Generative Engine Optimization: Measure and Improve AI Visibility

Traditional search engine optimization (SEO)Search Engine Optimization (SEO) covers all measures that help a website become more visible in unpaid (organic) search results.Open glossary entry → asks: Does my website rank in Google?
Generative Engine Optimization also asks: Do AI assistants name my brand in their answers, and do they link my website?

Many people now also ask their questions in ChatGPT, Gemini, Claude or Perplexity. They get a written answer instead of a list of results. Brands, providers and sources appear in that answer, or they are missing.

RankScan measures this visibility with the AI VisibilityAI Visibility describes whether and how visible a brand, website or person is in AI-generated answers. It measures mention, position, description and source linking across multiple platforms and prompts.Open glossary entry → Score. This article explains what Generative Engine Optimization (GEO)Generative Engine Optimization (GEO) covers measures that improve the chances of a brand or website being found, understood, extracted, classified and used as a source by generative AI answer systems. GEO extends classic SEO rather than replacing it.Open glossary entry → is, how the score is calculated and how to read it. It is a foundation article: the score is not a measure, but the goal above the measures in the AI presence area.

GEO does not guarantee a mention in AI answers. GEO improves the conditions for AI assistants to find your content, understand it and use it as a source.

What Is Generative Engine Optimization? #

Generative Engine Optimization (GEO) covers measures meant to improve the visibility of a brand, website or source in generative answer systems.

Such systems generate answers instead of just listing links. Many of them retrieve web sources, summarize them and name some of them as sources.

The term comes from a research paper published in 2023. It describes generative engines as systems that combine information from several sources into one answer with large language modelsA Large Language Model (LLM) is an AI model that learns from large amounts of text and generates formulated answers from it. LLMs form the core of generative answer systems such as ChatGPT, Perplexity or Gemini.Open glossary entry → (LLM). For website owners this means: visibility no longer depends only on whether a URL ranks. It also depends on whether an answer picks up the content and names the source.
Source: arXiv: GEO: Generative Engine Optimization

In practice, GEO asks new questions:

Not only:
Which keywordA keyword is the search term users enter into a search engine to find information, products or providers. Keywords are the basis against which rankings and visibility are measured.Open glossary entry → does my page rank for?

But also:
Do AI answers name my brand?
Do they link my website?
Which brands do they name more often?
Which sources do they use?

SEO, AEO, GEO, LLMO: What Is the Difference? #

The terms overlap. But they set different priorities.

A page that is accessible, clearly structured and well supported has better conditions for all three disciplines.

A top ranking in Google does not guarantee a mention in an AI answer. And a mention does not guarantee a click.

How AI Answer Systems Use Content #

Many AI assistants combine a language model with web search or other sources. How exactly they do this differs by provider and is often not public. Website owners can still distinguish four stages. The stages are general expertiseExpertise is the subject-matter competence behind a piece of content. It is the second E in E-E-A-T and should be visible through author details and supporting evidence.Open glossary entry →, not a RankScan measurement.

1. Access #

Can the system retrieve your content? Obstacles include a block in robots.txtThe robots.txt is a publicly accessible text file in a site's root directory. It tells crawlers which URLs they may fetch and therefore controls crawling.Open glossary entry →, a firewall, a login, an error status or content that only appears after JavaScript.

OpenAI documents several crawlers such as GPTBot, OAI-SearchBot and ChatGPT-User. They serve different purposes and can be controlled via robots.txt.
Source: OpenAI Platform: Overview of OpenAI Crawlers

Perplexity documents the crawlers PerplexityBot and Perplexity-User.
Source: Perplexity Docs: Perplexity Crawlers

2. Understanding #

Does the system recognize what the page is about? A clear heading, subheadings, semantic HTML, internal linksInternal linking refers to links that connect pages within the same website. They control how search engines and users find and thematically classify content.Open glossary entry → and a clear description of your brand help.

