Email Us WhatsApp Telegram Microsoft Teams

Blog

AI Search Visibility: Evidence, Mechanisms and Measurement of How Websites Are Cited in Generative Answers (2023–2026)

By Emily Thompson, Content Writer Published Updated 49 min read

1. Abstract and Keywords

Figure 1. The five-layer AI Visibility Stack proposed in this paper.

Between 2023 and 2026, a growing share of information-seeking moved from lists of links to generated answers. Google introduced AI Overviews and AI Mode, OpenAI added search to ChatGPT, and other assistants built retrieval into their responses. For site owners this created a new objective, which the industry calls AI search visibility, generative engine optimization (GEO) or answer engine optimization (AEO): to be mentioned, cited or recommended inside a generated answer. The field has grown faster than its evidence. Vendors publish large studies with incompatible methods, practitioners circulate tactics that no platform has confirmed, and the one peer-reviewed benchmark is based on a simulated engine and a fixed query set.

This paper reviews the evidence and builds a framework from it. Its method is documentary and comparative: it reads platform documentation, the peer-reviewed GEO study, a behavioral study by the Pew Research Center, vendor citation studies, and crawler-traffic data from Cloudflare, and it compares them on what each can and cannot show. The paper reaches five findings. First, the relation between organic rank and citation in AI answers is real but unstable: the share of Google AI Overview citations that come from the top ten organic results was reported at about 76 percent in mid-2025 and about 38 percent in early 2026 by the same vendor, and other vendors report lower figures, so rank is a weak and changing proxy. Second, the best-documented content tactics, adding sources, statistics and quotations, come from a single academic benchmark and should be treated as promising, not settled. Third, the technical conditions for eligibility are simple and documented, while many popular technical “optimizations”, such as the llms.txt file, have no confirmed effect on the major systems. Fourth, the commercial effect of AI answers on clicks is negative on average according to the best behavioral evidence, so visibility must be valued in terms beyond traffic. Fifth, measurement is the weakest part of the field, because generated answers vary from run to run and no platform publishes complete citation data.

The framework that follows from these findings, the AI Visibility Stack, has five layers: Access, Retrievability, Citability, Corroboration and Conversion. The paper presents an audit protocol, a ninety-day plan and a statement of what cannot be promised. In particular, no method guarantees inclusion in an AI answer, and claims of guaranteed placement should be treated as a warning sign.

Keywords: AI search visibility; generative engine optimization; answer engine optimization; AI Overviews; AI Mode; ChatGPT search; retrieval-augmented generation; query fan-out; citation; crawler access; llms.txt; zero-click search; measurement.


2. Introduction

2.1 From ranking to being used

Classic search presents a ranked list, and the unit of success is a position and a click. Generative search composes an answer from several sources and presents it, usually with links. The unit of success changes: a site can be used by the system without being visited by the user. Three outcomes therefore need to be distinguished.

  • Mention: the answer names the brand, product or site, with or without a link.
  • Citation: the answer links to or attributes a statement to a specific page.
  • Click: the user follows a link from the answer to the site.

A fourth outcome, recommendation, applies to answers that choose among options, such as “which tool should I use”. A brand can be mentioned without being recommended, and cited without being mentioned in the text. Because the economics differ, any serious discussion of visibility must say which outcome it means.

2.2 Why the topic needs a careful treatment

The subject attracts strong claims. Many of them cannot be tested, because the systems are proprietary, probabilistic and frequently changed. A claim such as “adding schema markup increases citations by a fixed percentage” can be repeated for years without anyone checking, and a claim such as “this file makes AI crawlers read your site” can be falsified by server logs only if someone looks. The aim of this paper is to separate four kinds of statement: what platforms document, what independent research shows, what vendors measure, and what practitioners assert. Each has a different weight, and the paper labels them.

2.3 Research questions

RQ1. How do the main generative search systems select and present sources, according to their own documentation?

RQ2. How strongly is citation in AI answers related to organic ranking, and is that relationship stable?

RQ3. What content and technical practices have empirical support for improving citation, and how strong is the evidence?

RQ4. What is the effect of AI answers on click-through and on the economics of publishing?

RQ5. How should AI visibility be measured, given non-deterministic answers and limited platform data?

2.4 Method and source classes

The paper uses documentary synthesis. Sources fall into five classes. Platform documentation includes Google Search Central pages on AI features and its May 2026 guide to generative AI features, and OpenAI’s publisher documentation on its crawlers and ChatGPT search. Peer-reviewed research is chiefly the GEO study of Aggarwal and colleagues, presented at the KDD 2024 conference. Independent behavioral research is the Pew Research Center’s analysis of browsing data from March 2025. Vendor studies are large citation-overlap studies from SEO software companies, which have large samples but commercial interests and methods that differ. Infrastructure data is Cloudflare’s published crawl-to-referral ratios and bot traffic shares. Trade press is used to locate and date these sources.

The paper does not generate original citation data. Its contribution is a critical synthesis, a framework, and a measurement protocol that others can test.

2.5 Limits

Four limits qualify everything below. Platform behavior changes frequently, so every figure is dated. Vendor studies measure different things under the same label. The peer-reviewed benchmark uses a research pipeline and one deployed engine, not the live systems of every provider. And the framework proposed is the author’s synthesis, which has not been tested in a controlled field experiment.

2.6 Structure

Section 3 describes how the systems work. Sections 4 and 5 review the evidence on ranking and on content tactics. Section 6 covers technical eligibility. Section 7 covers content, entities and corroboration. Section 8 addresses measurement. Section 9 discusses economics. Section 10 covers risks and ethics. Section 11 presents the framework and a plan. Section 12 concludes.


3. How Generative Search Systems Select Sources

3.1 The common architecture

Figure 2. Mention, citation and click: the outcomes of a generated answer, and what each is worth.

Most generative search products combine a language model with a retrieval step. The user’s question is interpreted, one or more searches are issued against an index or a search provider, the retrieved passages are supplied to the model, and the model composes an answer, usually attaching links to the passages it used. This pattern is often called retrieval-augmented generation. Its significance for site owners is that the model’s general training is not the only route to visibility. A page can influence an answer in two ways: through the model’s training data, where it is a faint and unattributed influence, and through retrieval at the time of the question, where it can be cited.

