Generative Engine Optimization (GEO): A Practical Guide to AI Visibility, Readiness and Measurement

Generative Engine Optimization, usually shortened to GEO, is the work of improving whether and how your brand is retrieved, understood, cited and represented inside AI-generated answers. That is a longer definition than most articles give it, and the length matters. Four separate things have to go right. A model has to be able to find your material, it has to understand what your organisation is and what it does, it has to pull your page into the working set for a particular question, and then it has to describe you accurately enough that the description helps rather than harms you.

Because those are four different problems, GEO cannot be reduced to a single file or a single tag. It is not adding llms.txt. It is not pasting in schema markup. Those are useful contributions to a much wider picture that includes crawlability, content structure, entity consistency, evidence, authority and the ordinary technical health of a website.

It is also worth being honest about the nature of the outcome. Nothing described in this article guarantees a citation. Retrieval and answer generation are probabilistic. The same prompt can produce different sources on different days, on different platforms, for different users. What good GEO work does is improve the conditions under which retrieval and citation become likely. That is a real commercial benefit, and it is a different claim from a promise.

GEO is not a ranking trick. It is the practice of making your organisation legible, retrievable and quotable to systems that answer questions instead of listing links.

The easiest mistake is to treat GEO as a replacement for search engine optimisation. It is not. The overlap is large, and most of the foundations are shared.

Classic SEO optimises for a ranked list. The unit of visibility is a position, and the measure of success is a click. GEO optimises for a synthesised answer, where the unit of visibility is a mention, a citation or a recommendation, and there may be no click at all. That changes what you write and how you structure it, but it does not change the plumbing underneath.

What carries over unchanged

• Crawlability and indexability. If a crawler cannot reach and read the page, nothing downstream can happen.
• Content quality and depth. Thin, derivative pages were never competitive and are no more competitive now.
• Authority and reputation. Independent references to your organisation still influence how much weight your claims carry.
• Clear site architecture. Sensible URLs, coherent internal linking and clean canonicals help every consumer of your site, human or machine.
• Performance and mobile experience. Slow, fragile pages are penalised by users first and systems second.

What genuinely changes

• You are writing passages that can be lifted whole, not pages that must be clicked.
• Entity clarity matters more, because a model has to resolve who you are before it can recommend you.
• Ambiguity is more expensive. A vague claim is a claim a system will not repeat.
• Provenance carries weight. Original data and primary sources are easier to cite than a summary of somebody else's summary.

In practice, an organisation with weak technical SEO does not have a GEO problem yet. It has an SEO problem that is also a GEO problem, and the fix is the same work in the same order.

It helps to stop thinking about AI visibility as one event and start thinking about it as a sequence of stages, each of which can fail independently. This is the model that sits behind the audits at GEO Engine Tools, and it is the most useful mental model we have found for explaining results to clients.

1. 01Crawl and discover
2. 02Understand and resolve entity
3. 03Retrieve
4. 04Select and rerank
5. 05Cite or use
6. 06Represent and recommend

Reading the stages

• Crawl and discover. Can automated agents reach the content at all? Robots rules, rendering dependency, authentication walls and broken links all decide this.
• Understand and resolve entity. Given the page, can a system work out which organisation, service, place or person it describes, and match that to what it already knows?
• Retrieve. For a given question, does your content enter the candidate set? This depends on semantic alignment between the question and your wording.
• Select and rerank. Among the candidates, is your page chosen? Clarity, specificity and perceived reliability matter here.
• Cite or use. Is your material actually quoted or linked, or silently absorbed into a generic answer?
• Represent and recommend. Is what the answer says about you accurate, current and commercially useful?

Passing one stage does not guarantee the next. A perfectly crawlable site can fail entity resolution because its name appears three different ways across the web. A well-understood brand can fail retrieval because its pages never use the language customers actually use. A page can be retrieved and still not cited, because a competitor stated the same thing more clearly. This is exactly why single-signal advice is unreliable, and why a diagnosis should tell you which stage is failing.

The following framework mirrors the check categories used by GEO Engine Tools. It is deliberately unglamorous. Most AI visibility problems we see are ordinary website problems that were tolerable when the only audience was a human with patience.

Crawlability and indexability

Robots directives, sitemap accuracy, canonical correctness, HTTP status hygiene, broken internal links, redirect chains and whether important content is reachable without a login or an interaction.

Structured data

Valid, non-contradictory schema on the pages where it applies, describing what the page really contains rather than a wishlist.

