The 2026 AI Citation Study

Executive summary

Two years into the AI search era, most of what's published about how AI engines pick their citation sources is either marketing speculation or single-engine micro-studies that don't generalize. This research synthesizes what we've actually observed across our 2025–2026 client engagements — patterns consistent enough across enough programs that we're willing to publish them as findings.

Ten findings, in order of how much they matter for an SEO program planning the next 12 months:

• Citation eligibility tracks classical ranking — but the correlation is loose, not deterministic

  Pages ranking outside the top 20 classical results are rarely cited; pages ranking in the top 5 are cited disproportionately often. But the citation-or-not call between top-5 candidates is decided by signals beyond classical rank — and those signals are where program optimization actually has leverage.
• Citation share is dramatically lopsided across engines

  AI Overviews and ChatGPT command the largest citation universe by a wide margin; Perplexity and Gemini are smaller but rising; Copilot is meaningfully smaller still. Resource allocation decisions inside an SEO program should reflect this lopsidedness rather than treating engines as equal-weight.
• AI Overview trigger rates vary enormously by query intent

  Informational and comparison queries trigger AI Overviews far more often than transactional or navigational queries. The implication is that AIO optimization is mostly a content-team conversation, not an ecommerce-product-page conversation.
• Schema-rich pages are cited more often than schema-thin pages of comparable rank

  Article, FAQPage, HowTo, Person, and Organization schema all correlate with citation lift in our data. The lift is directionally meaningful, not 10x — but it's consistent enough that schema is the single most leveraged tactical move we see in client work.
• Author E-E-A-T markup correlates with citation lift

  Pages with explicit, verifiable author bylines, sameAs Person markup, and author bio pages are cited more than uncredited or thinly credited equivalents. The lift is most pronounced in YMYL content.
• Freshness matters more in Perplexity than in ChatGPT

  Perplexity strongly prefers content updated within the last 6–12 months for most query types. ChatGPT is more forgiving of older content if the underlying expertise signals are strong. Editorial calendars should reflect this difference.
• Citation share is unstable week-to-week — and most tools hide that instability

  Same query, same week, repeated runs return different citation sets. Tools that report binary cited-or-not on single runs overstate signal quality. Our recommendation is to look at citation frequency over multiple runs rather than point-in-time citation status.
• Brand-name co-occurrence searches lift citation rate

  Pages on domains with high brand-name search volume (people Googling 'BrandName + topic') are cited more often than equivalent pages on lower-brand-volume domains. This is the AI engine version of the classical brand-strength signal.
• fr-CA citation share lags en-CA by a wide margin

  Across our bilingual Quebec engagements, French Canadian content is cited materially less often than English Canadian content, even when content quality and schema parity are matched. The gap is closing slowly but remains real.
• 'Information gain' is observable in citation patterns

  Pages with original data, proprietary frameworks, or unique perspectives that aren't already present in the top-cited cluster get cited at higher rates. The corollary: rewritten or paraphrased content is increasingly invisible to AI engines as the citation cluster matures.

Methodology and disclosure

Editorial research, not academic research. We say this up front so readers can calibrate trust appropriately.

Sample composition. The patterns synthesized here draw on dozens of active client engagements across the 2025–2026 period — primarily Canadian B2B, professional services (legal, healthcare, accounting), ecommerce, and SaaS verticals. Engagement size ranges from small in-house programs (under 100 priority queries tracked) to enterprise programs (1,500+ priority queries tracked). The sample is intentionally not random — it reflects the businesses that hire a Toronto-headquartered senior SEO agency, which skews toward higher-quality websites with intentional SEO programs.

Observation period. Patterns observed across roughly 18 months of active tracking, ending Q1 2026. Where engine behavior shifted materially during the observation window, we've noted the shift rather than averaging across it.

Engine coverage. Google AI Overviews and ChatGPT (including ChatGPT search) are observed most thoroughly. Perplexity is observed regularly. Gemini and Microsoft Copilot have less observation depth — findings on those engines are noted as such in the relevant sections.