3.2 Google AI Overviews and AI Mode

Google’s documentation says that AI Overviews and AI Mode may use a technique it calls query fan-out, in which the system issues several related searches across subtopics and data sources to build a response. It says that the two features may use different models and techniques, so the links they show can differ, and that AI Overviews appear only when its systems judge them additive to classic Search. It states that there are no additional requirements for appearing in these features and no special optimizations beyond existing SEO fundamentals, and that a page must be indexed and eligible to be shown in Search with a snippet to be a supporting link. In May 2026 Google published a consolidated guide on generative AI features, reported in the trade press as the first such single resource; it repeats the focus on helpful, unique content and technical eligibility, and, according to the trade summaries, it does not provide a way to control participation in AI features separately from classic Search.

The documentation implies that Google’s generative features draw on the same index and ranking systems as ordinary search. The vendor studies reviewed in Section 4 suggest that the relationship between the ordinary results for the user’s exact query and the cited pages has loosened, which is consistent with fan-out: if the system issues sub-queries, the pages it cites may rank for those sub-queries and not for the original one.

3.3 ChatGPT search

OpenAI documents separate crawlers for separate purposes. OAI-SearchBot is described as the crawler for search, used to surface websites in ChatGPT’s search features; GPTBot is described as relating to training of generative models; and ChatGPT-User is described as an agent that acts on a user’s behalf. OpenAI’s help documentation tells publishers that to be included in summaries and snippets in ChatGPT, they should not block OAI-SearchBot, and that their host or content delivery network should allow traffic from OpenAI’s published IP ranges. It states that ChatGPT ranks results using multiple factors intended to find relevant and reliable information, and that placement is not guaranteed. Referrals from ChatGPT carry a utm_source=chatgpt.com parameter, which allows analytics tools to attribute the traffic. A change to robots.txt can take roughly a day to take effect, according to OpenAI. Blocking OAI-SearchBot is described as meaning that a site will not be shown in ChatGPT search answers, though it can still appear as a navigational link, a distinction OpenAI does not define in detail.

Two points deserve emphasis. The documentation is specific about eligibility and silent about selection: it does not describe how sources are chosen among eligible pages. And the separation between training and search crawlers means that a publisher can decline training use and remain eligible for search citations, although publisher practice and the legal environment are changing.

3.4 Other systems

Other assistants and search products, including Perplexity, Microsoft’s Copilot and Bing, and various browser-based assistants, follow the same broad architecture with differences in index, crawler and citation style. The Bing Webmaster Tools offering is reported, in secondary sources, to have introduced an AI performance report in February 2026; that report is a rare example of a provider giving publishers citation data, and its exact scope should be checked in Microsoft’s documentation. Because public documentation for several systems is thin, this paper concentrates on Google and OpenAI, where the documentation is richest, and treats claims about others with more caution.

3.5 What differs across systems

The systems differ on several dimensions that matter for strategy: whether they use their own index or a third-party provider; whether the crawler is separate from the training crawler; how many sources they cite; how they treat freshness; and whether publishers can see any citation data. A tactic validated on one system cannot be assumed to work on another, and measurement should be run per platform.


4. Evidence I: Does Organic Ranking Still Predict Citation?

4.1 Why this question matters

If ranking well in classic search were enough to be cited in AI answers, then the existing SEO discipline would carry over, and AI visibility would be a by-product. If the relationship is weak, a separate discipline is justified. The evidence is mixed and changing.

4.2 The vendor studies

Figure 3. Reported share of AI Overview citations that also rank in the top 10, by study (methods differ; not directly comparable).

Several vendors have published overlap studies. The most-cited is from Ahrefs. In mid-2025 it analyzed about 1.9 million citations drawn from about one million AI Overviews and reported that roughly 76 percent of cited pages ranked in the top ten for the same query, with about 86 percent somewhere in the top one hundred, and a median organic rank of 2 for the first cited URL. In March 2026 the same vendor repeated the analysis on a larger sample of about 863,000 keyword results and 4 million AI Overview URLs and reported that only about 38 percent of cited pages ranked in the top ten, with the remainder split almost evenly between positions 11 to 100 and pages outside the top 100. The vendor noted that Gemini 3 had become the default model for AI Overviews in January 2026 and suggested that fan-out queries were playing a larger role in source selection. It also found that among cited pages outside the top 100, a large share, about 18 percent, were YouTube videos, and that YouTube accounted for about 6 percent of all citations in its dataset.

Other studies report lower overlap, using different definitions. A BrightEdge analysis reported by a trade source found that only about 17 percent of AI Overview citations came from pages in the organic top ten, even as the overall overlap with organic results grew over time. A Surfer study reported that about 68 percent of cited pages did not rank in the top ten for the query, counting a page as ranking if it appeared for the query or any of its fan-out queries. A small independent study of 198 AI Overviews found that about 31 percent of cited pages were on the first page of the same search, with “best” queries following the organic results more closely than “how to” queries.

4.3 Why the numbers differ

The figures cannot be placed on one scale, for at least five reasons.

  1. Different definitions of “ranking”. Some studies count only standard organic listings; others count any block in the results, including featured snippets, “People also ask” and video packs.
  2. Different query samples. Sample composition by topic, query length, intent and country changes the overlap.
  3. Different time points. AI Overviews are updated frequently, and the underlying model changed in January 2026.
  4. Different tracking methods. Each vendor parses the results page with its own tools, and these capture different numbers of citations.
  5. Different handling of fan-out. Counting a citation as matched if it ranks for a fan-out query raises the overlap relative to counting only the original query.

The honest summary is that the studies agree on direction, not on magnitude. Organic rank for the exact query predicted citation reasonably well in 2025, predicts it less well in 2026, and in all studies a substantial minority of citations come from pages that do not rank for the exact query.

4.4 Implications

Three implications follow. First, ranking well remains useful because it still correlates with citation, and because the same quality systems seem to underlie both. Second, ranking for the head query is no longer sufficient, and pages that cover related sub-questions can be cited without ranking for the main query. Third, formats other than text pages matter: the repeated finding about YouTube suggests that video, with titles, descriptions and transcripts that state clear answers, is an independent route to citation. The vendor also reports that mentions of a brand on YouTube correlate strongly with AI Overview visibility. This is a correlation reported by a vendor, and it does not show that producing videos causes citation.