3. Extraction #

Can the system take over concrete statements? This works better with definitions, direct answers, tables, figures with a source and short sections that make sense on their own.

4. Selection and Attribution #

Does the system choose your page as a source and link it? Topical expertise, evidence, mentions on other websites and the strength of competitors matter here.

The AI Visibility Score in RankScan #

The AI Visibility Score shows which share of the achievable visibility in AI answers your brand has won. It ranges from 0 to 100. RankScan calculates it for the whole property and for every topic.

Which Answers Count #

RankScan regularly sends the prompts of your topics to the AI assistants you choose: 2 to 4 of ChatGPT, Gemini, Claude, Perplexity and Copilot, depending on your plan. Prompts are the questions that RankScan asks.

Only these count towards the score:

• Answers to prompts without your brand name. An answer to a prompt with your name almost always names your brand. RankScan shows it separately as “Brand awareness”.
• Real answers. Follow-up questions and refusals do not count.
• Answers from the last seven days. If that period holds fewer than three scan days, the window reaches back to the third most recent scan day.

Points per Answer #

Each answer earns points for each brand:

So one answer earns at most 100 points. The link counts even without a mention. A link to a page that was unreachable when checked does not count. A brand's points are the average over all counted answers.

Leader, Reference and Target #

The score compares your points with two yardsticks:

• Leader: the brand with the most points. The leader can be a tracked competitor or a brand that RankScan classifies as a provider in your market. Associations, unrelated companies and plain terms are never the leader.
• Reference 20: If the leader has fewer than 20 points, RankScan calculates against 20. This way nobody reaches 100 just because nobody in the market is well visible. The comparison table then says so explicitly.
• Target 60: 60 points are an average across all counted answers. If you are named in every answer, you reach at least 60. But the same average can also come about if you are missing from one answer: three answers with 100, 80 and 0 points also give 60.

The formula:

Score = 100 × ( ½ × min(1, your points ÷ max(leader, 20))
+ ½ × min(1, your points ÷ 60) )

The first half says how close you are to the leader. The second half says how close you are to the target. You reach 100 if you are at least tied with the leader and reach around 60 points on average. You do not need a lead over the leader. Even at 100, your brand can be missing from single answers. The sentence on the score card shows how often it is named: “named in X of Y AI answers to prompts without the brand name”. This way even a leader sees how much room is left.

Each Assistant with Its Own Leader #

RankScan calculates the score for each AI assistant separately, against that assistant's leader. The AI Visibility Score is the average of these values. An assistant only counts from three answers in the window. If none reaches that number, RankScan calculates over all answers together.

The reason: a brand can lead in one assistant and be almost missing in another. A single value over all answers would hide that gap. The score card therefore names the assistant with the largest gap and its leader.

Example with Made-Up Numbers #

The AI Visibility Score is the average of 40, 20 and 75, so 45. In Claude, the brand is ahead of the leader but reaches only half of the target. The largest gap is in Gemini.

How RankScan Rates the Score #

AI Visibility Metrics in RankScan #

The score combines several values. RankScan also shows them individually on the “AI Visibility” page.

The competitive field covers your brand, your tracked competitors and the other brands named in the answers. Brands you have removed from the comparison do not count. Depending on the setting, RankScan also leaves out associations, unrelated companies and plain terms. The article on AI reputation and share of voice explains more about share of voice and reputation.

RankScan measures Google AI Overviews through Google Search, not through a prompt. That is why they do not count towards the score.

Further Metrics from Practice #

GEO literature mentions further metrics. RankScan does not measure them as a separate value:

• Accuracy: Does the answer describe your brand correctly? You check this yourself in the answers to a prompt.
• ToneSentiment describes the tone in which a brand is portrayed in AI answers, for example positive, neutral, mixed, negative or uncertain. It measures the quality of a mention, not just its presence.Open glossary entry →: How does the answer judge your brand? RankScan detects negative mentions. They earn no points for an early mention.
• Coverage per prompt type: For which kinds of questions do you appear? You can read this from the prompt list.

