AI search products can retrieve web pages, use other information sources, and show links alongside generated answers. Their methods differ, and most selection details are not public. We can describe what a provider documents and measure which links appeared in a defined sample; we cannot infer one universal citation formula from a crawler visit or a handful of answers.
That distinction matters for anyone working on AI visibility. A brand can be mentioned without its website being cited. A web crawler can visit a page without that page appearing in an answer. And a citation study can show a pattern without proving that a particular tactic caused it.
What does “choosing a source” mean in an AI answer?
A visible citation is a source link or reference displayed with an answer. It is an observable output—not a complete record of every page a system considered, every source behind a model’s response, or the weights used to select a link.
Keep these events separate:
- Crawling: a bot or fetcher requests a page.
- Search eligibility: a page can be indexed or surfaced under a platform’s documented requirements.
- Retrieval or source use: information is obtained for a particular task or answer. Providers disclose different amounts about this stage.
- Citation: a page is visibly linked or named as a source in an answer.
- Brand mention: the answer names a brand, whether or not it links to the brand’s site.
- Referral: a user clicks through and arrives at the site.
These signals answer different questions. None can stand in for all the others. For measurement definitions and denominators, see AI Visibility Metrics Explained and How to Measure AI Visibility.
What does Google document about AI Overviews and AI Mode?
Google says its AI Overviews and AI Mode use the same foundational Search eligibility and SEO practices as other Search experiences. Its documentation describes query fan-out: the system may issue related searches across subtopics and data sources while developing a response. Google also says the two features can use different models and techniques, and that the sources shown can vary. A page must be indexed and eligible to appear with a Search snippet; that eligibility does not guarantee inclusion.
Google’s public guidance does not tell site owners to add a special AI schema, create an AI-only file, or divide every page into tiny chunks to earn a citation. Structured data can still support ordinary Search features when it accurately matches visible page content, but it is not a documented shortcut to AI inclusion. Google’s own guidance is the right source for Google-specific claims; it should not be generalized into a description of ChatGPT or Perplexity.
What do crawler documents tell us about ChatGPT and Perplexity?
Crawler documentation is useful for understanding access controls. It is not a disclosure of the complete answer-generation or source-ranking system.
OpenAI documents separate controls for OAI-SearchBot, which helps surface sites in ChatGPT search, GPTBot, which relates to potential training use, and ChatGPT-User, which can fetch pages in response to a user action. Perplexity likewise distinguishes PerplexityBot, used to surface and link sites in search results, from Perplexity-User, which can fetch pages for a user-requested answer. The providers describe those roles separately in their OpenAI crawler documentation and Perplexity crawler documentation.
That distinction gives site owners a practical choice about access. Allowing a search crawler can make a site eligible for that provider’s search experience; it does not guarantee a mention, citation, recommendation, or click. A successful request in a server log proves a request was made—not that the page was later retrieved, selected, or displayed.
For provider-specific access checks, see Technical AI Search Optimization and AI Crawler Analytics. Do not treat one platform’s bot names or controls as a universal AI crawler policy.
What have researchers measured about AI search citations?
Research is beginning to quantify source visibility, but each paper measures a particular system, dataset, and outcome. The results are useful as bounded evidence, not a universal ranking formula.
Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024)
What it measured: Experiments using GEO-Bench, a 10,000-query benchmark, and generative search settings
Relevant finding: Some tested content changes improved visibility by up to 40%; results varied by domain.
What this does not establish: A 40% lift for every site, engine, query, or tactic.
Zhang et al., Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines (2025 preprint)
What it measured: 55,936 queries across six LLM-based search engines and two traditional search engines
Relevant finding: 37% of domains were unique to the LLM-based systems in that study’s comparison.
What this does not establish: That “37% of citations overlap,” or that 37% is a stable web-wide rate.
Einarsson et al., Curated Retrieval versus Open Web Search in Public AI Information Services (2026 preprint)
What it measured: Five experts evaluated 449 answers from one government-funded public information service; cited sources were flagged separately
Relevant finding: At least one source was flagged in 65 of 187 reviewed web-search answers (35%). A trusted-domain list in the system prompt raised the share of citations to listed domains from 12% to 21%.
What this does not establish: That 35% of all citations across consumer AI products are untrustworthy, or that prompt steering has the same effect in commercial tools.
Kumar, Generative Engine Optimization at Scale (2026 preprint; author affiliated with Ranqo)
What it measured: More than 100,000 prompt responses across 100+ brands tracked by the vendor during March–May 2026
Relevant finding: The report describes brand-maturity differences and says best-of listicles represented about 21% of citations in its dataset.
What this does not establish: A causal result that creating or entering a listicle will earn citations; the author notes causal tests as future work.
The GEO paper is a foundational benchmark study, but its “up to 40%” result is conditional on its methods and tested settings; its authors also report that tactic effectiveness varies by domain. The source-coverage paper’s 37% figure means unique domains in its comparison, not overlap. The curated-retrieval paper’s 35% figure is a share of reviewed answers that had at least one flagged source—not a share of all citations. Those distinctions are essential when turning research into advice.
Can a study show why a source was selected?
Sometimes, but the study design determines how far the conclusion can go. A controlled intervention can test an effect in the system and setting it controls. An observational study can identify associations or describe repeated patterns. Neither automatically reveals how a separate commercial platform works.
For example, the curated-retrieval study tested a particular public information service and found that adding a trusted-domain list to the system prompt increased the listed domains’ citation share from 12% to 21%. That is evidence about that intervention in that system. It is not proof that brands can prompt ChatGPT, Perplexity, or Google to cite them by publishing on a particular site.