4.5 A note on causality

None of the overlap studies can show that a given page was cited because of some property. They show that cited pages differ on average from other pages. The distinction matters because many recommendations in the industry take the form “pages that were cited had feature X, therefore add X”. Without controlled comparison, such recommendations may confuse the feature with the underlying quality that produced both the feature and the citation.


5. Evidence II: What Content Tactics Have Support?

5.1 The GEO benchmark

The central academic source is the paper “GEO: Generative Engine Optimization” by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, first circulated in November 2023 and accepted at the ACM KDD 2024 conference. The authors define a generative engine as a system that retrieves sources and produces a cited response, propose a framework for content creators to improve visibility in such responses, and introduce GEO-bench, a benchmark of about ten thousand queries across several domains. They test a set of content modifications against a baseline and report that the best methods improved visibility by up to about forty percent. The most effective were adding citations to sources, adding quotations from relevant sources and adding statistics. The authors also report that keyword stuffing offered little or no improvement, and they find that effectiveness varies by domain. They also validated a subset of methods on Perplexity, a deployed engine, with improvements of up to about thirty-seven percent on one of their measures. A secondary report of the study states that lower-ranked sites benefited more than top-ranked ones, which, if accurate, suggests that such tactics could help sites with weaker positions; this detail should be checked in the original paper.

5.2 How to read the benchmark

The study is valuable and should be read with care.

  • The engine is partly simulated. The main experiments use a research pipeline of a language model over retrieved sources, which resembles but is not identical to any commercial product. The Perplexity test provides external validation for a subset of methods.
  • The visibility metrics are constructed. They measure how much of the response text is attributed to a source and how it is positioned, which approximates but is not the same as being cited and clicked.
  • The modifications are applied to existing text. The tactics were tested by editing content and observing the effect on its inclusion. They do not test what happens when a site creates new content with these features from the start.
  • The time frame is early. The experiments reflect the systems of 2023 and 2024, and the systems have changed.
  • Gaming risk. If a tactic works by making content look more authoritative to a model (for example by inserting quotations), it may be copied at scale and then countered, as happened with earlier SEO tactics.

The reasonable conclusion is that the study supports a modest, plausible claim: content that is well-sourced, specific and quotable is more likely to be used in generated answers than content that is vague. That is consistent with Google’s guidance on helpful, unique content.

5.3 Vendor studies of correlates

Vendors have published correlational studies of brand-level visibility. Examples reported in the trade press include claims that brand mentions across the web predict AI citation better than backlinks, that content cited in AI answers tends to be fresher than content that merely ranks, that mentions on YouTube correlate with AI Overview visibility, and that distributing content through reputable third-party publications increases citations relative to publishing on one’s own site alone. A reported academic paper from September 2025 is described as finding that AI search systems show a strong bias toward earned media, meaning third-party sources, over brand-owned content. These findings share a theme, corroboration by independent sources, which fits the way retrieval systems favor material that appears in many places. But they are either vendor studies or secondary reports, and their methods have not been reviewed independently. The paper treats them as hypotheses to be tested, with the corroboration hypothesis the most plausible.

5.4 Practices with documented support and those without

The table summarizes the strength of evidence for common practices. The categories are the author’s assessment based on the sources above.

Table 1. Strength of public evidence for common AI-visibility practices (author’s assessment).

Practice Evidence class Strength
Allow the relevant crawler; keep pages indexable and snippet-eligible Platform documentation Strong (eligibility, not selection)
Provide helpful, original, accurate content Platform documentation Strong as a principle; no mechanism
Add sources, statistics, quotations Peer-reviewed benchmark, one engine validation Moderate
Cover related sub-questions (fan-out) Vendor studies, platform description of fan-out Moderate
Earn independent mentions and citations Vendor correlations, secondary report Moderate to weak
Publish video with clear titles and transcripts Vendor correlation Weak to moderate
Add structured data for AI inclusion Google states not required; vendor claims Weak (no public evidence of AI effect)
Create an llms.txt file Google states not used; logs and correlation studies Very weak
Keyword stuffing GEO benchmark Negative or none

5.5 Summary

The strongest content conclusion is not a trick. It is that content that gives a system something precise, attributed and verifiable to use is more likely to be used. The paper’s framework builds on this.


6. Technical Eligibility: Access, Retrievability and the Files That Do Not Matter

6.1 The eligibility principle

Figure 4. Crawler purposes and what blocking them does (based on platform documentation; verify current names and rules).

A page cannot be cited by a retrieval system that cannot fetch it, index it and show a snippet from it. The first layer of AI visibility is therefore access, and it is the layer where most of the clear, documented guidance exists. The technical requirements are modest but frequently violated by accident.

For Google, the documentation states that a page must be indexed and eligible to appear in Search with a snippet to be a supporting link in AI Overviews or AI Mode. The practical checks are the ordinary ones: the page is not blocked by robots.txt or a noindex directive; it returns a successful response; the main content is available to the renderer; the canonical URL is the one intended; and snippet-restricting directives have not been applied without intent.

For OpenAI, the documented requirements are to allow OAI-SearchBot in robots.txt and to ensure that the host or content delivery network does not block OpenAI’s published IP ranges. Several practitioner analyses report that accidental blocking at the CDN or firewall level is common, since bot-management rules can block unfamiliar crawlers by default.

6.2 Training crawlers versus search crawlers

The distinction between crawlers that collect training data and crawlers that retrieve pages for answers is central to a publisher’s decisions. Cloudflare reported that, in May 2026, about 52 percent of AI crawler requests on its network were for training and about 9 percent for search, with training crawlers returning little or no referral traffic. In response, many publishers now block training bots while allowing search bots, according to analyses based on Cloudflare data, a pattern that corresponds to the separation in OpenAI’s documentation between GPTBot and OAI-SearchBot.

Google’s situation differs. A trade report quoting Cloudflare observes that most AI companies separate discovery crawlers from training crawlers, which allows publishers to choose, but that Google does not offer the same separation for its main crawler. Google provides a separate token, Google-Extended, for controlling use of content in some of its AI products, but the documentation for AI features states that eligibility for them depends on being indexed and snippet-eligible in Search. Publishers should consult Google’s current documentation for the precise effect of each control, since the options have evolved.