Identity and entity consistency

One name, one legal entity, one address, one set of contact details, expressed the same way on your site, in your markup and on the third-party profiles that describe you. Conflicts here quietly undermine everything else.

Content clarity

Descriptive headings, clean definitions near the top of a page, specific service descriptions, plain statements of what you do, where you do it and for whom.

Trust and evidence

Named authors where relevant, dates, sources, original material, verifiable claims and case studies with real detail rather than adjectives.

Internal architecture

A logical hierarchy, meaningful anchor text, related content properly linked, no orphaned pages and no near-duplicate pages competing with each other.

Machine-readable surfaces

A working sitemap, sensible robots rules, clean feeds where you have them, and llms.txt as a supporting summary rather than a substitute for the site itself.

Rendering and performance

Whether primary content exists in the served HTML or only appears after client-side JavaScript executes, plus mobile performance and layout stability.

Platform limitations

Some site builders and legacy themes constrain what you can control, particularly around head markup, structured data and server rendering. It is better to name that constraint honestly than to keep prescribing fixes the platform will not allow.

There are two sensible depths at which to look at this, and confusing them leads to poor decisions.

The free AI Readiness Quick Test at GEO Engine Tools is a fast technical indication. It looks at a page and its immediate signals and tells you whether the obvious foundations are in place. It is the right tool for a first opinion, for a sanity check before a meeting, or for deciding whether a deeper look is worth paying for.

A Complete Website Audit is a different exercise. It samples the pages that actually matter across the site, evaluates patterns rather than one-off issues, records the evidence behind each finding, and reports coverage so you can see how widespread a problem is. That distinction between an isolated fault and a site-wide pattern is usually the difference between a two-hour fix and a rebuild conversation.

Neither can prove that an AI system mentions you. A readiness assessment measures the observable condition of your website. Actual AI presence is a separate measurement, discussed further below. Any tool that claims to score both from one crawl is overstating what it can see.

A readiness score is only useful if it is repeatable and explainable. GEO Engine Tools scores from observable website signals and shows why each check passed or failed, so the result can be argued with, verified and fixed. If a score cannot be traced back to a specific observation on a specific URL, it is an opinion wearing a number.

Two consequences follow. First, the same website in the same state should produce the same score, which rules out scoring methods that depend on a model's mood. Second, the age of a website must never itself affect the score. A site built in 2016 that renders clean server-side HTML, carries valid schema and loads quickly is in better condition than a 2026 build that hides its content behind a client-side framework and contradicts its own business identity.

Score the evidence, not the era. Website age is context for a decision, never an input to a technical result.

This is the question we are asked most often, usually by an owner who has been told by somebody that their site is too old to work with AI. The honest answer is that age is guidance only. A ten-year-old website can be technically excellent, and a brand new one can be badly built. The decision should rest on evidence.

The evidence that should decide it

• Rendering dependency. Does the meaningful content exist in the HTML, or only after scripts run?
• Structural limitations. Can the platform express clean URLs, head markup, canonicals and structured data at all?
• Identity conflicts. Does the site contradict itself or your third-party profiles on name, address, services or ownership?
• Structured data. Present, valid and truthful, or absent and unaddable?
• Mobile performance. Real-world speed and stability on a phone, not a lab score in isolation.
• Duplication. Multiple near-identical pages, legacy campaign clones, parameter variants and thin tag archives.
• Crawlability and broken links. Accumulated rot across years of edits and migrations.
• How widespread the problems are. One broken template is a fix. The same fault on every page is a pattern.

Four outcomes, not two

GEO Engine Tools resolves that evidence into one of four recommendations, and framing the decision this way removes most of the emotion from it.

• Maintain. The site is in good condition. Keep it healthy and spend your budget on content and evidence instead.
• Optimise. The foundations are sound and the gaps are specific. Fix schema, headings, internal links, metadata and performance in place.
• Modernise. The structure holds but parts of the stack are holding you back. Upgrade templates, rendering or key sections without starting again.
• Consider rebuilding. The platform or architecture prevents the fixes the evidence demands, and the cost of working around it exceeds the cost of a new build.

Most sites we assess land on Optimise. Rebuilding is a legitimate answer, but it should be the conclusion of a diagnosis rather than the opening pitch. If you would like that decision made with evidence, we are happy to run through it with you on the contact page.

llms.txt is a plain text file at the root of a site that summarises what the site is and points to its most important pages. We publish one, we think it is worth having, and it is routinely oversold.

