AI visibility
The GEO guide: getting cited by AI answer engines
Generative engine optimization is the work of becoming a source that AI assistants name and recommend. Here is what the research actually shows, and where the honest limits of it are.
What is generative engine optimization (GEO)?
Generative engine optimization, usually shortened to GEO, is the practice of making a business or a page more likely to be surfaced, quoted and attributed inside AI-generated answers, such as ChatGPT responses, Perplexity answers and Google's AI Overviews.
The term comes from a 2024 research paper by Aggarwal and colleagues, presented at KDD, which tested content changes against generative engine responses and reported visibility improvements of up to 40% from the methods it studied.
GEO overlaps heavily with SEO rather than replacing it. The differences are that it optimises for being named inside an answer rather than for a click, and that no engine publishes how it selects or attributes sources, so no one can guarantee a citation.
Where the term came from, and what it means
GEO is not a vendor coinage. It was introduced in a peer-reviewed paper, which is unusual for a marketing term and worth knowing, because it means there is something real underneath the hype.
In 2023 a team of researchers published a paper titled GEO: Generative Engine Optimization, later presented at KDD in 2024. They built a benchmark of real user queries, tested a set of content modifications against generative engine responses, and measured how visible a source became inside the generated answer. The headline result was that the methods they tested could improve visibility in generative engine responses by up to 40%.
Read the full breakdown: Where the term came from, and what it meansHide the full breakdown: Where the term came from, and what it means
What the paper found useful is instructive and slightly unfashionable. The changes that moved visibility most were things like adding citations to authoritative sources, including relevant quotations, and adding statistics, rather than keyword density tricks. In other words, the things that make a passage genuinely quotable also make it more likely to be quoted.
That is the whole idea in one sentence. A generative engine is assembling an answer from sources. Your job is to be the source that is easiest to assemble from: clearly written, factually checkable, unambiguous about who you are, and structured so a self-contained chunk of it can be lifted without dragging in three paragraphs of context.
It is worth separating GEO from answer engine optimization, because the two get used interchangeably and they are not quite the same job. AEO is about structuring a page so an engine can extract a clean answer from it, which wins featured snippets, voice answers and summaries. GEO is about becoming an entity that generative systems recommend by name, which depends as much on what the rest of the web says about you as on what your own page says.
The click is disappearing, and it is measurable
The reason GEO matters now is not that AI is exciting. It is that the mechanism the whole SEO industry was built on, a ranking producing a click, is carrying less traffic than it used to. SparkToro's analysis of a Similarweb clickstream panel measured US Google searches that ended without any click at all.
SourceSparkToro, zero-click search analysis 2026 (Similarweb clickstream panel)
Some of that is perfectly benign. A searcher who gets your opening hours from your Business Profile got what they wanted, and you may well get the visit. But a searcher who gets a full answer to a research question and never reaches anyone's website is a genuine loss of contact, and it is the pattern GEO exists to respond to.
What an AI summary does to where people click
The strongest evidence here is not from a marketing vendor. Pew Research Center studied the actual browsing behaviour of 900 US adults who shared their data, covering 68,879 unique Google searches in March 2025, of which 12,593 produced an AI summary.
Pew found that when an AI summary appeared, 8% of users clicked a traditional search result, compared with 15% when no summary appeared. Clicking a link inside the summary was rarer still, at 1% of visits. That is the clearest published picture of what an AI answer does to onward traffic, and it comes from observed browsing rather than from a survey or a modelled estimate.
Two conclusions follow, and they pull in different directions. The first is that traffic from queries that trigger AI answers is genuinely smaller, so a strategy that measures success purely in sessions will look like it is failing even when the business is doing fine. The second is that being named inside the answer is now worth something distinct from being clicked, because a recommendation read and acted on later does not leave a referral trail at all.
This is also why measurement in this area is honestly difficult, and anyone telling you otherwise is selling something. A person who asks an assistant for a recommendation, hears your name, and calls you the next day arrives in your records as a direct call with no source. That is not a tracking failure you can fix with better tags.
How much clickthrough the top positions lose
Ahrefs compared aggregated Search Console data across 300,000 keywords, 150,000 with AI Overviews present and 150,000 without, between December 2023 and December 2025. Their 2026 figures show the loss concentrated at the top of the page.