Tracking stack. The observations are not the output of a single tool. Our standard client tracking stack combines a dedicated AI rank tracker (typically Otterly or Peec for mid-market clients, Profound for enterprise clients), a brand mention monitor (typically AthenaHQ), the AI module of whichever classical SEO platform the client uses (Semrush AI Toolkit or Ahrefs Brand Radar), and weekly Search Console AI Overview impression review. Findings are cross-referenced against tools and verified with manual SERP audits on priority queries.

What we did not do. We did not publish exact percentages. We did not control variables to academic-experiment standards — this is observational research drawn from real client work, not a randomized controlled trial. We did not include client names, vertical-specific volumes, or any data that could reverse-identify a client engagement.

What we did do. We synthesized patterns that appeared consistently enough across enough engagements to be willing to publish under our own name. Where patterns were ambiguous or vertical-specific, we either omitted them or noted the ambiguity in the finding.

Finding 1: Citation eligibility correlates with classical ranking — but not perfectly

The single most useful framing for thinking about AI citation behavior is: classical ranking determines candidacy; other signals determine selection.

In observation after observation, pages ranking outside the top 20 classical Google results are rarely cited in AI Overviews. Pages ranking in the top 5 are cited disproportionately often. But within that top-5 candidate pool, the citation-or-not call appears to be decided by signals beyond classical rank — and those signals are where SEO program optimization has the most leverage.

Implication: programs that focus exclusively on classical rank improvement and ignore the secondary signals are leaving citation share on the table. Conversely, programs that obsess over secondary signals while ignoring classical rank are optimizing past their candidacy ceiling.

Finding 2: Citation share is lopsided across engines

One of the persistent mistakes we see in AI SEO planning is treating engines as roughly equal-weight when allocating optimization effort. Our observation across client work is that the engines are dramatically lopsided in citation universe size and in audience reach. Resource allocation should reflect that.

Implication: when prioritizing AI SEO investment for the typical Canadian B2B or professional services program, the resource split should disproportionately favor AI Overviews and ChatGPT, with Perplexity earning more weight in research-driven verticals and Copilot earning more weight in Microsoft-enterprise audiences. Treating all engines as equal-weight in resource allocation produces under-optimized outcomes on the engines that drive most of the value.

Finding 3: AI Overviews trigger unevenly by intent

AI Overview presence is not uniform across query types. The pattern is consistent enough across our client work to be predictable, which is useful when prioritizing what to optimize for.

Implication: AIO optimization is largely a content-team conversation, not an ecommerce-product-page conversation. The pages that benefit most from AIO citation lift are educational, comparison, and how-to content — which is exactly the category most existing SEO programs underinvest in relative to commercial pages.

Finding 4: Schema-rich pages get cited more often

Across our observation, pages with extensive structured data markup are cited more often than schema-thin pages of comparable classical rank. The lift is directionally meaningful, not 10x — but it's consistent enough that schema is the single most leveraged tactical move we see in client work.

The schema types that correlate most strongly with citation lift in our observations:

• Article + author markup with sameAs Person attributes

  Particularly strong for editorial and research content. The lift is most visible when authors have verifiable digital footprints (LinkedIn, professional bios, prior publications) referenced in sameAs.
• FAQPage and QAPage markup

  Strong correlation with citation in answer-format AI surfaces (AIO especially). Pages structured around explicit Q&A pairs appear to be easier for AI engines to extract and cite at the snippet level.
• HowTo markup for instructional content

  Stronger lift in AIO than in ChatGPT, but observable across both. Step-numbered HowTo content is structurally well-suited to AI summarization.
• Organization markup with hasCredential, awards, and trust attributes

  Particularly important for YMYL content where E-E-A-T weighting is heavier in citation selection.
• Product + Review markup for ecommerce

  Lift is more visible in shopping-intent surfaces (ChatGPT shopping, AIO shopping) than in informational AI surfaces.