What it can genuinely help with

• Giving a concise, authoritative summary of your organisation in one place
• Pointing to the pages you consider canonical for each topic
• Reducing ambiguity about what you do and where you do it
• Acting as a stable machine-readable surface alongside your sitemap

What it is not

It is not a ranking file. Adoption is not universal, support varies by platform and it is not a documented input to any ranking or citation decision. It does not replace crawlability, it does not replace content, it does not replace structured data and it certainly does not replace authority. A site with an immaculate llms.txt and unreachable content is still invisible. Treat it as a courtesy note to a reader who may or may not open it, and put the real effort into the site it describes.

Structured data is a standardised way of stating facts about a page in a form software can parse without guessing. It does not communicate privately with a language model, and it does not instruct one. It reduces the amount of interpretation required, which lowers the chance of being misunderstood or ignored.

The types that usually earn their place

• Organization or LocalBusiness. Your identity, and for a business with premises, its location, hours and contact details. LocalBusiness only where it is genuinely accurate.
• Service or Product. What you actually sell, described specifically enough to be matched to a need.
• Article. Editorial content, with author, publication date and modification date.
• FAQPage. Only where the questions are real and the answers are useful on the page itself.
• BreadcrumbList. Position within the site hierarchy, which helps establish structure.
• WebSite. Site-level identity and search behaviour where applicable.

Two rules keep schema honest. It must match what a human sees on the page, and it must not contradict itself across the site. Invalid, exaggerated or duplicated markup is worse than none, because it introduces the ambiguity you were trying to remove.

If you want to be quoted, write things that can be quoted. That is less about style than about structure. A passage is extractable when it stands alone, states something specific and does not need three paragraphs of context to make sense.

Practices that make content easier to lift

• Define the term early. Put a clean one or two sentence definition near the top of the page.
• Lead with the answer, then expand. Do not bury the conclusion beneath the reasoning.
• Use descriptive headings that read like questions or claims, not labels like Overview.
• Add FAQs only where somebody genuinely asks the question, and answer it fully in the answer.
• Prefer factual statements to adjectives. Specific numbers, named places, real constraints.
• Use tables and lists where the content is genuinely comparative or enumerable.
• Publish original data, case studies and frameworks. Original material is easier to attribute.
• Date your content and update the date honestly when you revise it.
• Name entities in full at least once. People, companies, places, products and standards.
• Attribute sources, and link to the primary one rather than a repost.

None of these are new writing advice. They are simply the parts of good editorial practice that happen to matter more when your reader is summarising rather than browsing.

Answer engines have an accuracy problem and they know it. Anything that makes a claim easier to verify makes it safer to repeat. That is the practical reason provenance matters commercially, rather than a moral one.

Original evidence is the strongest asset here. If you have measured something, publish the measurement and the method. If you have delivered a project, describe what was actually built, when, and what changed. If you are relying on somebody else's finding, link to their original publication rather than an article about it. And if you do not know a number, do not invent a plausible one, because unverifiable specifics are the fastest way to make a page look untrustworthy to both a person and a system.

The same logic applies to your own identity. Consistent, corroborated facts about your organisation across your site, your markup and independent profiles make you a safer entity to name. Contradictions make you a risk, and a system with a choice will usually choose the source it does not have to reconcile.

There are two distinct things people mean by measuring GEO, and conflating them produces bad reporting.

Technical AI readiness

The observable condition of your website against the framework above. This is deterministic, repeatable and improvable, and it is what a readiness audit measures. It tells you whether you have removed the obstacles within your control.

Actual AI presence

Whether AI systems mention, cite and describe you in practice. This requires querying the platforms themselves over time, and the useful metrics look like this.

• Brand mentions. How often you are named in answers to relevant questions.
• Source citations. How often your pages are linked or referenced as a source.
• Citation share. Your share of cited sources against named competitors for the same prompts.
• Accuracy of representation. Whether what is said about you is correct and current.
• Change over time. Direction of travel across repeated prompts, dates and platforms.

Two caveats belong in every report of this kind. Results vary by platform, because the systems retrieve differently. Results also vary between runs on the same platform, because generation is probabilistic. A single prompt on a single day is an anecdote, not a measurement. Repeat the prompts, record the dates, and read the trend rather than the instance.

It is tempting to talk about AI search as one thing. It is not. The major systems differ in how they discover sources, when they choose to search at all, and how visibly they attribute what they use.