The strategic consequence is a trade-off that has no universal answer. Blocking a training crawler can protect content from use in model training with no loss of search visibility on systems that separate the two. Blocking a search crawler forfeits citation on that platform. An analysis reported by an industry source asserts that sites blocking AI bots receive effectively no referrals from ChatGPT, which is consistent with OpenAI’s description of the search crawler but is a secondary report.

6.3 Rendering and content availability

Retrieval systems vary in whether they execute JavaScript. Google renders pages. Other crawlers may fetch raw HTML and extract text without running scripts. A page whose main content is injected by client-side scripts may therefore be visible to Google and invisible to a simpler crawler. The safe practice is to deliver the primary content in the initial HTML, using server-side rendering or static generation, and to reserve scripts for enhancement. The same practice benefits accessibility and speed.

6.4 Structured data and AI inclusion

Google’s documentation states that no special markup is needed for AI features. Structured data remains useful for rich results eligibility and for clarifying content to machines, as treated in the companion paper on on-page SEO, but no public evidence from the platforms shows that it raises the probability of citation. Several trade analyses make the same point, describing structured data as a lever for rich results and not for AI citation. The recommendation is to implement markup that is accurate and justified on its own merits, and not to expect AI visibility from it.

6.5 The llms.txt file

The llms.txt proposal, launched in September 2024, suggests that sites publish a markdown file at a standard location to help language models find and understand key content. It has been adopted by a large number of sites, with one tracking service reporting more than eight hundred thousand implementations, concentrated in developer-tool and software companies. The evidence on its effect is consistently negative for the systems that matter to most publishers.

  • Google’s guide to generative features includes a section dispelling myths, and the documentation states that Search does not use machine-readable AI files, and that keeping one neither helps nor hurts. A Google search advocate stated publicly that no AI system currently used the file, and a note was added to Google’s documentation in June 2026, according to trade reports.
  • Server-log analyses by several independent parties report that the major AI crawlers rarely or never request the file. One vendor reported that OpenAI fetched it seven times across eighty-three sites in twelve weeks.
  • A correlation study of about 300,000 domains by SE Ranking is reported to have found no relationship between the presence of the file and citation rates.
  • The one legitimate use found is documentation for developers, where coding assistants and agents may read such files.

The conclusion is that llms.txt is, for most sites, a low-cost but low-value activity, and that its promotion is an example of enthusiasm outrunning evidence. The cost of maintaining a manually curated file can exceed its benefit, and a site with limited resources is better served by access control, content quality and corroboration. The file does no harm, and a docs-heavy site may still choose to publish one for agent use.

6.6 Performance and reliability

Retrieval systems that fetch pages in real time have time limits. A slow or unreliable server may fail to deliver a page within the system’s window, and the page is then skipped. The ordinary measures, fast server response, caching, a stable hosting configuration and correct status codes, serve AI retrieval as well as classic search. Because AI crawlers can produce heavy load, as Cloudflare and ETH Zurich researchers have discussed in relation to cache behavior, site owners should also monitor bot traffic and apply rate limits that do not block legitimate search crawlers.

6.7 A technical checklist

  • Confirm that the primary content is present in the initial HTML.
  • Check robots.txt rules for each relevant crawler by name, and check CDN and firewall rules.
  • Verify that pages intended for citation are indexable, canonical and not restricted from snippets.
  • Review server logs for crawler requests and response codes.
  • Keep sitemaps accurate and update dates truthful.
  • Do not rely on files or markup that platforms have said they do not use.

7. Content, Entities and Corroboration

7.1 Citability: what a system can use

Figure 5. Features of a passage that a retrieval system can use and a reader can verify.

Section 5 established that content with specific, attributed and verifiable information is more likely to be used. This section turns the finding into editorial practice. A page is citable when a passage can be extracted and used without loss of meaning, and when the source of the information is clear.

Six properties help.

  1. Self-contained passages. A passage should answer a sub-question without depending on earlier text. Use the full name of the subject, state the condition under which the answer holds, and avoid pronouns that point to distant context.
  2. Precise statements. A concrete statement with a quantity, a date, a condition or a named entity is more useful to a system than a general one. The benchmark’s finding about statistics is consistent with this.
  3. Attribution. State where the information comes from. Cite primary sources and, where original, say so. This helps both the system and the human who follows the citation.
  4. Definitions. Define terms at first use. Definitions are among the most frequently extracted passage types.
  5. Structure that matches content. Use tables for comparisons, lists for sequences and headings that state the question the section answers.
  6. Dates and versions. State when a fact was true and keep pages current. Freshness appears in vendor reports as a distinguishing feature of cited content, and it is plausible that systems prefer current information for time-sensitive queries.

None of this requires writing for machines. It is a description of good reference writing.

7.2 Original information as the strongest asset

Information that exists nowhere else cannot be replaced by a competitor’s page, and cannot be reproduced by a model from general training. Original data, first-hand tests, proprietary surveys, case records with permission, and expert interviews all qualify. Google’s guidance on generative features, as summarized in trade sources, emphasizes valuable, unique, non-commodity content. For a retrieval system, original information has a further advantage: if the system needs the specific fact, only the original source can supply it, so citation is more likely to point there.

Commodity content, such as a standard definition or a summary of the first five search results, faces the opposite problem. A system can generate it from many sources, so no individual source is necessary, and a page that offers nothing else is the most exposed to being synthesized away.

7.3 Coverage of sub-questions

Because fan-out breaks a complex question into sub-queries, a site that covers a topic thoroughly, with distinct pages or well-delineated sections for each sub-question, has more chances to match. This favors topical depth over breadth of thin pages. A practical method is to take a target question, list the sub-questions an expert would consider, test which of them the site answers with a self-contained passage, and fill the gaps. The tools that simulate fan-out, which several vendors provide, give approximations only, since fan-out queries differ between runs and are not exposed by the platforms.

7.4 Entities and brand recognition

Generative systems answer questions about brands, products and people, and the way a brand is described depends on how it is represented across sources. Four practices support accurate representation.

  • Consistent description. Use the same name, category description and key facts across the site, profiles and directories.
  • An authoritative home for facts. Provide an about page, product pages and documentation that state the essential facts clearly.
  • Presence in reference sources. Where the organization qualifies under the rules of reference works, ensure that accurate information exists there. The Pew study found that Wikipedia links were somewhat more common in AI summaries than in standard results, which indicates that such sources carry weight.
  • Correction of errors. Monitor how systems describe the brand, and correct errors at the source where possible, since the system’s description is derived from the sources it finds.