Implication: schema is the highest-ROI tactical investment in most AI SEO programs because it's relatively cheap to implement, durable across model updates, and consistently correlated with citation lift. We typically prioritize schema audits in the first 30 days of any new AI SEO engagement.

Deep dive: we published a focused companion study on this finding alone — see The 2026 Schema Citation Lift Study, which observes ~600 pages, breaks the lift down by schema type, ranks engines by schema sensitivity (Perplexity weighs schema most), and includes a 90-day implementation program.

Finding 5: Author E-E-A-T markup correlates with citation lift

Beyond pure schema, the broader Author E-E-A-T signal — explicit bylines, author bio pages, credentials, sameAs Person markup, and verifiable digital footprint — correlates with citation lift in our data. The effect is most pronounced in YMYL content (medical, legal, financial) but observable across editorial content broadly.

The pattern that emerges from our work:

• Uncredited content is cited less often

  Articles with no author byline are cited less than otherwise-equivalent articles with named authors. The gap is widest in YMYL.
• Named-author-with-bio outperforms named-author-without-bio

  An author byline that links to a real bio page with credentials and a digital footprint outperforms a named author without that infrastructure.
• sameAs Person markup is a force multiplier

  Linking authors to verifiable identities (LinkedIn, ORCID where applicable, prior publications, professional associations) appears to compound the byline signal materially.
• Editorial review chains help YMYL content specifically

  Pages with explicit reviewedBy markup (a credentialed reviewer separate from the author) are cited more often in medical, legal, and financial verticals.

Implication: every editorial program should have author infrastructure as a baseline requirement, not a nice-to-have. The cost of building an author bio page with proper markup is small; the citation lift is consistent.

Finding 6: Content freshness matters more in Perplexity than ChatGPT

Perplexity has the strongest observable freshness preference of the major engines we track. For most query types, Perplexity preferentially cites content updated within the last 6–12 months over older content of comparable quality. ChatGPT is more forgiving of older content when the underlying expertise signals are strong. AI Overviews fall somewhere in between, with notable variation by vertical.

Implication: editorial calendars should reflect engine-specific freshness preferences rather than treating refresh as one-size-fits-all. For programs heavily exposed to Perplexity (research-driven B2B, technical content), quarterly refresh of priority content is the cadence we recommend.

Finding 7: Citation share is unstable week-to-week — and the tools mostly hide it

This is the finding that most contradicts the way the AI rank tracking tool category currently markets itself. AI engine citation behavior is genuinely noisy. Same query, same week, repeated runs return different citation sets. Tools that report binary cited-or-not on single runs are overstating the precision of their signal.

What we observe in practice when running repeated queries against the same engine in the same week:

• Citation set overlap between runs is consistently incomplete

  Even within a single week, repeated identical queries return citation sets that share the majority of sources but vary in 1–3 positions out of 5–8 cited sources. The variance is real and is not a tool bug.
• Position within the citation set is even more unstable

  Whether your domain is the first cited source vs the fourth varies more than whether you're cited at all. Position-tracking tools that report exact position are tracking noise as much as signal.
• Frequency-over-runs is a more honest metric than binary citation

  Reporting 'cited in 7 of 10 repeated runs this week' is a more honest signal than 'cited' or 'not cited' on a single run. The best tools (Profound notably) report this; most tools don't.
• Trend lines smooth out the noise — but require enough observation period to be meaningful

  Weekly noise washes out at the monthly trend level. We recommend judging citation share movement against monthly trends, not week-over-week deltas.

Implication: AI rank tracking dashboards should be read with the same skepticism you'd apply to any noisy metric. Trend over time matters; point-in-time snapshots matter much less than the dashboards make them appear. Programs making weekly tactical decisions off single-run citation tracking are making decisions on noise.

Finding 8: Brand-name co-occurrence searches lift citation rate

One of the more durable findings across our observation: pages on domains with high brand-name search volume (users searching for 'BrandName' alongside topic queries) are cited more often than equivalent pages on lower-brand-volume domains. This is the AI engine version of the classical brand-strength signal — and it appears to compound over time.