A single answer is not a reliable measurement. Answers vary by assistant, prompt, time and language. A research paper on measuring AI visibility therefore recommends looking at visibility together with its uncertainty.
Source: arXiv: Quantifying Uncertainty in AI Visibility

The Score in the List of Measures #

The AI Visibility Score is not a measure. It has no status, and you do not report it as done. It is the goal in the header of the chapter “AI presence” in the list of measures:

AI presence
3 of 5 topics at 50 or above · Score 45

• “x of y topics at 50 or above” counts the active topics with a score that reach at least 50.
• “Score” is the AI Visibility Score of the whole property.
• “Review” names topics below 50 for which RankScan cannot derive a measure. Then you check the prompts, the keywords and the page of the topic yourself.

The measures of the chapter follow below the header. They are the steps that raise the score.

How to Read the Score #

Step 1: Find the Weakest Assistant #

Open the “AI Visibility” page. The table “Per AI assistant” shows how often each assistant names you, its score and its leader. The sentence on the score card names the assistant with the largest gap.

Step 2: Compare with the Leader #

The section “Comparison with your competitors” shows for each brand: named, named early, link to website and points. This tells you whether you lack mentions or only early mentions and links. If a brand you have not tracked is named often, RankScan suggests it as a competitor.

Step 3: Check the Prompts #

A score is only as good as its prompts. Check whether your prompts ask for providers. A pure knowledge question or a yes-no question rarely names providers. An answer without providers counts like an answer without your brand. RankScan gives hints on this when you add a prompt.

Step 4: Work Through the Measures in the Chapter “AI presence” #

The measures name concrete pages, questions and sources. They are listed in order of relevance:

Technical prerequisites are in the chapter “Technical”, for example “AI crawlersAI crawlers are automated bots that fetch web pages for AI providers in order to find content, make it available for search functions, retrieve it for answers, or, depending on the provider, use it for model training.Open glossary entry → blocked in robots.txt”. The article on AI crawlers and robots.txt explains more.

Step 5: Watch the History #

The history on the “AI Visibility” page shows two lines: “Named” and “Score”. Judge changes over several scans, not after a single answer. If your set of prompts changes, RankScan marks this in the history.

The Five Most Important GEO Levers #

These levers are general GEO expertise. RankScan's measures apply them concretely to your pages and topics.

1. Ensure Technical Accessibility #

GEO starts with access. If an AI assistant cannot retrieve your page, it cannot cite it either.

Check:

• Are important pages indexableIndexing is the step in which a search engine processes crawled content and decides whether a URL is added to the search index. Only indexed pages can appear in search results.Open glossary entry → and reachable with status 200?
• Does robots.txt block search crawlers or AI crawlers?
• Does a firewall or a content delivery network (CDN)A content delivery network (CDN) is a network of distributed servers that delivers static content closer to the user, reducing loading times.Open glossary entry → block bots by mistake?
• Is the content in the delivered HTMLHyperText Markup Language (HTML) is the markup language for web pages. It defines the structure and content of a page, such as headings, paragraphs, links and images.Open glossary entry →, or only after JavaScript?

Google names no additional requirements for AI Overviews and AI Mode. The usual basics of Google Search apply.
Source: Google Search Central: AI features and your website

You do not need an llms.txt for this. RankScan lists it as a hint with no action needed, because no AI system has been shown to read it.

2. Create Citable Content #

AI assistants are more likely to take over clear, verifiable statements.

Weak:

Our solution is innovative and helps companies become more successful.

Stronger:

RankScan measures visibility in AI answers in the assistants you choose from ChatGPT, Gemini, Claude, Perplexity and Copilot. It counts mentions, early mentions and links per answer.

The second statement is specific and verifiable. Put important facts such as prices, dates and figures on the page as text, not only in a PDF or an image. The article on high-quality content and fact density explains more.