SourceAhrefs, AI Overviews and clickthrough rate, 2026 (300,000 keywords)
Note that this is a vendor study of aggregated accounts rather than independent research, and that Ahrefs' own earlier measurement of the same effect, eight months prior, was 34.5%. Whichever figure you take, the direction is consistent and the practical implication is the same: the value of position one is falling on the queries where AI answers appear, and the response is not to chase position one harder.
How engines choose what to cite, and what nobody actually knows
Here is the thing most pages on this topic will not tell you. No generative engine publishes how it selects, ranks and attributes the sources inside an answer. Google documents that AI features exist, describes broadly that they draw on its index and that ordinary search best practices apply, and explicitly does not publish a ranking mechanism for them. OpenAI, Anthropic, Perplexity and the rest publish less than that. Anyone presenting a definitive ranking model for AI citations has inferred it, and should say so.
What can be said with reasonable confidence, because it follows from how these systems are built rather than from a leaked formula, is this. These systems work from text they can retrieve and from a model's learned associations. Content that is clear, specific and self-contained is easier to retrieve a usable passage from. An entity that is described consistently across many independent sources is easier to identify with confidence. A claim with a checkable source attached is safer for a system that is penalised for being wrong.
Read the full breakdown: How engines choose what to cite, and what nobody actually knowsHide the full breakdown: How engines choose what to cite, and what nobody actually knows
That is genuinely useful guidance and it is also unglamorous. It is more or less the advice that good editors have given for a century, with the addition that a machine is now one of your readers. The GEO paper's own findings point the same way: citations, quotations and statistics improved visibility more than stylistic optimisation did.
What cannot be said is that any specific action will get you cited. We will not tell you otherwise, and we would be suspicious of anyone who does. The honest promise is that the work is designed to make citation more likely, that it is the same work that improves the page for human readers, and that it is measurable in aggregate over time rather than guaranteeable on any single query.
How to do generative engine optimization
Ordered so that each step makes the next more effective. Nothing here is a trick, and nothing here depends on a mechanic we cannot evidence.
Be unambiguous about who you are
A generative engine has to be confident that the business it is describing is one business before it will name it. That means one consistent name, address and phone number across your site, your profiles and the directories that quote you, and an entity description that does not drift from surface to surface.
- One canonical business name, used identically everywhere
- Organization and LocalBusiness structured data that matches the visible page
- The same description of what you do on every profile you control
Write answer-first, in self-contained chunks
Lead each section with a direct answer of two or three sentences, then expand underneath. A passage that is correct and complete when lifted out of its page is worth far more than a paragraph that only makes sense after the two above it.
- Use the real question as the heading, phrased the way people ask it
- Answer immediately underneath, before context and caveats
- Make each section survive being read alone
Attach sources to every claim that carries a number
This is the finding from the research that most people skip. Citations, quotations and statistics improved visibility in the original GEO study more than surface-level rewriting did. It also happens to be the thing that keeps you honest.
- Name the publisher and year in the sentence, not only in a footnote
- Link the primary source, not a blog summarising it
- Cut any number you cannot trace to a named publisher
See the remaining steps: How to do generative engine optimizationHide the remaining steps: How to do generative engine optimization
Add structured data, and keep it truthful
Structured data does not make an engine cite you and it is not a ranking boost. What it does is remove ambiguity about what a page is, who published it and what it asserts, which is useful to any system trying to summarise you correctly.
- Organization, LocalBusiness, Service, FAQPage and Article where each genuinely applies
- Markup that matches the visible content exactly, with no invented fields
- No review or rating markup that you did not collect yourself
Earn mentions on sources that are not yours
What other people say about you is a large part of how confidently any system can describe you. Genuine coverage, accurate directory records, professional associations and real reviews all contribute to an entity being well enough attested to be recommended.
- Accurate listings on the platforms your industry actually uses
- Coverage earned by being worth covering, not bought
- A steady flow of genuine first-party reviews
Measure visibility, not just traffic
Ask the assistants the questions your customers ask, on a schedule, and record whether you are named and who is named instead. This is a sampling exercise rather than a metric, because answers vary between sessions and users, but the trend over months is informative.
- A fixed list of real customer questions, asked repeatedly over time
- Record who gets named, not only whether you do
- Track branded search and direct enquiries alongside it
None of these steps is exotic, and that is the point. GEO done honestly is mostly editorial discipline, entity hygiene and earned authority, applied with an awareness that a machine is now reading alongside the human.