Be especially careful with claims that a GEO score “predicts nothing,” that Reddit presence or backlinks are the “#1 driver,” or that listicles always win. To publish those as findings, an article needs the exact source, sample, definition of visibility, comparison group, statistical method, and limits. A correlation coefficient is not a causal effect, and statistical significance alone does not establish a useful or generalizable effect.
What can a brand team measure reliably?
A team can measure what it observed if it records the conditions and denominator. A simple repeatable protocol is more useful than an opaque score with an unknown sample.
- Write a fixed prompt set. Group prompts by intent—discovery, comparison, and direct brand questions—and record the date the set was created.
- Record the platform and mode. Note the product, search or answer mode, locale, language, account state when relevant, and date/time.
- Repeat a sample. Answers may change across runs. Keep successful, failed, and unavailable runs visible in the data.
- Record outputs separately. Capture brand mentions, visible citations to the brand’s own domain, third-party citations, recommendation context, and referral traffic as distinct fields.
- Publish raw counts with rates. For example, report “12 of 40 valid answers included a brand mention,” then state what counted as a valid answer. Do not compare that rate to another provider unless the prompt set and collection method are comparable.
- Save evidence. Keep the prompt, answer, linked URL, screenshot or response record, timestamp, and coding notes, subject to platform terms and privacy requirements.
Google now documents a Search Console report for impressions from supported generative Search features. That is valuable first-party data for those Google features, but it is not the same measure as sampled citations in ChatGPT or Perplexity. A site may also monitor logs and analytics, but crawler requests, impressions, citations, and sessions describe different points in a discovery path.
See How to Measure AI Visibility for the baseline workflow and AI Visibility Metrics Explained for definitions. If you compare tracking platforms, use their own documentation to verify each product’s sources and metric definitions; vendor scores are not automatically comparable.
What practical actions are supported by current evidence?
The most defensible actions improve ordinary discovery, accuracy, and measurement—not promise a citation:
- Keep important pages accessible to the crawlers and search systems your organization chooses to serve.
- Make core facts easy for people to find and verify; keep public descriptions consistent with the product or service.
- Use clear page structure and internal links to help users navigate related material.
- Earn accurate, legitimate coverage and references from relevant third parties; do not manufacture mentions or forum posts.
- Check what each platform officially documents before changing robots.txt, firewall, or content settings.
- Measure mentions, citations, source URLs, impressions, and referrals separately.
- Rerun a stable sample after meaningful changes, and describe the result as an observation unless the study design supports a causal claim.
Google says its generative Search features use standard Search eligibility and helpful-content foundations. OpenAI and Perplexity document separate crawler controls for their products. Those are actionable boundaries. A precise “citation formula,” secret weighting system, or universally winning page format is not established by those documents.
Common claims about AI source selection that need qualification
“This tactic boosts AI citations by 40%.” The GEO paper reports up to 40% in specific experimental settings. State the paper, tested method, and limits; do not turn the maximum result into a promised lift.
“Adding schema makes AI systems cite a page.” Google says no special schema is required for AI Overviews or AI Mode. Structured data should match visible content and serve its ordinary supported purposes.
“A crawler visited, so the page was used in an answer.” A request is evidence of access, not evidence of retrieval or citation.
“37% overlap proves AI search uses a different web.” The cited study reports 37% of domains unique to its LLM-based search sample relative to its traditional-search comparison. The number belongs to that study’s dataset and definition.
“35% of AI citations are untrustworthy.” The cited curated-retrieval study found a flagged source in 35% of a specific set of web-search answers. That result is not a cross-platform citation rate.
“Listicles, Reddit, or backlinks are guaranteed ranking levers.” A vendor-associated preprint reports that best-of listicles represented about 21% of citations in its 100-brand dataset. That is a descriptive result, not evidence that creating a listicle causes a brand to be cited. Treat such claims as hypotheses unless a controlled study demonstrates the effect in the target setting.
Frequently asked questions
Do AI search engines use the same source-selection rules?
There is no public evidence of one shared rule set. Providers publish different documentation, and the visible answers and links can vary by platform, query, and run. Describe each provider’s documented behavior separately.
Does allowing an AI crawler guarantee that a brand will be cited?
No. Crawler access can affect whether a provider can fetch or surface a page, but it does not guarantee that a page will be selected or cited in a particular response.
Is a brand mention the same as a citation?
No. A mention is text naming the brand. A citation is a source reference or link shown with an answer. Record both because one can appear without the other.
Can I measure AI visibility with one score?
You can define a composite score for a specific workflow, but there is no universal standard. State the inputs, weights, prompt sample, platform, and denominator so readers can interpret it. Keep raw mention, citation, and referral counts available.
Sources and further reading
- Google Search Central: AI features and your website
- Google Search Central: optimizing for generative AI features
- OpenAI: Overview of OpenAI crawlers
- Perplexity: Perplexity crawlers
- Aggarwal et al.: GEO: Generative Engine Optimization
- Zhang et al.: Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines
- Einarsson et al.: Curated Retrieval versus Open Web Search in Public AI Information Services
- Kumar: Generative Engine Optimization at Scale: Measuring Brand Visibility Across AI Search Engines
Editorial note: This article explains documented behavior and bounded study results. It does not claim firsthand testing of commercial AI answer engines or promise that any tactic will earn a citation. Platform documentation and study versions were checked September 28, 2026; recheck before republishing later.
Research and platform documentation checked September 28, 2026. AI answer visibility varies by query, platform, and run; this article describes documented practices and bounded study findings.