Research
Original research on AI answers, citations, retrieval, source patterns, and representation across models and topics.
A fixed-panel study examines freshness, length, headings, images, lists, tables and schema signals on pages cited by Claude, Gemini, OpenAI and Perplexity.
Answers: How recent, long and structurally complex were the pages cited by Claude, Gemini, OpenAI and Perplexity in the same B2B research benchmark?
Read researchA fixed-panel study compares independent, commercial, competitor, first-party and page-format sources cited by Claude, Gemini, OpenAI and Perplexity.
Answers: How much of the evidence cited by Claude, Gemini, OpenAI and Perplexity came from independent publishers, commercially interested sources, competitors and first-party pages?
Read researchA fixed-panel study of Claude, Gemini, OpenAI and Perplexity separates citation volume, material-claim citation coverage and reviewed source support.
Answers: Do AI systems that produce more citations also cite a larger share of material claims and provide stronger evidentiary support?
Read researchA fixed-panel study across Claude, Gemini, OpenAI and Perplexity found very low URL, domain and publisher overlap for the same B2B buyer questions.
Answers: When Claude, Gemini, OpenAI and Perplexity answer the same buyer question, how often do they cite the same URLs, domains and publishers?
Read researchA matched-question study of how Claude, Gemini, OpenAI and Perplexity differ in citation density, source mix, freshness, overlap, author visibility and candidate selection.
Answers: When Claude, Gemini, OpenAI and Perplexity answer the same B2B buyer questions, how different are the sources and evidence environments they cite?
Read researchA cross-platform analysis finds AI citation domains have a concentrated head and broad long tail, with different top-k concentration across ChatGPT, Gemini and Perplexity.
Answers: How concentrated are observed AI citations across source domains, and how does that concentration differ across platforms and observation periods?
Read researchThe capstone defines where citation-rank findings apply, why adjusted associations remain observational, and what randomized experiment is needed next.
Answers: What can observational citation-rank research establish about AI-generated answers, and what requires a randomized experiment?
Read researchOnly 25 recommendations occurred across 1,576 eligible candidates, with no credible Rank-1 advantage over Ranks 3–5.
Answers: Does a source appearing at Rank 1 make its associated entity more likely to be explicitly recommended in an AI-generated finance answer?
Read researchHigher-ranked cited-source entities appeared more often and earlier in finance answers, but most Rank-1 entities were still unnamed.
Answers: Does higher cited-source position correspond to greater source-entity visibility and earlier first-mention position in generated finance answers?
Read researchShared Rank-1 sources did not make finance answers converge, while Rank 1 showed a modest 1.47-percentage-point semantic-alignment premium.
Answers: Why can Rank 1 show a modest within-response semantic-alignment premium without making answers that share the same first cited source converge?
Read researchA study of 1,500 finance responses finds modest Rank-1 alignment and greater entity visibility, but no credible similarity or recommendation effect.
Answers: Does a source appearing at Rank 1 shape AI-generated finance answers through shared-source similarity, semantic alignment, entity visibility or explicit recommendation?
Read researchA cross-platform study finds that final cited answers commonly look answer-ready, while finalist-level readiness decisions and exact claim-to-citation mappings remain unobserved.
Answers: Can surviving AI-cited evidence support a clean, specific final answer, and can that readiness decision be validated from final outputs?
Read researchA cross-platform study finds that final AI citation sources are query-specific and temporally reproducible, but the hidden candidate-level quality gate remains unobserved.
Answers: Do final AI citation sources show evidence of selective quality filtering, and can the underlying candidate-level quality gate be validated from final outputs?
Read researchA matched-decoy study finds that selected Gemini citation fragments align much more strongly with their originating prompts than with plausible alternatives, while the hidden relevance gate remains unobserved.
Answers: Do passage-like fragments selected in final AI citations semantically match the query that produced them?
Read researchA cross-platform study finds emitted AI citation URLs describe final citation structure but do not prove historical page fetching, parsing or text extraction.
Answers: Does an emitted AI citation prove that the platform successfully accessed a source and converted it into usable text?
Read researchA large cross-platform study of final citation source composition found Brand-owned evidence in roughly two-thirds of observed answers and Comparable-vendor evidence in about half, with the source mix changing by platform and buyer-question type.
Answers: Which source families appear in final AI citations, and how does the source mix differ across ChatGPT, Gemini, Perplexity and different buyer-question types?
Read researchA stress test finds more than three-quarters of citation misses had a contemporaneous citation elsewhere; candidate sufficiency remains unknown.
Answers: When an AI platform produces no visible citation, how often does the same contemporaneous question receive citations elsewhere, and what does that reveal about ecosystem availability versus candidate sufficiency?