The mechanism we hypothesize (which we can't directly verify but which fits the observed pattern): AI engines weight source authority, and brand-name search volume is one of the strongest available proxies for real-world authority. A domain with rising brand search volume is, in aggregate, a domain users trust enough to seek out by name — and AI engines appear to encode that trust in citation selection.

Implication: PR, brand-building, and direct-traffic-generating activities are not separate from AI SEO — they feed it. Programs that obsess over technical SEO and ignore brand investment are optimizing one variable while ignoring an equally important one.

Finding 9: fr-CA citation share lags en-CA by a wide margin

Across our bilingual Quebec engagements, French Canadian content is cited materially less often than English Canadian content, even when content quality, schema parity, and link profile are roughly matched. The gap is real, observable, and slowly closing.

What we see in practice:

• fr-CA AI Overview trigger rates are lower than en-CA equivalents

  Identical-intent queries trigger AI Overviews less often when issued in French Canadian than in English Canadian. The gap is notable across most verticals we track.
• fr-CA citation source pools are smaller

  When AI engines do cite for fr-CA queries, the cited source pool is smaller and more concentrated — the same handful of high-authority French Canadian sources show up repeatedly.
• ChatGPT and Perplexity handle fr-CA content better than AI Overviews do

  Among the engines we track, AI Overviews has the largest fr-CA gap; ChatGPT and Perplexity narrow it somewhat.
• Auto-translated French content underperforms native French content meaningfully

  Programs that translate English content into French via machine translation see noticeably lower fr-CA citation rates than programs commissioning native French Canadian writing. The signal AI engines pick up on is real, even when surface quality looks similar.

Implication: programs targeting Quebec audiences need native French Canadian content investment as a baseline. Auto-translation is not a substitute, and the citation rate penalty for treating fr-CA as a translation problem rather than a content problem is large enough to outweigh the cost savings.

Finding 10: 'Information gain' is observable in citation patterns

Google's published "information gain" framing — the idea that pages providing unique value beyond what's already in the top-cited cluster are advantaged — appears to be observable in AI engine citation patterns as well, not just in classical ranking.

What we observe:

• Pages with original data, frameworks, or proprietary perspectives are cited more often

  Where the top-cited cluster on a query is converging on the same five sources saying the same five things, a sixth source bringing a genuinely novel data point or framework gets cited at higher rates than a sixth source paraphrasing the existing five.
• Rewritten or paraphrased content is becoming progressively less visible

  As the citation cluster on a query matures, AI engines appear to deduplicate paraphrased content more aggressively. Content that contributes nothing the existing cluster doesn't already have gets cited less and less over time.
• First-mover advantage on emergent topics compounds

  Original content on emergent topics earns disproportionate citation share early, and that share appears to be sticky as the topic matures and the citation cluster fills in.
• Synthesis still wins, even without original data

  Where original data isn't possible, synthesis that genuinely connects points the existing cluster doesn't connect well is also cited more often than pure rewriting. The threshold for 'information gain' is lower than 'must be original research' — but it's not zero.

Implication: editorial programs that lean heavily on rewriting competitor content are optimizing into invisibility. Original data, proprietary frameworks, expert perspectives, and synthesis that adds connection — all lift citation rate in observable ways. This is the structural argument for why agencies and in-house teams should invest in research, expert-author content, and original frameworks rather than treating SEO as a content-volume game.