What makes a page easy to quote
Apply this to any page you want an engine to be able to use. It is also, not coincidentally, a checklist for a page that a hurried human can read.
- The first two or three sentences answer the page's core question completely
- Headings are real questions, phrased the way a person would type them
- Each section can be read on its own without the section above it
- Every statistic names its publisher and year in the visible sentence
- Definitions are stated plainly before any elaboration
See the full checklist: What makes a page easy to quoteHide the full checklist: What makes a page easy to quote
- Comparisons are laid out as comparisons rather than buried in prose
- The business, its locations and its services are named explicitly, not implied
- Structured data is present and matches what a reader can see
- Dates are shown, and the content is actually kept current
- Nothing on the page asserts a claim the page cannot source
The last item does the most work. A page that overstates is a page a careful system has reason not to repeat, and a page a careful reader has reason not to trust.
Being one business, described one way
Most of the difficulty a machine has in describing a small business comes down to a single problem: it cannot tell whether the records it has found describe one entity or several.
Consider a practice trading under a shortened name on its website, its full legal name on one directory, a previous partnership name on an old professional listing, and two phone numbers because the second line was added in 2021 and never propagated. Every one of those records is defensible on its own. Together they make it hard to be certain they are one organisation.
Structured data helps here, not because it is a ranking signal but because it states relationships explicitly rather than leaving them to be inferred: this organisation, at these locations, offering these services, with these profiles. That is exactly the kind of assertion a system trying to summarise you needs.
The practical work is unromantic and finite. Decide the canonical form of your name, address and phone. Fix the largest and most frequently quoted records first. Make your own site the clearest and most complete statement of what you do, because it is the one source you control entirely.
SEO, AEO and GEO side by side
These get used interchangeably and they are not the same. The distinction is worth holding because it determines what you optimise and what you measure.
| Item | SEO | AEO | GEO |
|---|---|---|---|
| Goal | Rank, and earn the click | Be the extracted answer | Be the named recommendation |
| Surface | Organic results and the map pack | Featured snippets, voice, AI Overviews | Assistant answers and AI summaries |
| Main lever | Relevance, authority, technical health | Structure and answer-first writing | Entity clarity and corroborated authority |
| Measured by | Rankings, sessions, conversions | Snippet and summary capture | Whether you are named, sampled over time |
| How well documented | Extensively, by Google | Partly, through snippet guidance | Barely, by anyone |
| Can it be guaranteed | No | No | No |
The bottom row is not a joke. None of these can be guaranteed, and the third least of all. The reason to do the work is that the inputs are the same inputs that make a business easier to find, describe and trust, which pays whether or not any particular engine names you this week.
The specific questions inside GEO
Each of these takes one question further than a guide can. Start with the first if you want the practical version of this page.
How to get your business cited by ChatGPT
The practical version of this guide, focused on being named rather than being ranked.
How to get your business cited by ChatGPTWhat is answer engine optimization
Structuring a page so an engine can lift a clean answer from it, and how that differs from GEO.
What is answer engine optimizationSEO vs AEO vs GEO
The three jobs compared properly, with what each one actually optimises for.
SEO vs AEO vs GEOContent that ranks and gets cited
How one well-built page can rank, win the answer box and be quoted by an assistant.
Content that ranks and gets citedWhat is schema markup
What structured data does, what it does not do, and which types are worth the effort.
What is schema markupThe local SEO guide
The foundations GEO sits on for any business that serves a place.
The local SEO guideGEO is present tense, but the claims around it are not
It is reasonable to ask whether this is a service or a bandwagon. The most useful data point is that it is already normal practice rather than an experiment: The CMO Survey, fielded with 308 US marketing leaders in January 2026 by Duke's Fuqua School of Business, found roughly four in ten companies already doing generative engine optimization. Whatever you think of the label, the work is being done.
Consumer behaviour is moving in the same direction on the local side. BrightLocal's 2026 survey of 1,002 US consumers found 45% saying they use ChatGPT or similar tools for local business recommendations, and 40% saying they trust AI platforms for those recommendations. That is self-reported and it is a steep year-on-year change, so treat it as a direction rather than a precise measurement, but the direction is not ambiguous.