These practices overlap with the on-page and off-page work discussed in the companion papers on those subjects.

7.5 Corroboration

The hypothesis that independent mention increases the likelihood of citation fits the mechanics of retrieval and of the language model. A claim that appears consistently on several credible sites is more probable to be retrieved and more probable to be treated as reliable. Corroboration is also the on-web counterpart of the “authoritativeness” in E-E-A-T. The practical forms are earned coverage in credible publications, expert contributions, genuine reviews, citations of original research by others, and participation in communities where the audience asks questions. The vendor and secondary-source evidence summarized in Section 5.3 points in this direction but is not conclusive.

Two warnings apply. First, corroboration cannot be manufactured without risk. Paid link schemes, fabricated reviews, undisclosed sponsored content and networks of low-quality sites are addressed by Google’s spam policies and by platform rules, and they create reputational and legal risks. Second, the more a system is trained to detect artificial corroboration, the less value such schemes have. The durable method is to do things that deserve mention.

7.6 Video and other formats

The repeated finding that YouTube is a large source of citations, including for queries where the video does not rank organically, suggests that AI visibility is not confined to text pages. For topics where demonstration matters, a video with an accurate title, a clear description and a transcript that states the main answers is a plausible route. The caution is the same as before: the evidence is correlational, and it is more relevant to some query types, such as how-to queries, than to others.

7.7 Freshness and maintenance

Pages cited in AI answers, according to vendor reports, tend to be newer than pages that merely rank. The mechanism is likely that retrieval systems favor current information for queries that need it. Maintenance is therefore a visibility activity: review pages for accuracy on a schedule, update facts and dates when they change, and remove or redirect obsolete pages. Changing a date without changing content is deceptive and counterproductive.


8. Measurement: How to Know Whether Visibility Exists

8.1 Why measurement is hard

Three properties of generative systems make measurement difficult. First, answers vary. The same question asked twice can produce different text and different citations, and answers differ by user, location, time and conversation context. Second, platforms disclose little. Google reports AI feature traffic within overall Search Console data, and publishers should check Search Console for any dedicated reporting of AI features, since its scope has changed over time; OpenAI publishes no impression or citation data to publishers; and Microsoft is reported to have introduced a performance report for AI in Bing Webmaster Tools in early 2026. Third, the outcome of interest, influence without a click, leaves no trace in the publisher’s analytics.

8.2 Metrics

Figure 6. A measurement loop for AI search visibility.

A defensible measurement program combines several metrics, each with a clear definition.

  • Citation rate: the share of sampled answers, for a defined set of prompts, in which a page from the site is cited.
  • Mention rate: the share of sampled answers in which the brand is named, with or without a link.
  • Share of voice: the site’s citations or mentions as a proportion of all citations or mentions among a defined set of competitors for the same prompts.
  • Position or prominence: whether the citation appears early in the answer or in a prominent link, if the interface allows this to be judged.
  • Accuracy: whether the answer describes the brand and its facts correctly.
  • Sentiment or framing: whether the answer recommends, neutrally describes or criticizes.
  • Referral traffic: visits attributed to AI sources, using referrer data and parameters such as utm_source=chatgpt.com that OpenAI adds to ChatGPT referrals, and the AI features’ share of traffic in Search Console where reported.
  • Downstream outcomes: conversions, signups and sales from AI-referred sessions, which Google has encouraged site owners to treat as more important than raw clicks.

8.3 Designing a prompt set

The quality of the measurement depends on the prompts. A good prompt set reflects the real questions of the target audience, drawn from search queries, support tickets, sales conversations and community discussions. It covers the funnel: awareness questions (“what is”), consideration questions (“best”, “compare”) and decision questions (“how much”, “how to buy”). It includes brand and non-brand prompts, since the latter reveal discoverability. It is stable over time, so trends can be compared, and it is reviewed periodically for relevance. A set of fifty to a few hundred prompts, segmented by topic, is typical for a mid-sized site, but the right size depends on the variability of the answers and the precision required.

8.4 Sampling and statistics

Because answers vary, a single run is not a measurement. Each prompt should be run multiple times, on a schedule, across the platforms of interest, and the results aggregated. The variability can be estimated from repeated runs, and a confidence interval can be attached to the citation rate. Changes smaller than the interval should not be interpreted. Location, language and logged-in state should be controlled or recorded, since they affect results. When comparing before and after a change, the comparison should include similar pages or prompts that were not changed, to reduce the risk of attributing a platform-wide shift to the intervention. Vendor tools can automate this, but their sampling methods and prompt databases should be understood before their figures are trusted.

8.5 Attribution of traffic

AI-referred traffic is small in many analytics accounts and is often misattributed. Some assistants strip referrer information, some apps open links in in-app browsers, and some visits appear as direct traffic. Using the identifiable signals, such as the ChatGPT parameter, gives a lower bound. Surveys that ask new customers how they found the business, and branded-search trends after AI exposure, give complementary evidence of influence that analytics cannot see.

8.6 Reading the vendor dashboards

Commercial AI visibility tools publish scores, such as “AI visibility index” or “share of AI voice”. These are useful for trends and for competitive comparison, but they are proprietary composites. Before using one, the analyst should ask: what prompts does it run, how often, on which platforms and in what location; how does it count a mention; how does it handle variability; and can the raw responses be inspected? A score that cannot be traced to raw responses should be treated as an indicator and not as a measurement.

8.7 A minimal audit protocol

  1. Define the objective (citation, mention, recommendation or conversion) and the audience.
  2. Build the prompt set and tag by funnel stage and topic.
  3. Select platforms and record settings (location, language, logged-in state).
  4. Run repeated samples over at least a few weeks.
  5. Code the responses for citation, mention, accuracy and framing, ideally with two coders for a subset to check agreement.
  6. Analyze rates with intervals, compare with competitors, and identify gaps by topic and platform.
  7. Diagnose each gap using the layers of the stack in Section 11.
  8. Intervene and re-measure, with control prompts where possible.

9. Economics: Clicks, Value and the Publishing Bargain

9.1 What the best behavioral evidence shows

Figure 7. Share of Google visits with a click, by presence of an AI summary (Pew Research Center, March 2025 browsing data).