In its tests, the GEO paper found that source citations, statistics and clearer wording could increase visibility in generated answers. The effects varied by topic. They are not a guarantee.
Source: arXiv: GEO: Generative Engine Optimization

3. Describe Your Brand Clearly #

AI assistants must understand who you are and what you stand for. Describe your brand in one clear sentence on the home page and the about page:

[Brand] is a [category] for [target group] focusing on [topic] in [market].

Use the same description in structured dataSchema markup is code that describes a page's content in a structured way, so search engines understand it not just as text but as clearly classified information.Open glossary entry → and on your profiles. Schema.org defines the type Organization for this.
Source: Schema.org: Organization

The article on entities and entity SEO explains more.

4. Improve Structure #

Clear subheadings, short sections, tables and summaries make extraction easier. Structured data makes information machine-readable.

Google describes structured data as a standardized format for providing information about a page.
Source: Google Search Central: Introduction to structured data

Structured data does not replace visible, helpful contentHelpful content refers to content created primarily for people that delivers a clear answer and shows substance, originality and trust, rather than being optimized mainly for search engines.Open glossary entry →.

5. Win Mentions on Other Websites #

AI answers often rely on other sources: comparison portals, trade media, associations, directories and forums. If these sources do not name your brand, the answer often does not name it either. RankScan shows you the sources that AI answers cite for your topics.

Quality matters more than quantity. Bought reviews, forum spam or invented awards are risky. Google describes such practices in its spam policies.
Source: Google Search Central: Spam Policies for Google Web Search

Choosing Good Prompts #

The score only measures what the prompts ask. A good prompt looks for a provider, the way a prospect would ask.

Keep two separate prompt sets for this reason. Brand-free provider questions measure visibility (discovery) and count towards the score. Branded questions measure reputation and belong in a separate evaluation. Record the region, language, the specific model, web access and the number of repetitions. A model queried through an APIAn application programming interface (API) is an interface through which software systems communicate and exchange data. It defines how requests are made and answered.Open glossary entry → can answer differently from the end-user product. Only treat such results as equal with this caveat.

RankScan gives hints when you add a prompt. For example, it warns when a prompt names no region, none of your keywords or no role. The hints never block: a prompt that deliberately measures a gap is legitimate.

Look at AI Assistants Separately #

Every assistant works differently. That is why RankScan calculates the score per assistant.

• ChatGPT answers with or without web search, depending on the mode. OpenAI's crawlerCrawling is the process by which a search engine bot fetches a URL and downloads its content. It is the first step before a page can be indexed.Open glossary entry → documentation matters for website owners.
• Perplexity usually shows its sources. You can see well there which websites shape the answers.
• Gemini and Claude cite more or fewer sources depending on the answer.
• Google AI Overviews are part of Google Search. RankScan counts them separately via the keywords, not in the score.

The statements about the individual assistants are observations, not documented rules of the providers. They can change.

Realistic Time Frames #

GEO has no fixed time to impact. It depends on many factors:

• how often the assistants retrieve your pages,
• whether an assistant answers with web search,
• how strong your competitors are,
• whether other sources name your brand,
• how large the change to your pages is.

You can check technical fixes right away. New mentions in AI answers usually take longer. For “Page for AI answers missing”, for example, RankScan says that the impact often takes several weeks. Do not expect fixed deadlines such as “first mention after four weeks”.

GEO Roadmap in Three Phases #

The order is a RankScan recommendation. How long each phase takes depends on your website and your team.

Phase 1: Measurement and Basics #

• Check the prompts of each topic: do they ask for providers?
• Track competitors, including suggestions from the comparison.
• Fix the technical measures in the chapter “Technical”, such as blocked AI crawlers or content that only appears after JavaScript.
• Choose a “Page for this topic” for every topic.

Phase 2: Pages and Content #

• Align the page of each topic with the questions of the prompts.
• Add clear statements, exact names and facts as text.
• Build comparison pages factually and as a table where RankScan suggests it.

