Research library
Original research and practical guidance on AI search, with a note on what each source can—and cannot—tell you.
Start with platform documentation for platform rules. Read studies for their methods as well as their findings. University web-team advice is distinct from academic research; commercial studies can be useful without proving that a tactic caused a result.
This is a curated collection, not a ranking of the most AI-cited websites or an exhaustive literature review. We reviewed the published material at the depth shown below; we have not independently reproduced the studies. Inclusion does not imply endorsement of our agency.
Learn how to evaluate an AEO claim · Back to the knowledge base
SEO fundamentals apply; special AI files and special schema are not prerequisites.
Keep in mind: Google-specific. Eligibility is not guaranteed indexing, citation or traffic.
Accessible, indexed, snippet-eligible content can be considered; Search Console includes AI-feature traffic in Web reporting.
Keep in mind: Do not equate Search Console Web totals with isolated AI traffic or Gemini app performance.
Original value, clear sourcing and an identifiable responsible publisher support useful content.
Keep in mind: E-E-A-T is not a purchasable badge or a numeric score we can promise to increase.
Creating many low-value pages to manipulate visibility can violate policy.
Keep in mind: Automation itself is not the only issue; purpose and value matter.
FAQ rich results were retired in May 2026; newer llms.txt guidance clarifies no Google Search benefit.
Keep in mind: A visible FAQ can still be useful. Feature retirement is not a ban on question-and-answer content.
OAI-SearchBot, GPTBot and ChatGPT-User have different functions.
Keep in mind: Crawl access is not a citation guarantee; user-triggered access differs from automatic crawling.
PerplexityBot supports search; Perplexity-User handles requested page visits.
Keep in mind: The user fetcher generally ignores robots.txt; verify current published IP ranges rather than trusting a user-agent string.
Site owners can inspect citation activity across supported Microsoft and partner experiences.
Keep in mind: Aggregated citation counts do not establish placement, authority or visits; query data is sampled. Preview features may evolve.
Reports improvements up to 40% within evaluated conditions, varying by domain.
Keep in mind: Not a 40% promise for live clients, traffic or revenue; benchmark visibility differs from organic discovery.
Separates whether claims have citation support from whether cited material supports a claim.
Keep in mind: Historical engine versions; do not reuse reported accuracy rates as current performance.
Treats answer quality and citation quality as separately evaluable properties.
Keep in mind: A model benchmark does not identify which website edits earn live commercial citations.
External retrieval can inform generated answers alongside model parameters.
Keep in mind: Not documentation of every modern search product, and not a website-ranking experiment.
Argues for separate discovery, citation, content-use and economic measures.
Keep in mind: Peer review not established; abstract-level review here. Its corpus-level conclusions need full-text checking before quotation as settled evidence.
Highlights accurate, accessible information and readable organization.
Keep in mind: Practical advice, not controlled university research proving citation lift. Broad statements about AI behavior need provider-specific checks.
Recommends useful headings, lists and organized answers.
Keep in mind: Screenshots do not establish causality. Its llms.txt discussion is not evidence of a Google ranking benefit; do not repeat its restriction of AI Mode to informational searches.
Observed misattribution, incorrect URLs and confident errors.
Keep in mind: News-identification task and historical versions; not an error-rate estimate for local contractor recommendations.
Web mentions had a reported correlation of 0.664, greater than backlinks at 0.218.
Keep in mind: Association is not causation. Brand mentions are not equivalent to website citations; broad brands may not represent a local pilot.
No clear positive citation uplift across the studied surfaces.
Keep in mind: Pages were already heavily cited; concurrent edits, recrawl timing and pooled schema types limit attribution. Not proof that schema is useless for new sites.
Reports 58% lower position-one CTR relative to its modeled comparison.
Keep in mind: Observational, selected keyword groups and two-year gap. Not a 58% loss forecast for every business, device or query.
Observed differences in cited-page age between platforms.
Keep in mind: First-seen date is a proxy, not certain publication time. Descriptive age differences do not prove that changing a date improves citations.
Citation-source mix changed over time and differed across three search surfaces.
Keep in mind: All-topic domain popularity is not an AEO-topic source ranking. Proprietary prompt selection limits generalization.
Cited and comparison pages differed on several scored content qualities.
Keep in mind: Reported percentage differences are feature-score differences, not percentage gains caused by edits. Commercial scoring method is not a randomized intervention.
Describes technical and engagement patterns among cited URLs.
Keep in mind: Cited-only samples and raw frequencies do not establish an optimal slug length. Post-click engagement does not establish a citation input signal.
Distinguishes citation-only, mention-only and combined appearances.
Keep in mind: The detailed 61.7% denominator is all appearances, despite the headline saying citations. Do not reproduce that headline statistic without reconciliation.
Traditional-result clicks occurred on 8% of visits with summaries versus 15% without; summary-link clicks on 1%.
Keep in mind: Observational panel and query differences; reconstructed SERPs need not exactly match the original user view. Not a causal local-business forecast.
Shows how an agency positions AEO services for universities.
Keep in mind: Carnegie is a vendor here, not Carnegie Mellon research. Service promises are not independently validated results.
Make accurate information accessible, keep an evidence record, and test a specific change with the same measurement protocol. A study of large brands or historical engine versions does not establish what will happen to a new service-business website.
See our measurement methodology or use the website checklist.
Which service do you want more customers for—and how do they find you today?