Read researchA cross-platform study finds query-type citation differences on Gemini, where a raw Educational deficit shrinks after controlling for retrieval need.
Answers: How does buyer-question type relate to visible AI citation exposure across platforms, and why does Gemini's raw Educational deficit shrink from roughly 29 points to about 5 points in the comparable Medium-retrieval-need cohort?
Read researchA capstone analysis of 496 finance AI responses showing why descriptive GEO scores cannot automatically become market rankings or authority effects, and what stronger experiments require.
Answers: What experimental design is required before descriptive GEO observations can support defensible company rankings, market-visibility claims, authority associations or causal conclusions?
Read researchA companion analysis of 496 finance AI responses found 923 recurring source-identity pairs. Support, Jaccard similarity and lift reveal repeated answer contexts, but not backlinks, partnerships, authority or causal influence.
Answers: Which inferred source identities repeatedly appear together in finance AI responses, and what can support, Jaccard similarity and lift responsibly tell us about those recurring answer contexts?
Read researchA companion analysis of 496 finance AI responses showing why 4,660 named-source objects supported response- and publisher-level measurement while only two canonical page identities supported exact page attribution.
Answers: When is an AI source record resolved well enough to support a page-level citation claim, and how should measurement change when identity stops at the source-name or publisher level?
Read researchA companion analysis of 496 finance AI responses found 3,028 external named-source objects across 1,197 inferred publisher identities, revealing a fragmented evidence ecosystem spanning press distribution, video, communities, reviews and specialist sources.
Answers: Which external source environments recur around target-oriented finance AI answers, how fragmented is that ecosystem, and what can publisher response incidence actually tell us?
Read researchA companion analysis of 496 finance AI responses explaining why 90.7% target-owned source presence measures branded discoverability, not general AI market visibility, share of voice or recommendation probability.
Answers: What does a 90.7% target-owned source rate in target-oriented finance AI prompts actually measure, and why is it not the same as general market visibility or AI share of voice?
Read researchA 496-response finance GEO study finds high first-party citation presence, a fragmented external publisher ecosystem and major limits on page-level conclusions.
Answers: When a finance company is already part of a target-oriented question, how often does a source attributable to that company appear, and what can the surrounding publisher ecosystem support?
Read researchAn analysis of 167 role-framed marketing responses found more financial language in leadership-labelled prompts, but no clean strategy/execution divide.
Answers: When marketing roles are explicitly named in AI prompts, which decision criteria become more prominent in the resulting advice?
Read researchA follow-on analysis of 1,494 finance-persona responses tests whether residual persona differentiation survives removal of exact role and profile language.
Answers: After accounting for prompt design, how much persona differentiation remains when exact persona labels and profile phrases are removed, and where does the residual signal appear most strongly?
Read researchA 750-response study found persona-specific shifts in search, pipeline, AEO and SEO language, with important causal limitations.
Answers: How do five persona prompt conditions change AEO, SEO, search and pipeline language across matched business scenarios?
Read researchA 1,500-prompt study found that most raw persona similarity in AI responses was explained by prompt construction; the residual signal was modest.
Answers: After accounting for prompt construction and experimental design, how much persona-related signal remains in grounded AI responses and search queries?
Read researchA large cross-platform study finds retrieval need strongly associated with visible AI citations on Gemini, while ChatGPT shows little separation and Perplexity operates near citation saturation.
Answers: How does retrieval need relate to visible citation exposure across ChatGPT, Gemini and Perplexity, and why is the Gemini result a Low-versus-rest pattern?
Read researchKojable analysed 180 grounded Gemini prompts and 1,620 search-query instances to measure how prompt wording changes AI retrieval behaviour.
Answers: When similar prompts are phrased differently, does Gemini search for similar information?
Read researchKojable tested 180 B2B finance prompts in Gemini to measure whether similar prompts produce similar answers for AEO monitoring.
Answers: How many AI prompts do teams need to monitor when similar prompts produce similar answers?
Read researchExplore 250 high-coverage cited URLs across 13 technical topic neighborhoods and three executive citation systems.
Answers: Which publishers and brands control observed AI citations within each topic?
Read researchA research note challenging the idea that Reddit is a universal AI visibility lever for fintech. The evidence points toward vertical media, analyst content, and review platforms as stronger sources for regulated financial topics.
Answers: Is Reddit really a universal visibility lever for fintech AI search?
Read researchData-backed analysis of how search-result position bias may carry into AI-generated answers, affecting which companies are surfaced, cited, and recommended.
Answers: How does search-result position bias shape AEO visibility?
Read research