Phase 3: Other Sources #

• Go through the sources of the AI answers.
• Check your entries on portals and directories.
• Contact neutral sources with a concrete occasion.
• Watch the history of “Named” and “Score”.

General Checklist: Checking AI Visibility #

This checklist is general expertise. It goes beyond what RankScan measures. Items that RankScan measures are marked.

• Are there fixed prompts that ask for providers? (RankScan: prompts per topic)
• Are competitors measured too? (RankScan: comparison)
• Are the assistants evaluated separately? (RankScan: Per AI assistant)
• Is the brand named, named early and linked? (RankScan: points)
• Which other sources do the answers cite? (RankScan: “Sources” tab)
• Does the answer describe the brand correctly?
• Is the tone of the answers appropriate?
• Are important pages accessible to AI crawlers?
• Is the content in the delivered HTML?
• Is there a clear description of the brand?
• Is there fact-rich, well-supported content?
• Do other websites name the brand?

Example: Timber Builder with a Weak AI Visibility Score #

The numbers are made up.

Starting Point #

A timber builder ranks well in Google for “timber construction Central Switzerland”. It rarely appears in AI answers. The AI Visibility Score is 28. The chapter “AI presence” shows “1 of 4 topics at 50 or above”.

Analysis #

• The table “Per AI assistant” shows: no Gemini answer names the timber builder.
• The comparison shows: the leader is often named early and frequently linked.
• Two prompts are knowledge questions and almost never name providers.
• The measure “Page for AI answers missing” suggests a comparison page for one topic.
• The measure “Relevant sources without your brand” lists an industry portal without an entry.

Approach #

1. Replace the two knowledge questions with prompts that ask for providers.
2. Build a factual comparison page with a table and report it with its URL.
3. Create the entry in the trade directory and report the measure.
4. Put references with project, year and location on the service pages as text.

Result #

The score does not rise right away. The prompt change shifts the value first, and RankScan marks it in the history. The impact of the new page only shows over several scans.

Common Mistakes in Generative Engine Optimization #

Mistake 1: Treating GEO as a Quick Trick #

GEO combines technology, content, brand description and other sources. A single trick does not replace that.

Mistake 2: Giving Up SEO #

GEO builds on SEO. Accessible pages, internal links and helpful content remain important.

Mistake 3: Overrating Single Answers #

One answer is not a measurement. Judge the score over several scans.

Mistake 4: Counting Prompts with the Brand Name #

A prompt with your name almost always names you. It shows awareness, not visibility. That is why it does not count towards the score.

Mistake 5: Marketing Language Instead of Facts #

AI assistants are more likely to take over concrete, supported statements than general promises.

Mistake 6: Ignoring Other Sources #

If only your own website explains your brand, confirmation from outside is missing.

Mistake 7: Blocking Crawlers Across the Board #

If you block all AI crawlers, you may lose mentions and links. Separate crawlers for training from crawlers for search.

Mistake 8: Overrating llms.txt #

An llms.txt is not a standard that AI systems have been shown to read. RankScan sees no need for action.

Conclusion: GEO Makes AI Visibility Measurable #

Generative Engine Optimization adds one question to SEO: do AI assistants name your brand, and do they link your website?

RankScan's AI Visibility Score answers this question with a number. It measures each assistant against the leader and the target. It shows where the largest gap is.

How to proceed:

1. Check that prompts ask for providers.
2. Find the weakest assistant and the leader.
3. Remove technical obstacles in the chapter “Technical”.
4. Work through the measures in the chapter “AI presence”.
5. Watch the history over several scans.

Sources and Further Reading #

• Google Search Central: AI features and your website
• Google Search Central: Introduction to structured data
• Google Search Central: Spam Policies for Google Web Search
• OpenAI Platform: Overview of OpenAI Crawlers
• Perplexity Docs: Perplexity Crawlers
• Schema.org: Organization
• arXiv: GEO: Generative Engine Optimization
• arXiv: Quantifying Uncertainty in AI Visibility