• ChatGPT. Very widely used for research. It may answer from model knowledge or run a live search depending on the question, and its browsing and citation behaviour has changed repeatedly.
• Google Gemini and AI search experiences. Built on Google's own index and infrastructure, which makes traditional technical SEO health directly relevant, while synthesis and source selection behave differently from the classic ten blue links.
• Claude. Common in professional and knowledge work contexts, with its own search behaviour and a generally conservative approach to unverified claims.
• Perplexity. Search-first by design and unusually explicit about citing sources, which makes it a useful place to observe which of your pages are actually being pulled.

Do not assume a tactic that appears to help on one platform transfers to another. The sensible strategy is to fix the shared foundations, which help everywhere, and then observe each platform separately rather than optimising for whichever one you happened to test.

Visa-Pattaya is a useful illustration of the framework applied end to end, and we present it as an implementation example rather than a visibility result. We have not published measured AI citation data for it, so we will not claim any.

What was implemented

• Entity-first information architecture, with one consistent identity across the site and its markup
• Structured data across service pages, with markup that matches the visible content
• Answer-shaped content, with clean definitions and genuine questions answered directly
• Original guides to Thai visa categories rather than rewritten summaries of other sites
• Machine-readable surfaces, including a working sitemap, sensible robots rules and clean canonicals
• Server-rendered content, so the substance of every page exists in the HTML

The reason to describe it this way is that the implementation is verifiable and the outcome would need measurement. That is the standard we would ask any agency to hold itself to, including us. You can see the full case study in our work, and the same approach underpins our digital services.

None of this requires a rebuild to begin. In rough order of return on effort:

• Run a free AI Readiness Quick Test at GEO Engine Tools and read the evidence behind each failed check.
• Confirm your important content is present in the served HTML without JavaScript.
• Fix crawl basics. Robots rules, sitemap accuracy, canonicals, broken links and redirect chains.
• Standardise your identity. One name, one address, one set of details, everywhere including third-party profiles.
• Add or correct Organization schema, and LocalBusiness only if you genuinely have premises.
• Add Service or Product schema to the pages that describe what you sell.
• Rewrite the top of your five most important pages to define plainly what you do, for whom and where.
• Add a short, honest FAQ section to service pages where real questions exist.
• Publish one piece of original material. Data you have, a project you delivered, a method you use.
• Tighten internal linking with descriptive anchor text and remove near-duplicate pages.
• Publish an accurate llms.txt, then stop thinking about it.
• Ask ChatGPT, Gemini, Claude and Perplexity about your brand and your category. Record what is said, what is wrong and who is cited instead of you.
• Repeat that prompt set monthly and keep the answers, so you have a baseline to compare against.

If the audit surfaces problems your current platform cannot fix, that is the point at which a rebuild conversation becomes rational rather than speculative.

"llms.txt guarantees AI visibility"

It does not. It is an optional, unevenly supported summary file and no platform documents it as a ranking or citation input. Useful, minor, not decisive.

"Schema guarantees citations"

Structured data reduces ambiguity. It does not create authority, relevance or trust, and it cannot make a system quote a page it never retrieved.

"We rank first on Google, so AI will cite us"

Sometimes, and sometimes not. Answer engines select and rerank sources with their own logic, and several of them do not use Google's index at all. Strong rankings help. They do not transfer automatically.

"More content is always better"

Volume without substance adds duplication, dilutes your entity and gives systems more chances to retrieve your weakest page. Fewer, better, more specific pages consistently outperform bulk.

"A high readiness score means AI already recommends us"

A readiness score measures the condition of your website. Presence in AI answers is a separate measurement taken from the platforms themselves. Readiness improves your odds. It is not evidence of the outcome.

GEO rewards patience and evidence rather than tricks. The organisations that will be named as defaults inside AI answers over the next few years are the ones fixing their foundations now, publishing material worth citing and measuring what is actually said about them.

Begin with a diagnosis. The free AI Readiness Quick Test at GEO Engine Tools gives you a fast technical indication, and a Complete Website Audit samples the pages that matter, identifies site-wide patterns with the evidence behind them and resolves the result into a clear recommendation to maintain, optimise, modernise or consider rebuilding.

If the fixes need building rather than listing, that is our side of the work. We handle schema, rendering, structure, content architecture and implementation on both modern AI-assisted builds and existing WordPress sites. Have a look at our digital services or get in touch and we will talk through what your evidence actually says.

Technical readiness is what you control. AI visibility is what you measure. Do not let anybody sell you the first as proof of the second.