The most cited independent evidence on the commercial effect comes from the Pew Research Center, which analyzed browsing data from about 900 U.S. adults who agreed to share it, covering 68,879 Google searches in March 2025. Pew reported that about 18 percent of the searches produced an AI summary. When an AI summary appeared, users clicked a traditional result link in 8 percent of visits, compared with 15 percent when no AI summary appeared. Clicks on the links inside the AI summaries were rarer still, at about 1 percent of visits. Users were also more likely to end their browsing session after a page with an AI summary. Pew noted that the study did not address revenue effects or effects outside U.S. English-language search.

Google’s position is different. Its documentation states that clicks from results pages with AI Overviews are of higher quality, in the sense that users are more likely to spend more time on the site, and it says that AI Overviews lead people to visit a greater diversity of websites for complex questions. These statements are about the quality and diversity of clicks and are not directly inconsistent with a lower number of clicks. The two sources therefore describe different things, and neither establishes the net effect on a particular publisher’s revenue. The paper’s reading is that the average effect on clicks to informational queries is negative, that the effect varies by query type and by site, and that site-level data are necessary for any decision.

9.2 The crawl-to-referral imbalance

Cloudflare publishes a crawl-to-referral ratio, the number of pages a bot operator crawls for each visit it refers. In 2025 figures cited by Cloudflare and trade sources, Google’s ratio was about 14 to 1, OpenAI’s about 1,700 to 1 and Anthropic’s about 73,000 to 1, and later 2026 analyses report ratios that fluctuate widely and, for some operators, fell substantially as consumer products that link out were launched. Trade analyses of 2026 data report Google’s traditional-search ratio at roughly 5 to 1 in early 2026, and ClaudeBot’s ratio in the thousands to tens of thousands in different periods. The figures depend on the period and the method, and they are quoted here to illustrate the order of magnitude and the direction of change, not as precise values. The disclosure in this paper that one of these operators is the developer of the model writing this text is intended for transparency; the data are Cloudflare’s, and the same caution applies to all operators.

The imbalance explains why many publishers block training crawlers while allowing search crawlers, and why Cloudflare moved to blocking certain AI crawlers by default for new domains from September 2026 and, according to trade reporting, shifted from per-crawl charges to compensation tied to citations from July 2026. These are infrastructure-level responses to the broken bargain, and their effects remain to be seen.

9.3 What visibility is worth

If clicks fall on average, the value of AI visibility must be measured in other terms. Four sources of value are plausible.

  1. Brand influence. Being named in the answer shapes consideration, even without a click. This is the least measurable and, for brands with a sales cycle, potentially the most valuable.
  2. Higher-intent referrals. Users who do click from an AI answer may be further along in their decision, which is consistent with Google’s claim of higher-quality clicks, although the evidence is contested.
  3. Defense of existing traffic. A vendor study reported by a trade source found that brands cited in AI Overviews had higher organic click-through than brands not cited. The causal direction is unknown, and the figure is a vendor claim.
  4. Accuracy. Being the source of the facts in the answer reduces the risk of misrepresentation of the brand.

For each business, the relative weight differs. A publisher whose revenue depends on page views is more exposed than a service business whose revenue depends on inquiries.

9.4 Which pages are exposed

The pages most exposed to loss of clicks are those that answer simple informational questions that a system can answer from a snippet: definitions, short facts, conversions, basic how-tos. The pages least exposed are those that offer what an answer cannot: tools, calculators, original data, community, transactions, in-depth analysis and the personal voice of a trusted author. Strategy follows from this. For each page, ask what a reader who has seen only the summary still needs, and make sure the page provides it.

9.5 Strategic options for publishers

Publishers have several options, which can be combined.

  • Compete for citation on the queries where visibility is valuable, using the practices of this paper.
  • Differentiate toward content that is hard to summarize and valuable to visit.
  • Restrict training crawlers while allowing search crawlers, where the trade-off suits the business.
  • Build direct relationships through newsletters, memberships, apps and communities, which reduce dependence on any intermediary.
  • Seek licensing or partnership terms where the scale justifies them, noting that the terms of such arrangements are largely private.
  • Measure and report to inform decisions, instead of reacting to anecdotes.

None of these options is a guaranteed remedy. The common thread is to treat AI search as one channel with its own economics, to be evaluated like any other.


10. Risks, Ethics and the Limits of Optimization

10.1 Manipulation of generative systems

Figure 8. A spectrum from legitimate practice to manipulation in AI visibility work.

Wherever a system selects sources, attempts to manipulate the selection follow. Several techniques have been described in research and in the trade press: hidden text addressed to the model, instructions embedded in pages that try to steer an assistant’s recommendation, mass-produced pages that mirror the phrasing of common prompts, fabricated reviews and fake “best of” lists that rank the publisher first. Their effect varies and they decay as systems adapt, but the more important point is that they carry costs for the publisher.

  • Policy risk. Google’s spam policies cover scaled content abuse, site reputation abuse and deceptive practices, and these apply to material intended to influence AI features as to classic results.
  • Reputational risk. A brand exposed as having planted fabricated comparisons or reviews loses trust that is hard to recover.
  • Legal risk. Fake reviews and undisclosed endorsements are regulated in many jurisdictions.
  • Obsolescence risk. A tactic that exploits a current weakness of a model is lost when the weakness is fixed.

The line this paper draws is simple. Practices that make real information easier to find, verify and use are legitimate. Practices that create false signals, or that address the system instead of the reader, are not. Figure 8 shows the spectrum, with the middle area, such as aggressive formatting for extraction, requiring judgment about whether the reader is also served.

10.2 Accuracy and the duty of care

Generated answers can be wrong, and a page that is cited may be misquoted or taken out of context. A publisher cannot control this, but it can reduce the risk by making statements unambiguous, stating conditions and exceptions next to the claim, and avoiding sentences that are accurate only in context. For regulated topics such as health, finance and law, the duty is higher. Content in those areas should be reviewed by qualified people, cite primary sources, and state limits clearly, and the publisher should monitor how assistants summarize it.

10.3 Misrepresentation of the brand

Systems sometimes state wrong facts about companies: incorrect prices, discontinued products, invented features or confused identities. The measurement protocol in Section 8 should include an accuracy check, and errors should be traced to the sources from which the system could have drawn them, such as outdated pages, directory listings or third-party articles. Correcting the source is the reliable remedy. Platforms provide feedback channels, but their response is not guaranteed.