What this means for the next 12 months

If the ten findings above hold — and we expect they will, with engine-specific evolution — the implications for SEO programs over the next 12 months sort into a fairly clear priority order:

• Schema audit and remediation as a 30-day priority

  The single highest-leverage tactical investment in most programs. Article + Author + sameAs + FAQPage + HowTo + Organization markup, audited rigorously and remediated where missing or thin.
• Author infrastructure as a baseline requirement

  Every editorial program needs author bio pages, credentials, sameAs Person markup, and an editorial review chain for YMYL content. The cost is small; the citation lift is consistent.
• Engine-specific resource allocation

  Stop treating engines as equal-weight. Heavy investment in AI Overviews and ChatGPT optimization, calibrated investment in Perplexity and Gemini, light investment in Copilot unless audience-relevant.
• Original research and information-gain content as a strategic priority

  The agencies and in-house teams that win citation share over the next 24 months will be the ones publishing original data, proprietary frameworks, and expert perspectives — not the ones publishing higher-volume rewritten content.
• Editorial cadence calibrated to engine freshness preference

  Quarterly refresh of priority content for Perplexity-exposed programs; vertical-specific refresh for AIO-exposed programs; annual refresh as a floor for everything.
• Brand investment as part of AI SEO, not separate from it

  Brand search volume feeds citation rate. Programs that ignore brand-building are optimizing one variable while ignoring an equally important one.
• Native French Canadian investment for Quebec-targeted programs

  Auto-translation is not a substitute. The fr-CA citation gap is real, and closing it requires native-language content investment.
• Tracking discipline, not tracking dashboards

  Read citation tracking dashboards as noisy signals at the trend level. Don't make weekly tactical decisions off single-run citation snapshots. The teams winning are the ones with a consistent measurement ritual, not the ones with the prettiest dashboards.

Limitations and what we'd study next

Honest research includes its own limitations. The ones we'd flag for readers of this study:

• Sample bias toward well-resourced SEO programs

  Our sample skews toward businesses that hire a senior agency. Findings about citation lift from schema, author markup, and information gain may overstate the marginal effect for sites that aren't already at a baseline of editorial and technical quality.
• Engine coverage gaps

  AI Overviews and ChatGPT are observed most thoroughly. Gemini and Copilot have less observation depth in our sample. Findings on those engines should be treated as more provisional than findings on AI Overviews and ChatGPT.
• Vertical concentration

  Our client base concentrates in Canadian B2B, professional services, ecommerce, and SaaS. Findings may generalize less well to consumer goods, news media, and entertainment categories where we have thinner observation.
• Engine non-determinism limits precision

  We've discussed this in finding 7, but it's worth restating as a methodological limitation: the underlying signal is noisier than tools and dashboards typically convey, and that noise affects the precision of any aggregate finding.
• Locale concentration

  Findings are Canadian-context anchored. Structural findings (schema, author markup, information gain) likely generalize globally; locale-specific findings (fr-CA gap, AIO trigger rates) are Canadian-specific.

What we'd study next, given more research capacity:

• Citation-to-click ratios

  We track citation appearance well; we don't yet have a robust framework for measuring how often citations actually drive downstream traffic. This is the most important measurement gap in the category.
• Voice and agentic search citation patterns

  Voice surfaces (Siri, Alexa, Google Assistant) and emergent agentic search experiences are under-studied. Sample sizes there are still too small in our work to publish findings, but it's the next frontier.
• Vertical-specific deep dives

  Finer-grained findings within specific verticals (legal, healthcare, ecommerce, SaaS) would be more actionable for vertical-specific programs than the synthesis we've published here. We'd publish those as follow-up studies.
• Multi-engine citation correlation

  How often does a page cited in AI Overviews also get cited in ChatGPT, Perplexity, and Gemini? We have observations on this but not yet enough rigor to publish; it's a candidate for the 2027 update of this study.

If you'd like to discuss this research, request a brand-specific AI citation audit, or explore working with us on an AI SEO program, request a free AI visibility audit. We deliver written, brand-specific findings in three business days. Or call (437) 900-3626 to talk to a senior strategist.

Citation: Toronto SEO, "The 2026 AI Citation Study: How AI Engines Actually Pick Their Sources" (April 2026). torontoseo.com/research/ai-citation-study-2026.

Study Methodology FAQ

The questions journalists, analysts, and other agencies have asked us about how this research was conducted.