Generative engine optimization examples from cited pages

A page earns an AI citation when it gives an engine a clean piece of evidence for a particular claim. Length, schema, and brand polish can help a page get discovered, but none of them gives the engine a reason to attach that page to a sentence. The cited pages I keep seeing do something more basic: they answer a narrow question, make the answer easy to extract, and show where it came from.

That changes the job. You are not optimizing one page for one keyword and one ordered results page. An AI engine can break a question into several searches, retrieve different sources for each subtopic, and cite only the passages that support its final wording. Good generative engine optimization examples therefore look less like keyword tricks and more like careful technical documentation, original research, or a decisive field note.

I will dissect three real pages that surfaced as supporting sources while researching this article: Google Search Central's documentation on AI features, Microsoft's announcement of AI Performance in Bing Webmaster Tools, and the KDD 2024 paper that formalized GEO. They are different kinds of pages with different authority. Their shared patterns are useful, but the differences matter more than most GEO checklists admit.

Citation is a separate outcome from retrieval

A page can influence an answer without receiving a citation, and a citation can appear without sending a visit. Treating those events as one metric makes GEO reporting look better than it is. I separate the pipeline into eligibility, retrieval, use, citation, prominence, click, and conversion. Each stage can fail while the previous one succeeds.

The distinction has a practical consequence. If an engine never retrieves a page, rewriting its opening paragraph will not fix a blocked crawler, a weak internal path, or a mismatch between the query and the page. If the engine retrieves the page but cites another source, the problem may be evidence quality, passage clarity, source authority, or simple context competition. If the citation appears but nobody clicks, the answer may already satisfy the reader or the citation label may sit out of view.

The original GEO paper by Pranjal Aggarwal and colleagues is often summarized as proof that a few writing tactics raise AI visibility by up to 40 percent. Read the experimental setup before turning that number into a forecast. The researchers evaluated methods on GEO-bench with supplied source material and defined visibility measures inside generated responses. That is useful evidence that presentation can affect how a model uses available material. It does not prove that adding a quotation will make an unknown page crawlable, retrieved, cited across every engine, or commercially useful.

Microsoft now makes part of this distinction explicit. Its AI Performance report describes total citations and page-level citation activity, then warns that those counts do not indicate placement, importance, or ranking. That warning belongs in every internal GEO dashboard. Count citations as citations. Put referrals and qualified conversions beside them rather than quietly treating them as the same thing.

A workable measurement model uses at least four columns: citation rate for a fixed query set, share of cited answers where your page appears early enough to notice, visits carrying an identifiable AI referrer, and conversions from those visits. The first two diagnose source visibility. The latter two tell you whether that visibility matters to the business.

Three cited pages reveal three different reasons to cite

The strongest examples do not share a magic page template. They supply different kinds of evidence for different claims. During the research pass, each page below surfaced as a supporting source for the query it was built to answer. That is a small observed sample, not a universal ranking study, so I use it to inspect page construction rather than declare winners.

Google Search Central answers the eligibility question

Google's page titled "AI features and your website" answers a high-stakes operational question: what must a site owner do to appear in AI Overviews or AI Mode? The answer arrives near the top. Existing SEO practices still apply, there are no extra technical requirements, and a supporting page must be indexed and eligible to appear in Search with a snippet.

The page is unusually citable because each claim has a clear owner. Google explains its own product requirements. A third-party consultant can interpret those requirements, but cannot outrank the product owner on source authority merely by writing a smoother paragraph. The document also divides the topic by user task: how the features work, how to appear, how to measure, how to control content, and how to troubleshoot preview controls. An engine can retrieve one section without asking the model to untangle the whole page.

Notice what the page refuses to do. It does not invent an AI schema type, recommend a special machine-readable file, or promise inclusion. It explicitly says no special schema.org markup is needed and that eligibility does not guarantee serving. Those negative boundaries make it safer to cite because they prevent an answer from expanding a limited fact into a broad promise.

Microsoft supplies definitions tied to a product screen

Microsoft's "Introducing AI Performance in Bing Webmaster Tools Public Preview" page answers a different query: how can a publisher see citations in Microsoft AI experiences? It names the supported surfaces, defines each dashboard metric, and states what each metric cannot tell you.

This page has strong passage-level structure. "Grounding queries" is followed immediately by a definition. "Page-level citation activity" is followed by the scope and limitation of that count. The section on using the data connects observations to actions, including improving clarity, structure, evidence, and freshness. A model does not need to infer whether a paragraph describes a feature, a recommendation, or a caveat because the page labels the job of each paragraph.