10.4 Content licensing, consent and the training question

Whether a publisher should permit its content to be used for model training is a business and ethical decision, not a technical one. The considerations include the lack of referral traffic from training crawls, the possibility that blocking reduces the brand’s presence in a model’s general knowledge, the existence of licensing arrangements for larger publishers, and the evolving law. This paper does not recommend a single policy. It recommends that the decision be made deliberately, per crawler, and documented, and that it be revisited as platforms change their controls.

10.5 Fairness and concentration

If a few systems mediate a large share of information access, the concentration of attention raises questions about the sustainability of the web’s open publishing model. Cloudflare’s move toward charging for or controlling crawler access, and proposals for compensation tied to citations, are attempts to rebuild an exchange of value. Whether they succeed is uncertain. For a practitioner, the implication is to avoid dependence on any single channel.

10.6 What cannot be promised

This section is a statement of limits that practitioners and clients should accept.

  1. No method guarantees a citation, mention or recommendation. OpenAI’s documentation says outright that placement is not guaranteed.
  2. No method guarantees traffic from AI features; the average click effect is negative.
  3. Results differ across platforms, queries, locations and time.
  4. Dashboards give indicators, not exact counts.
  5. Claims of guaranteed AI ranking, “AI SEO in 30 days” and similar are unsupported by the evidence reviewed here.

11. The AI Visibility Stack and a Ninety-Day Implementation Plan

11.1 Why a layered framework

The evidence reviewed in this paper points to several independent conditions that must all hold for a page to be used in an AI answer. A failure at a lower layer cannot be compensated by strength at a higher one: great content on a blocked page is invisible. The AI Visibility Stack is a diagnostic framework organized by these conditions. It is the author’s synthesis, offered as a working model to be tested and revised.

11.2 The five layers

Figure 9. The AI Visibility Stack with diagnostic questions per layer.

Layer 1: Access. Can the relevant systems fetch, render and index the page, and show a snippet? Diagnostics: robots.txt rules by crawler; CDN and firewall rules; status codes; rendering of primary content in HTML; noindex and snippet directives; server logs showing crawler requests. Evidence base: platform documentation (strong).

Layer 2: Retrievability. When a system searches for sub-questions related to the topic, does the page match? Diagnostics: coverage of sub-questions; clarity of headings and passages; topical depth; internal links connecting related pages; ranking for related queries. Evidence base: platform description of fan-out, vendor overlap studies (moderate).

Layer 3: Citability. Once retrieved, does the page offer a passage that the system can use and attribute? Diagnostics: self-contained answers; specific figures, dates and conditions; named sources; definitions; original data; current dates. Evidence base: GEO benchmark, vendor correlations (moderate).

Layer 4: Corroboration. Do independent sources support the same claims and describe the brand consistently? Diagnostics: third-party mentions, reviews, citations of original research, reference-source presence, entity consistency. Evidence base: vendor and secondary reports (weak to moderate; plausible mechanism).

Layer 5: Conversion. When the system uses or sends a user, does the site turn that attention into value? Diagnostics: what the visitor still needs after the summary; landing page fit; tracking of AI-referred sessions; branded-search and assisted-conversion trends. Evidence base: economics literature and behavioral data (varies).

11.3 Diagnosing a gap

The protocol in Section 8.7 yields prompts for which the brand is not cited. For each, the analyst walks up the stack. If the page is not accessible, fix Layer 1. If it is accessible but is not retrieved for related queries, address Layer 2. If it is retrieved, as seen by the fact that competitors with similar content are cited, but not cited itself, review Layer 3. If it is cited rarely but competitors have much wider third-party presence, work on Layer 4. If it is cited but the traffic does not convert, work on Layer 5.

11.4 A ninety-day plan

Days 1–14: Baseline and access.
– Define objectives and the audience; list five to ten priority topics.
– Build the prompt set (Section 8.3) and run the first sample across the chosen platforms.
– Audit access: robots.txt, CDN rules, rendering, indexability, snippet directives.
– Set up analytics segments for AI referrals (including utm_source=chatgpt.com) and connect available platform reports.
– Decide the crawler policy: allow search crawlers; decide on training crawlers.

Days 15–45: Retrievability and citability on priority pages.
– For each priority topic, map sub-questions and identify gaps.
– Rewrite or add sections so each sub-question has a self-contained, specific, sourced answer.
– Add original information where the organization has any: data, tests, case summaries, expert review.
– Update dates and facts; remove obsolete material.
– Add accurate structured data where it serves rich results, without expecting AI effects.

Days 46–75: Corroboration.
– Identify the sources that systems cite for the topic, using sample outputs, and consider how the organization can be accurately represented there.
– Pursue earned coverage: publish original research with a clear press hook; contribute expert commentary; support genuine reviews.
– Check entity consistency across profiles and directories.
– Consider video for topics where demonstration helps.

Days 76–90: Measure, learn, decide.
– Re-run the sample with the same prompts; compare against the baseline with intervals and control prompts.
– Review accuracy and framing.
– Review AI-referred traffic and conversions.
– Decide what to continue, drop or test, and schedule the next cycle.

11.5 Roles and effort

The work spans disciplines: technical SEO, content, public relations, analytics and often legal or compliance. In small teams one person may cover several, but the tasks should still be assigned explicitly. A reasonable allocation for the first cycle is a larger share of time on access and on the priority pages, with measurement set up early, because the baseline cannot be recovered later.

11.6 A page-level checklist

Before publishing or updating a page intended to be cited:

  • The answer to the main question appears near the top in a self-contained passage.
  • Each section’s heading states the question or topic it answers.
  • Claims are specific, dated and attributed to sources; original data is labeled as such.
  • Terms are defined at first use.
  • Comparisons are in tables and procedures in numbered lists.
  • The author, reviewer and date are shown and truthful.
  • The page is indexable, canonical, rendered in HTML and not snippet-restricted.
  • Related questions are answered on the page or linked.
  • The page offers something beyond the summary: a tool, data, example or depth.
  • The page has been checked for accuracy by someone qualified where the topic requires it.