Read the full breakdown: GEO is present tense, but the claims around it are notHide the full breakdown: GEO is present tense, but the claims around it are not
What is not established is any of the specific mechanics being sold around it. There is no published ranking factor list for AI citations. There is no verified technique that reliably produces one. There is no credible way to promise a client that an assistant will recommend them by a given date. When you read a GEO pitch, the test is simple: does it describe work, or does it promise a position?
Our position is that GEO is worth doing, that the honest version of it overlaps almost entirely with doing SEO and editorial work properly, and that the additional pieces are entity clarity and a genuine discipline about sourcing. We aim to make citation more likely. We do not, and cannot, promise it.
Want to know what assistants currently say about you?
We can sample the questions your customers actually ask and show you who gets named instead.
Straight answers.
Has AI made SEO pointless?
No, and the data does not support that framing. Search volume has not collapsed, and generative answers are substantially assembled from the same web content that search ranks. What has changed is that a smaller share of searches produce a click, so the value of a ranking is increasingly partly indirect.
The practical shift is in measurement rather than in method. If you judge the work purely on sessions you will misread what is happening, because being named in an answer produces business that arrives with no referrer attached.
How do I get ChatGPT to recommend my business?
There is no setting, no submission form and no reliable technique, and anyone offering one is guessing. What you can do is make yourself easy to identify and easy to describe accurately: consistent details everywhere, a site that states plainly what you do and where, genuine reviews, and accurate records on the platforms that get quoted.
Then sample it. Ask the questions your customers ask, write down who gets named, and repeat monthly. That tells you where you stand far better than any promise made in advance.
Is structured data worth adding for AI visibility?
It helps in the sense that it removes ambiguity about what your page says and who published it, which is useful to anything trying to summarise you. It does not work as a ranking lever, and no engine has published that structured data influences whether you are cited.
Mark up what is genuinely on the page and nothing else. Markup that does not match the visible content is a quality problem, not an advantage.
Is it better to block AI crawlers or let them in?
It depends on what you are optimising for, and it is a genuine trade-off rather than a best practice. Blocking reduces the chance of your content being used without a visit. It also reduces the chance of being named when someone asks for a recommendation in your category.
For most local service businesses, being findable and recommendable is worth more than protecting the text of a service page. For publishers whose product is the content itself, the calculation is different. Decide deliberately rather than by default.
When should I expect GEO work to show anything?
There is no established timeline, because there is no published mechanism to reason from. What we can say is that the underlying inputs, entity consistency, earned mentions and genuinely useful published content, are the same slow-moving inputs SEO depends on, and they compound over months rather than days.
Be sceptical of any specific window. Nobody has the data to support one.
Does GEO work for a local business, or only for publishers?
It applies to local businesses, arguably more directly. A large share of assistant queries about local services are asking for a recommendation, which is a question about entities rather than about documents, and entity clarity is something a small business can genuinely get right.
The foundations overlap almost completely with local SEO: an accurate profile, consistent details, real reviews and a site that says plainly what you do and where you do it.
How do I measure whether any of this is working?
With a combination of sampling and indirect signals, and with honesty about the limits. Sample assistant answers for your real customer questions on a schedule and record who is named. Watch branded search volume, which tends to move when people are hearing your name somewhere you cannot see.
Then watch the numbers that represent money: calls, forms and booked work. If those are rising while sessions are flat, that is consistent with visibility you cannot directly observe, and it is a more useful signal than any dashboard claiming to measure AI share of voice precisely.
Where this comes from.
Primary documentation and published research behind the guidance on this page.
- Google Search Central: AI features and your website (opens in a new tab)What Google publishes about AI Overviews and AI Mode, and what it does not.
- Google Search Central: Creating helpful, reliable, people-first content (opens in a new tab)Google's published self-assessment questions for content quality and experience.
- Google Search Central: Introduction to structured data markup (opens in a new tab)
- Google Search Central: Google Search Essentials (opens in a new tab)
- Aggarwal et al., GEO: Generative Engine Optimization, arXiv and KDD 2024 (opens in a new tab)The peer-reviewed paper that introduced the term GEO.
- Pew Research Center, browsing-panel study of Google AI summaries, 2025 (opens in a new tab)Observed browsing behaviour, not a survey.
Talk to the team
A short call, a look at how the business currently shows up, and a straight answer on what we would do first.