Generative Engine Optimization (GEO): method and measurement

GEO looks at how content is found, mentioned and cited in AI-generated search answers, and at how that can be observed. It does not replace classic SEO. GEO combines sound SEO, useful original content and clear supporting evidence with systematic tracking of mentions and citations in AI search. That is precisely why no provider should promise a citation or a ranking.

GEO begins with SEO fundamentals

Google's documentation for AI Overviews and AI Mode requires no special GEO file, no special schema and no extra technical hurdles. What matters are the familiar SEO fundamentals: a page must be accessible, indexed and eligible for a search snippet, and it must be understandable and useful to people. Google states explicitly that meeting these conditions guarantees neither crawling, nor indexing, nor serving in its AI features.

There is a practical consequence: before a team buys new “AI hacks”, it should check technical blockers, internal linking, canonicals, structured data that accurately reflects the page's visible content, sitemaps and editorial quality. Structured data can help machines understand a page, but it must not claim content that users cannot see on the page. An llms.txt experiment does not replace these fundamentals; Google documents neither an extra GEO file nor extra schema as a prerequisite for its generative search features.

Distinguish three different things cleanly

A classic ranking, a mention and a citation are not the same thing. A ranking describes a position in search results for a defined query, in a given location, at a given time. A mention only means that a brand or entity occurs in the generated text. A citation is a visible reference to a source. An answer can mention a brand without linking to its page; it can cite a page without naming the brand in the same sentence; and a classic ranking position can move without the same movement showing up in an AI answer.

Microsoft makes this distinction particularly clear in Bing Webmaster Tools: its AI Performance reporting shows citations, cited pages, trends and examples of grounding queries (the search queries Bing uses to retrieve source material). At the same time, Microsoft warns against reading citation figures as ranking, authority or placement. That kind of conceptual discipline should apply to every measurement system, even when another search interface is less explicit about it.

Check access rules for each platform

Crawler control is platform-specific. OpenAI documents OAI-SearchBot as the control over whether a website's content can appear in ChatGPT search answers, including possible summaries, snippets, citations and links. A website that excludes OAI-SearchBot does not appear in those search answers, but it can still turn up as a navigational link; the control covers search and is not a switch that rules out every conceivable mention across all ChatGPT uses. GPTBot governs possible use for training separately. The two settings are independent, and treating them as one can lead to an unintended decision. Even allowed access only means eligibility: it does not cause inclusion, selection or a citation.

For Google, its own rules for Search crawlers and indexing apply. For Bing, crawler access and current sitemaps remain important foundations. Sitemaps and truthful lastmod values make genuine updates recognisable; Microsoft names IndexNow as a way to report content changes promptly. Both help new or updated content to be discovered, but neither guarantees that a page will appear in generated answers. A technical register should therefore record, for each platform: the relevant bot, the robots rule, the indexing status, the sitemap, the canonical, the date of the last content update and the public visibility you expect.

Build content that extracts cleanly

A page does not get better by mechanically repeating search terms. A clear information architecture helps more. Answer the central question early, define ambiguous terms, use meaningful subheadings and keep facts, opinions and examples apart. Tables work for genuine comparisons, lists for steps, and paragraphs for explaining how ideas relate to one another. In our own implementation we serve everything decisive server-side instead of hiding it only in an interactive graphic – a robust choice on our part, not a universal Google requirement that would make every JavaScript page ineligible.

Evidence should sit directly next to the claim it supports. Name the source, date, geography and scope right where they support the statement. A Swiss legal question needs a Swiss primary source first. A product feature needs current documentation from the provider in question. A statistic needs a definition, a time period and a population. Avoid chains in which one blog cites another and the original evidence stays unclear.

Originality does not mean that every sentence has to be new. Value comes from original, verifiable work: a clearly documented method, a traceable decision aid, a precise local assessment or a carefully analysed data set. Summaries that add no further insight are easier to replace and help readers less. Google likewise emphasises useful, original content rather than pages mass-produced for search engines.

Do not inflate entities artificially

An entity is a clearly named unit such as an organisation, a product, a person or a subject area. Such units should be named consistently and connected clearly across your website. Contact details, author information and company details must be accurate. External profiles can support the same identity, but they should not be staged as a manufactured link scheme. Schema markup reflects what is visible on the page; it does not create expertise and does not by itself make a system recognise an entity.

The same applies to terminology. If “GEO” means Generative Engine Optimization in one part of the article and a region in another, that needs to be made clear. Synonyms such as AI search optimization and answer engine optimization can be explained without publishing three nearly identical pages. Every page should answer one clearly defined question for its target audience. Variants are explained on that page instead of creating cannibalisation through thin parallel pages.