12. Discussion, Limitations and Conclusion

12.1 What the evidence supports

The evidence supports a measured position. AI search visibility is a genuine and measurable outcome with documented eligibility conditions. The relation to classic ranking is real but weakening, so sites cannot rely on rank alone. Well-sourced, specific, current content is more likely to be used, which fits both the academic benchmark and the platforms’ own advice. Independent corroboration plausibly matters, and the economics are less favorable than in classic search because answers reduce clicks on average.

12.2 What the evidence does not support

It does not support the claim that a special set of AI-only techniques, separate from good publishing and sound technical practice, reliably wins citations. The evidence for files such as llms.txt is negative for the major systems. It does not support the claim that structured data alone drives citation. It does not support precise percentages for the effect of any tactic outside the settings in which they were measured. And it does not support promises of guaranteed placement.

12.3 Limitations of this paper

  1. Secondary reporting. Several figures, especially for 2026, come from trade-press summaries of vendor studies, not from the primary publications, and should be verified against the originals before being relied on.
  2. Fast change. Platforms changed repeatedly during the period reviewed, and some statements will be outdated by the time of reading.
  3. No original data. The paper does not report its own citation experiments.
  4. Framework untested. The AI Visibility Stack is a synthesis and has not been validated in a controlled field trial.
  5. Geographic and language scope. Much of the evidence concerns U.S. English-language search.
  6. Vendor interests. Several of the data sources are produced by firms that sell optimization or measurement products.

12.4 Directions for further research

  • Controlled field experiments in which matched pages receive a single content intervention, with citation measured over repeated runs.
  • Open, shared prompt sets and logging standards that would make vendor studies comparable.
  • Studies of the long-term effect of AI exposure on branded search and conversion.
  • Analyses of how citation behavior differs across platforms, languages and query types.
  • Research on the effect of counter-manipulation measures by platforms.

12.5 Conclusion

AI search changes the unit of competition from the position of a page to the use of its information, but it does not replace the foundations. A site that can be fetched, that answers real questions with specific and sourced statements, that is recognized by independent sources, and that offers something beyond a summary is in the best available position. The framework in this paper organizes that work into five layers and a measurement loop, and it insists on the limits: nothing guarantees inclusion, the numbers are noisy, and the best use of the next ninety days is to establish a baseline and learn. As the evidence improves, so should the framework.


References

The list below names the sources used. Because several 2026 items were located through trade-press reports and web search summaries, readers should verify each against the original publication before citing it, and should check publication dates and current versions.

  1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., and Deshpande, A. “GEO: Generative Engine Optimization.” arXiv preprint (2023); Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD 2024).
  2. Google Search Central. “AI features and your website.” Documentation (AI Overviews and AI Mode; eligibility; query fan-out).
  3. Google Search Central. Guide to generative AI features in Search (published May 2026, as reported in trade coverage).
  4. OpenAI. Publisher and crawler documentation: OAI-SearchBot, GPTBot, ChatGPT-User; help-center article on ChatGPT search and publishers.
  5. Pew Research Center. Analysis of Google search behavior with and without AI summaries, based on browsing data collected in March 2025 (68,879 searches; about 900 U.S. adults).
  6. Ahrefs. Studies of AI Overview citation overlap with organic rankings (mid-2025 analysis of about 1.9 million citations; March 2026 follow-up).
  7. BrightEdge. Analyses of AI Overview citation overlap with organic results, as reported in trade coverage.
  8. Surfer SEO. Study of AI Overview citations and fan-out queries, as reported in trade coverage.
  9. SE Ranking. Study of llms.txt presence and citation rates across about 300,000 domains, as reported in trade coverage.
  10. Cloudflare. Radar data and blog posts on AI crawler traffic, purposes of crawling and crawl-to-referral ratios (2025–2026).
  11. Microsoft. Bing Webmaster Tools AI performance reporting (early 2026), as reported in secondary sources.
  12. llms.txt proposal (September 2024) and adoption tracking reports.
  13. Google. Spam policies for Google web search (scaled content abuse; site reputation abuse).
  14. Reports on a September 2025 academic study of earned-media bias in AI search, as summarized in secondary sources.

Appendix A. Glossary

AEO (answer engine optimization): Optimization aimed at being the source of direct answers, including featured snippets, voice answers and AI answers.

AI Mode: Google’s conversational search experience that uses query fan-out to build answers.

AI Overview: A generated summary shown at the top of some Google results, with links to supporting pages.

Citation: A link or attribution to a specific page within a generated answer.

Crawl-to-referral ratio: The number of page requests a crawler operator makes for each visit it sends back to publishers.

Fan-out (query fan-out): The practice of issuing several related searches to answer one question.

GEO (generative engine optimization): Optimization aimed at visibility within generated answers.

llms.txt: A proposed markdown file intended to guide language models to a site’s key content.

Mention: A named reference to a brand or site in an answer, with or without a link.

OAI-SearchBot: OpenAI’s crawler for surfacing sites in ChatGPT search.

GPTBot: OpenAI’s crawler associated with training generative models.

Retrieval-augmented generation (RAG): A design in which a model answers using passages retrieved at the time of the question.

Share of voice: A brand’s proportion of citations or mentions among a defined competitor set for the same prompts.

Zero-click search: A search session that ends without a click to any result.

Appendix B. Source-Quality Grading Used in This Paper

Class Description Treatment
A Platform documentation Authoritative for eligibility and stated behavior; silent on selection
B Peer-reviewed or independent behavioral research Strong on method; may be dated or narrow
C Vendor studies with disclosed samples Directional; check definitions and period
D Trade-press summaries and secondary reports Locate leads; verify against primary
E Practitioner assertion without data Hypothesis only

Appendix C. Prompt-Set Template for Measurement

Field Example entry
Prompt ID P-014
Prompt text “What is the best way to [task] for [audience]?”
Funnel stage Consideration
Topic cluster [Cluster name]
Brand/non-brand Non-brand
Platforms Google AI Overviews, AI Mode, ChatGPT search, other
Runs per period 5 or more
Recorded outputs Cited URLs; brand mention; position; accuracy flag; framing
Notes Location, language, logged-in state

Leave a Message

Sent under NDA. We reply from leo@ or [email protected].

Emily Thompson

Content Writer

Writes EMA’s adult SEO and escort marketing coverage, working from the accounts the team runs. Commercial detail in every engagement stays under NDA.

Related locations

Services referenced