It also dates the announcement and calls the feature a public preview. That context matters. Product documentation without version or status information ages badly, and an engine may combine an old definition with a current interface. A citable product page states the feature name, status, scope, metric definition, and limitation in the same local neighborhood.

The GEO paper supplies provenance and bounded findings

The paper "GEO: Generative Engine Optimization" answers the origin and evidence questions. Its arXiv record names the authors, submission and revision dates, conference acceptance, subject categories, abstract, and stable paper identifier. The page can support claims about who introduced the framework, what benchmark they built, and what the experiment reported.

Its citable unit is not simply a concise paragraph. It is a claim attached to research provenance. The abstract states that the effect varies by domain, which limits the headline result instead of pretending that one method works everywhere. Anyone who copies only the largest reported gain and drops the domain qualification makes their own page less trustworthy than the source.

These three pages share scannable headings and direct language, but that is the shallow lesson. The deeper pattern is claim-source fit. Google owns the eligibility rule. Microsoft owns the metric definitions. The research team owns the experimental result. The page most worth citing is the one entitled to make the exact claim, not the page with the most FAQ markup.

Answer-shaped passages reduce the model's work

A citable passage makes one defensible claim, supplies the necessary qualification, and stays understandable when lifted away from the page. This is more demanding than writing short sentences. A vague sentence remains vague when shortened.

Suppose a page says: "Our approach helps businesses succeed in AI search through advanced optimization." The sentence names no action, condition, evidence, or result. An engine cannot safely use it to answer whether a page needs special markup, how citations are measured, or what changed after a test. It is marketing copy disguised as an answer.

A stronger passage says: "Google requires a page to be indexed and eligible for a Search snippet before it can appear as a supporting link in AI Overviews or AI Mode. Google lists no additional technical requirement for those features." The first sentence states the prerequisite. The second closes off a common false inference. A reader can verify both against the named owner's documentation.

Write in blocks of roughly one to three short paragraphs around one question. Put the answer in the first sentence after the heading, then give the mechanism, boundary, and evidence. Keep pronouns local. If a paragraph opens with "this" but the noun it refers to appears two screens earlier, extraction can strip away the meaning. Repeat the precise noun when ambiguity costs more than repetition.

Tables work when the reader needs exact comparison across repeated fields, such as metric, definition, source, and limitation. They fail when writers pour whole essays into cells. Lists work for independent checks. Prose works for causality and exceptions. Choose the structure that preserves the relationship, because a model can extract a bad table as efficiently as a good one.

The heading should match the question at the level the passage answers it. A heading such as "AI visibility" is too broad to guide retrieval. "A citation count does not measure citation placement" gives both the engine and the reader a specific claim. Do not turn every heading into a question for style. Use the wording that makes the section's scope unmistakable.

Original evidence gives a page something to own

Pages that only summarize the current consensus compete with every other summary. Original evidence gives an engine a reason to cite your page rather than the source you paraphrased or a larger site saying the same thing. The evidence does not need to be a giant industry survey. It needs a method, an observation, and enough detail for someone else to inspect the result.

For a GEO page, useful first-party evidence could be a fixed set of prompts tested across engines, a before-and-after revision with captured citations, an anonymized table of grounding queries, or a failure trace showing that a canonical URL was not the page the engine selected. State the engine surface, locale, account state when relevant, date window, number of repetitions, and what counted as a citation. Without those fields, a screenshot proves that one interface produced one answer once.

Separate observation from interpretation. "The page appeared in 7 of 20 captured answers" is an observation if you retained the captures. "The engine prefers our format" is an interpretation and probably an overreach unless the test controlled topic, authority, retrieval, and competing sources. Most real publishing tests cannot control all of that. Say what changed, show what remained uncertain, and resist converting correlation into a rule.

Named sources still matter when you do not own the primary evidence. Engage with them instead of decorating the article with citations. Google's documentation rejects the need for special AI files or special schema to appear in its AI search features. Microsoft's documentation recommends clear headings, tables, FAQ sections, evidence, and current information for pages visible in its AI answers. Those positions can coexist: ordinary SEO eligibility gets the page into consideration, while clear evidence helps a retrieved passage support an answer.

A strong page often combines both layers. It uses an authoritative source for the external rule and its own reproducible artifact for the local result. The external source tells readers what the platform claims. Your evidence tells them what happened in the setup you actually run.