Establish a baseline for visibility in AI answers before changes

A baseline is the first complete round of measurements, taken before you change anything; it does not start with a single screenshot. Define a small, stable set of information tasks: brand questions, category questions, problem-solving questions and high-stakes factual questions. For each check, record the interface, product, date, location or language setting, the exact prompt and the visible answer. Record separately whether the brand was mentioned, whether a page of yours was cited, which URL it was and whether the statement was factually accurate.

Generated answers can vary even with an identical prompt. A single run is therefore not a ranking position. Repeat at a rhythm you set in advance, not until a desired result appears. Changes to the prompt, the signed-in state, the model or the search interface are recorded as a new measurement condition. Where a provider's own tools, such as Bing AI Performance, supply data, record their definitions and coverage.

Visits can be attributed to an AI answer when reliable source information is available, such as a platform-provided label, referrer data or another valid tracking signal. If every trace is missing, the origin stays unknown and is not counted as zero. Such visits answer a different question than a citation: a citation may generate no visits. Conversely, a visit may come from an AI interface that your manual sample did not cover. Classic search performance, citations in AI answers, mentions, visits from AI answers and business outcomes therefore remain separate measurement series.

Prioritise changes with explicit hypotheses

After the first measurement, the answer is not to “optimise everything for GEO”. Formulate a concrete hypothesis: “The central definition only becomes visible after a long preamble; a direct answer plus a primary source could improve comprehension and retrievability.” A second hypothesis might concern an outdated sitemap or contradictory canonicals. Every change gets a clearly named person responsible, a date, the affected URLs and a note of what the team will watch in order to recognise a possible change.

Errors with broad impact come first: blocked important pages, incorrect canonicals, outdated core information, unsupported claims or content without a clear answer. Structural improvements follow. New pages are useful when they answer a distinct question your existing content does not cover. Mass-produced variants can harm quality, crawl budget and trust.

What a reliable report may and may not say

A reliable report may state that a page was technically accessible to a bot, that it appeared in an index, that it was mentioned or cited in certain documented tests, or that visits arrived through an identifiable referral from an AI answer. It should also state the sample, the interface, the date and the variability, and explain that the observations do not establish causation. What it must not do is turn a citation into general authority, a crawl into future selection or a short-term rise into a guaranteed business effect.

“Share of voice” also needs a precise definition. Which questions, systems, repetitions, languages and citation types count? Without that reference point, the percentage cannot be interpreted meaningfully. A small panel of prompts that your team defines can be useful for observing trends, but it remains a sample, not a complete picture of the market.

The decision to run a pilot

A GEO pilot is ready once the key question from your target audience, the existing pages, the technical access conditions, the evidence quality, the first measurement, the measurement definition and the change hypothesis are documented. Choose a few important pages and reversible changes. Check first whether the change informs people better; visibility in AI answers is an additional observation channel.

How maitiq helps: a citable answer needs an original contribution

Answer a concrete question from your audience directly, explain the conditions and add original evidence: a documented method, a useful comparison or a calculation readers can check. An article about lead costs, for example, can draw a clean distinction between cost per lead (CPL) and cost per qualified lead. Clearly named sources and a clearly identifiable publisher or named expertise make the statement verifiable.

maitiq connects such content with the actual offering and a suitable next step. Visibility in AI answers is observed, not guaranteed; indexable, helpful content remains the foundation. Alongside mentions, assess qualified visits and enquiries as far as these can be measured reliably.

How a first pilot with maitiq begins

In an initial assessment we clarify which questions your target audience asks, which of your pages already answer them soundly and where evidence is missing. For an agreed pilot project we record the pages involved, open questions about data and responsibility, planned changes and success criteria.

Decision rule: For each question that matters to your audience, define a short answer, a verifiable source and a clearly named person responsible. Do not promise visibility from special markup or guarantee a citation.

Sources and how to interpret them

The sources listed below are the primary documentation of the three platforms and describe capabilities and limits, not commitments to any individual website. Google explains which fundamentals apply to AI Overviews and AI Mode, that neither a special file nor special schema is required, and that meeting the conditions guarantees neither crawling, nor indexing, nor serving; traffic from these features appears in the overall Search Console report, not in a separate AI referral report. OpenAI separates the tasks of its crawlers: OAI-SearchBot for search in ChatGPT, GPTBot for possible use in training, and ChatGPT-User for visits that users trigger themselves. Microsoft presents AI Performance reporting in Bing Webmaster Tools, with citations, cited pages, trends and examples of grounding queries, and names IndexNow for reporting content changes; according to Microsoft, no ranking, authority or placement can be derived from those figures. Read such statements as a description of the relevant platform, not as evidence of a mention or a result.

Have maitiq review your specific case.