definitions and comparisons — "what is X", "X versus Y". This is the backbone of an answer: the engine takes a definition and rewrites it in its own words.
• Pricing logic. Even without figures, sections that honestly answer "what sets the price" get quoted. In our measurement one competitor was named with its actual numbers because it publishes price bands on its site.
• Ordered steps and checklists. Sections that lay out what to do, in order, feed straight into the "how to" half of an answer.
• Local and concrete facts: which districts you cover, what you delivered in which sector, on what timeline. Generic sentences are what everyone has; these are not.

What an answer leaves out cannot be measured; only what it quotes can. But a sentence that lists adjectives about itself — "innovative, dynamic and customer-focused" — has nothing an answer could use: it answers no question.

GEO for local businesses: the İzmir case

For a local business the centre of gravity shifts. In a national informational query, depth of content decides the outcome; in a "best … in İzmir" query, directories, map listings and reviews join the material the engine holds. In our October measurement, two of the seven service questions cited a directory and two cited the firms' own sites; for the remaining three the source type was not recorded. A small sample — but it shows both doors are open.

So the order for a local business is: the Google Business Profile and reviews first, then identical information in directory and association records, then content that genuinely names the city and the districts. Doing only the third — publishing pages that say "İzmir web design" and nothing else — leaves the engine with no external trace to confirm.

One more detail: when the phone number, the address and the way you work are stated correctly, a user can read a local answer and call without ever clicking. That makes consistent contact details one of the few things in GEO that connects directly to a sale — not a rival to content, but the step that finishes it.

What if another company shares your name?

Name collisions are more visible in generative engines than in classic search: a results page can simply list two links, but a single written answer can merge two businesses — or stop to ask which one you meant. That is exactly what happened in our own measurement.

The fix is to multiply the distinguishing signals the engine can see:

• Use the city and country alongside the brand name consistently — not only in page titles but in directory records too.
• Tie the map listing, the social profiles and the website together; say in structured data that these are the same brand.
• Name the founder on the page and in the structured data. A person's name is more distinctive than a company name.
• Register in sector directories: appearing under the right category makes the distinction easier to draw.

How do you see the traffic AI sends you?

A mention is not a result on its own; the question is whether it turns into a visit and a message. In a default analytics setup a visitor from ChatGPT usually disappears into referral or direct traffic. Two steps separate it out.

1. 1.Group AI assistants into their own channel in analytics: chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com, claude.ai and the rest.
2. 2.Record where the first visit came from, and attach that record to the lead when a form is submitted. Only then can you answer whether assistant traffic actually asks for quotes.

With those two in place, GEO stops being a visibility story and becomes a measured channel. On our own site the source of the first visit — ad click id, UTM parameters and the AI assistant, if any — travels with the form submission.

How is GEO measured?

Unmeasured GEO work is a matter of belief. Measurement runs on a fixed panel of questions, repeated identically every month. The method we use:

• Twenty questions, mixing service, sector, informational and brand queries, some in Turkish and some in English.
• The same questions every month. Change a question and the comparison breaks there, which is why the panel carries a version number.
• Three things recorded per answer: was the brand mentioned, which sources were cited, and who was named instead of you.
• An engine that could not be measured is not counted as "not mentioned". One that requires a login or shows a challenge is marked separately, or next month's comparison is meaningless.
• In analytics, visits arriving from AI assistants are tracked as their own channel — that is the only place you see whether mentions turn into real visits.

How do you read an AI answer?

Every answer we record carries four pieces of information, and each one is useful for something different.

• Were you mentioned: the basic measure. If not, the answer is being assembled from sources that are not you.
• Where in the answer: being named in the first sentence is not the same as appearing in a list at the end.
• Which source was cited: what is carrying you — your own site, a directory, a client's page? This decides next month's work.
• Who was named instead: when the same names keep recurring, looking at where those names appear is the fastest roadmap you will get.

The fourth is usually the most instructive. Across the twelve non-brand questions of our October measurement, three cited a directory (Armut and Edvido; in English DesignRush, Semrush Agencies and The Manifest), three cited competitors' own sites and two cited foreign publications; for the remaining four the source type was not recorded. The most instructive was the hotel question: İzmir Palas Oteli's social media and production are our work, and it still was not enough to put us in the answer — a directory was cited instead. Both doors were open; we held the key to one, our own site.

Our own measurement: the October 2026 baseline

The evidence for this article is our own result. We ran the panel for the first time on 23 September 2026. Of the sixteen questions we could measure on the engine that works without an account, we were mentioned in four — and all four were brand questions. Asked about us by name, we are described correctly. Asked about the service, we appear in no answer at all.

• Brand questions (4): four mentions. Address, phone, service list and client names all correct, sourced from our own site.
• Service questions (7): zero mentions. The answers named dozens of agencies in İzmir; we were not among them.
• Sector questions (2): zero mentions — for hotels and for tour and transfer operators, despite live work in both.
• Informational questions (3): zero mentions. For "what is GEO", all ten cited sources were non-Turkish.

The most instructive answer came from a brand question. Asked whether Maora is a trustworthy agency, the engine listed our portfolio and services correctly — and then hedged, saying it could not find customer reviews or verified references on independent platforms. It is not pleasant to read about yourself, and it is exactly right: on that date our Google Business Profile carried no reviews at all. Generative engines read trust from independent traces like these. Which makes the first item of GEO work social rather than technical: real reviews from real clients.

We will measure this table again every month and update this article. If the numbers rise, they will be written here; if they do not, that will be written here too.

Five questions to ask before buying GEO

The field is new, so the claims are loud. Five questions separate the people doing the work from the people describing it:

1. 1.Which questions, on which engines, how often? If the answer has no question list and no schedule, there is no measurement.
2. 2.What is my baseline today? The first job is measurement, not improvement; without a baseline, no claim of progress can be checked.
3. 3.Which external sources will we register with? Are directories, associations and sector lists part of the plan?
4. 4.Who builds the structured data, and what does it cover — business, services, founder, pages?
5. 5.Do you guarantee results? A proposal that says yes does not know that the engine writes the answer.

A 30-day starting plan

For a business starting from zero, this is a sensible order for the first month. None of it requires an advertising budget.

1. 1.Check crawler access: robots.txt and the firewall should not block AI bots.
2. 2.Complete the Google Business Profile: category, services, real photographs, correct address and phone.
3. 3.Start collecting reviews. No rewards, nothing invented — just ask, and answer every review that arrives.
4. 4.Write the business name, address and phone identically on your site and in every directory.
5. 5.Set up the structured data: business, services, founder and pages, with social profiles declared alongside.
6. 6.Pick the five questions you are asked most and write an answer-first section for each.
7. 7.Open a Bing Places listing. Microsoft's index is known as one of the sources Copilot and ChatGPT's web search draw on; the listing is free and can be imported from your Google profile.
8. 8.Register in your sector's directories — those are usually the sources engines cite.
9. 9.Set up the panel and take the first measurement, even if the result is poor. Without a baseline, progress is invisible.

Three common mistakes

• Filling pages with AI-written content. Engines see the same patterns everywhere; what sets you apart is what only you have — your real pricing logic, real timelines, real cases.
• Betting everything on one file or one piece of markup. GEO works through many small true things stacked up, not one lever.
• Selling or buying "AI optimization" without measuring it. A mention is a measurable thing; with no measurement there is no claim.

GEO is not a new magic replacing search engines. It is three long-familiar things — accurate information, a consistent identity and clear writing — done in a form machines can read. What changed is where you read the result: not in a click count, but in whether your name appears inside the answer.