Source identity must match the claim

AI engines cite documents, but users judge sources. A technically perfect paragraph about a tax rule, medical dose, or platform requirement still has an authority problem if the publisher cannot reasonably know the answer. Clear authorship cannot manufacture expertise, yet hidden authorship can waste real expertise.

Put the author, organization, publication date, meaningful update date, and editorial ownership where a reader can find them. Explain the basis of experience through the work itself: configurations tested, edge cases encountered, artifacts produced, and decisions made. A long biography does less than one accurate failure analysis.

Keep entity facts consistent across the page. If a product has two names, a plan changed status, or a metric changed definition, state the relationship plainly. Do not force an engine to reconcile conflicting dates in the introduction, FAQ, schema, and footer. Structured data must match visible text. Google says this directly, and it is good publishing practice even when no AI feature is involved.

Claim-source fit also determines when you should not compete. A company can be the best source for its own API limits and a poor source for a neutral comparison of competitors. A consultant can publish a strong implementation test and still need the vendor's manual for the supported behavior. Cite the primary owner for facts it controls, then add your own experience where it changes the decision.

Third-party mentions have a different job. They can establish that other parties observed or evaluated an entity, but volume alone does not prove accuracy. A copied error across twenty roundup pages remains an error. For high-consequence claims, trace the statement back to the manual, dataset, standard, filing, or named person who can support it.

Technical eligibility comes before clever formatting

A page cannot be cited from the open web if the relevant engine cannot fetch, index, or expose it as a source. Teams regularly rewrite headings while a CDN blocks crawlers, JavaScript hides the main answer, a canonical points elsewhere, or a snippet directive suppresses the content. Fix that layer first.

Google's current guidance says a supporting link for AI Overviews or AI Mode must be indexed and eligible for a Search snippet. It also says site owners do not need a new AI file or special schema.org type. Microsoft says Bing respects robots.txt and other supported owner controls. Neither statement promises that an eligible page will be retrieved or cited.

Run a small header and HTML check from the same environment you use for release diagnostics. Replace the variables with a real page and a crawler user agent you are authorized to test. The expected output shape below is illustrative; your headers will differ.

PAGE_URL='your-page-address'
CRAWLER_UA='approved-crawler-user-agent'
curl -sSIL -A "$CRAWLER_UA" "$PAGE_URL" | sed -n '1p;/^content-type:/Ip;/^x-robots-tag:/Ip;/^location:/Ip'
curl -sSL -A "$CRAWLER_UA" "$PAGE_URL" | grep -ioE '<meta[^>]+(robots|canonical)[^>]*>' | head

HTTP/2 200
content-type: text/html; charset=utf-8
<meta name="robots" content="index,follow">
<link rel="canonical" href="your-page-address">

Check the final status after redirects, the rendered main text, canonical target, robots directives, snippet controls, authentication walls, and whether internal links lead to the page. Then inspect the engine's own webmaster tools. A successful curl request proves only that your test client received a response. It does not prove that a particular crawler saw the same response or that the engine indexed it.

JavaScript is not automatically disqualifying, but making the central answer depend on a fragile client-side request creates another failure point. Put important content in textual form and verify what the crawler receives. If the visible page says one thing and structured data says another, repair the disagreement rather than adding more markup.

A citation audit needs repeated queries and retained evidence

A useful audit measures a stable question set repeatedly. One prompt, one engine, and one screenshot cannot distinguish a durable pattern from response variation. AI answers can change with query phrasing, locale, time, account context, retrieved corpus, and model updates.

Use this sequence for a baseline and every meaningful page revision:

1. Build 20 to 40 questions from customer calls, search queries, support tickets, sales objections, and product documentation. Label each by intent and the claim your page should support. Freeze the set before testing.
2. Create two natural paraphrases for each question without inserting your brand. Run every version on each engine surface you care about, in a clean and documented context. Repeat on at least three separated occasions.
3. Save the full answer, all displayed sources, source order, engine, locale, timestamp, and prompt variant. Record whether each cited page actually supports the sentence attached to it.
4. Map each citation to the exact passage that likely supplied the claim. Mark failures as eligibility, retrieval, use without citation, citation mismatch, low prominence, or no business response. Do not use "GEO issue" as one catch-all label.
5. Change one class of page feature at a time, such as the opening answer, evidence block, title, or technical access. Rerun the frozen set and compare citation rate and support quality, not merely the presence of your domain somewhere in an answer.

A compact capture format keeps the work auditable:

run_id,engine,locale,query_id,variant,cited_url,citation_position,supports_claim,landing_sessions,conversions
2026-08-08-a,engine-a,en,q07,b,page-id,2,yes,0,0

Use an internal page ID if your reporting copy should not contain addresses. Keep the raw captures somewhere access-controlled and define every column. For citation rate, divide answers containing at least one qualifying citation to your page by eligible answers run for that query group. Keep errors and unavailable answers in a separate denominator report rather than silently dropping them.

Review false positives by hand. Domain mentions without a source link are not citations. A source included in a general carousel may not support the adjacent claim. A page cited for the wrong fact can raise the count while damaging trust. Citation correctness matters as much as citation frequency.

For teams I advise through oleg.is, I treat this as an engineering measurement loop with content judgment inside it, not a campaign that ends when a dashboard turns green. The retained query set and captures make disagreements testable.

Special GEO tricks are weaker than their sales pitch

The popular recommendation is to add an llms.txt file, special AI schema, an FAQ block, and a layer of short answers to every page. It is popular because the work is easy to package and easy to verify in a CMS. The evidence does not justify treating that bundle as an admission ticket.

Google explicitly says no new machine-readable AI file and no special schema.org markup is required for AI Overviews or AI Mode. That does not make every experimental convention useless across every product. It means you should not sell an unsupported mechanism as a Google requirement. Test it against a defined surface and keep the result separate from claims about other engines.

FAQ sections help when readers truly have discrete follow-up questions and the answers add information. A pile of near-duplicate questions can dilute the page, introduce contradictions, and create stale facts in ten places. The same rule applies to tables. Structure clarifies good material; it cannot rescue claims that have no source or page that answers no specific need.

Another bad recommendation is to rewrite every sentence into a tiny, declarative block for machines. That produces monotonous pages and often destroys the reasoning a human needs to judge a claim. Keep the direct answer near the top of a section, then preserve causality, exceptions, and tradeoffs. Extraction is useful only when the extracted passage remains true.

Do not optimize for an engine by impersonating certainty. Dates, method notes, limitations, and explicit unknowns make a page safer to reuse. If a result applies only to one locale or one interface, say so beside the result. A smaller claim that survives inspection is more citable than a sweeping claim that forces the engine to hedge or find another source.

The publishing checklist follows the citation pipeline

A GEO checklist should help a team locate failure, not encourage it to paste the same template onto every article. Run these checks in order because later polish cannot compensate for an earlier broken stage.

At eligibility, confirm that the final URL returns the intended page to approved crawlers and allows indexing and snippets. Retain the header capture, rendered text, canonical target, and robots review.

For query fit, confirm that the page owns a specific question and that the title and main heading describe that scope. Retain a query map with one primary intent and its related subquestions.

For the direct answer, make each major section state its answer in the first sentence and define necessary terms nearby. Review every candidate passage outside the page context.

For claim-source fit, take platform rules from the platform owner, research findings from the research, and local results from your retained test. Keep source notes beside every consequential claim.

For original evidence, include a method, artifact, observation, or failure that another writer cannot honestly copy. Retain raw captures, sample data, commands, or change history.

For boundaries, put dates, versions, locales, sample limits, and exceptions beside the claims they constrain. Add an editorial check for overstatement.

For consistency, make visible copy, structured data, FAQ, author details, and product facts agree. Retain an automated field comparison and a human review.

For citation quality, verify that a displayed source supports the sentence it accompanies rather than merely discussing the broad topic. Record a manual support judgment in the audit sheet.

For repetition, use frozen queries, natural paraphrases, several runs, and documented contexts. Keep a versioned query set and run log.

For the business result, keep citation, prominence, referral, and conversion metrics separate. Use a dashboard with distinct definitions and denominators.

Use the checklist to decide what to repair. A page with no retrieval needs technical and query-fit work. A retrieved page that loses the citation needs a better answer block, evidence, or source match. A cited page with no visits needs a stronger reason to continue beyond the generated answer, not another paragraph repeating the definition. A page that drives qualified conversations may already be doing its job even if its raw citation share looks modest.

The pages worth copying are not those with the most visible formatting. Copy Google's restraint when it states eligibility without promising inclusion. Copy Microsoft's habit of defining a metric beside its limitation. Copy the GEO paper's provenance and domain qualification. Then add the one thing none of those sources can supply for you: evidence from your own work, recorded well enough that a skeptical reader can check it.

Publish the page, preserve the baseline, and rerun the frozen questions after the engine has had a fair chance to recrawl. If the citation moves, you have a result to investigate. If it does not, the failure label tells you where to work next. That is slower than installing a GEO template, but it produces knowledge your competitors cannot copy by viewing your source.