Understanding the Gap Between AI Citations and Traditional Rankings

What changed between traditional rankings and AI citations?

The change is not that rankings stopped mattering. The change is that rankings are no longer the only visible result of a search. Google AI Overviews can answer a question before a user reaches the classic list of blue links, then attach a selected set of supporting sources to that answer. A page can therefore be prominent in ordinary search while absent from the answer a user reads.

A new study by Chiang-Yu Cheng calls this gap Citation-Ranking Divergence, or CRD. The research examines whether generative-search citations mirror conventional rankings or reallocate source visibility. Its answer is more useful than a simple yes or no: citations are partly anchored to ranking visibility, yet the generative layer does not faithfully reproduce the ranked list. It can narrow the source set or surface sources that are not highly placed in traditional results.

That distinction matters because a ranking is a position in a results page, while a citation is a source selected to support a generated answer. They overlap, but they are different outcomes. A search team that reports only positions may be describing one surface while missing the surface where an answer is formed.

Google makes the separation clear in its own documentation. AI Overviews and AI Mode surface supporting links, may use different models and techniques, and may show varying sets of responses and links. Google also says that a page needs to be indexed and eligible for a Search snippet to be eligible as a supporting link, but eligibility is not a promise of serving. The floor still matters. The result above that floor can differ.

This is why an AI citation strategy should not treat citations as a renamed ranking report. It should treat them as a separate visibility measure. Track the prompts that matter, record the sources that appear, and compare that evidence with your conventional search performance. The point is not to chase a hidden trick. The point is to see the customer-facing answer surface as it actually behaves.

1. Traditional rankings order eligible pages in a search result.
2. AI citations identify selected supporting sources within a generated response.
3. A page can be strong on one surface and weak on the other.

What did the two-wave audit find?

The audit found a durable gap rather than a one-time oddity. Cheng examined Google AI Overviews using 1,500 base queries collected in an initial wave and the same queries collected again in a second wave. The design produced 3,000 wave-specific observations across five language-region locales before eligibility filtering.

The before-and-after design is the citable asset here. Before, the audit established the relationship between sources cited in AI Overviews and sources visible in conventional rankings for 1,500 queries. After, it repeated those same queries in a second collection wave. The repeated measurement showed that citation-ranking divergence could be studied as a stable pattern, not dismissed as a single volatile snapshot.

The study reports recurring forms of divergence. One is narrowing: the AI answer draws on a smaller pool of sources than the ranking layer presents. Another is promotion: a source that is not prominently ranked appears among the answer's citations. Neither pattern means rankings have become irrelevant. Both mean a rank report alone cannot establish whether a brand is visible in the generated answer.

The study's language-region coverage also limits a common shortcut. An English-only prompt set cannot tell a business whether the same visibility pattern holds in other locales. Search, source availability, and the answer experience vary by query and market. If a company serves multiple regions, its measurement should reflect the questions and languages that buyers actually use.

The practical conclusion is straightforward: a citation report needs repeated observations. One prompt run can reveal an opportunity. It cannot prove a stable pattern. Measure a defined prompt set over time, preserve the answer and cited domains, and compare the result with organic visibility. That creates evidence for an editorial decision instead of a screenshot collection.

1. Initial wave: 1,500 base queries.
2. Second wave: the same query set collected again.
3. Combined design: 3,000 wave-specific observations across five language-region locales before eligibility filtering.
4. What to do: repeat a fixed prompt set and compare cited domains with ranked domains over time.

Who does citation-ranking divergence affect?

Citation-ranking divergence affects any business whose buyer asks an answer engine to choose, compare, explain, or recommend. That includes B2B SaaS teams competing for category queries, service businesses competing for local discovery, publishers whose reporting is used as support, and ecommerce brands answering product questions.

The immediate risk is a false sense of security. A team can see a page rank well and assume it is winning search visibility, even though the relevant AI answer cites competitors, a trade publication, a directory, or a first-party documentation page. The opposite can also happen: a lower-ranked page can appear as a cited source and receive attention the rank report did not predict.

This changes how editorial teams should define coverage. A page built only to target a keyword may be too narrow for the question variations an answer engine explores. Google's documentation says AI has can use query fan-out, which means related searches across subtopics and data sources may contribute to a response. A useful page must therefore answer the central question clearly and provide support that holds up when the question expands.

For publishers, the issue is distribution. Being indexed is necessary, but it is not the same as being selected as support. For service businesses, the issue is buyer intent. A prospect who asks for the best provider in a city may see a compact answer with a limited set of sources before considering a conventional result. For SaaS teams, the issue is category definition. If a competitor appears in citations for comparison prompts, it may shape the buyer's shortlist even where your domain ranks for related terms.

None of this supports a claim that AI citations can be hacked or guaranteed. It supports a stronger operating standard: measure the visibility that exists, publish reliable coverage for the questions that matter, and refresh it when the evidence changes.

1. B2B SaaS teams need visibility on category and comparison prompts.
2. Local and multi-location businesses need visibility on service-and-location prompts.
3. Publishers need to know whether their reporting is selected as answer support.
4. Any brand using rankings as its only search KPI has an incomplete view.

Why can a high-ranking page still miss an AI answer?

A high-ranking page can miss an AI answer because selection for a supporting link is not identical to placement in the ranked results. Google says its AI has may use different models and techniques, and that the responses and links they show can vary. A conventional ranking is therefore an input to visibility, not a receipt for citation.

The audit supplies a useful explanation without pretending to know every model decision. It found partial alignment with rankings alongside structured divergence. That means the broad web hierarchy still has influence, but citation allocation can become more concentrated or introduce different sources. A team should resist two bad conclusions: that SEO is finished, or that a top-three rank guarantees the answer surface.

Content shape can also affect whether a page is easy to use as support. A page that buries its answer beneath broad marketing copy creates more work for a reader and for a system looking for a clear claim. A page that gives a direct answer, names its sources, explains limitations, and keeps facts current gives itself a stronger editorial foundation. That is not a special optimization requirement. It is solid publishing.

Technical access still belongs in the checklist. Google says supporting-link eligibility requires a page to be indexed and eligible to appear with a snippet in Google Search. If a page is not crawlable, not indexed, blocked from snippets, or materially weak on the query, there is no citation strategy that bypasses that problem.

The right response is disciplined diagnosis. Check whether the page is technically eligible. Compare the page's answer with the query's real intent. Review the cited sources that win the answer. Then decide whether the missing coverage is a content gap, a source-trust gap, a freshness problem, or simply a prompt where your site is not yet the best support.

1. Confirm indexing and snippet eligibility first.
2. Read the cited pages, not only their domains.
3. Compare the cited answer with the promise and evidence on your page.
4. Refresh claims when source material or customer questions change.

How should an AI citation strategy respond?

An AI citation strategy should add citation measurement to SEO, not replace SEO with a separate ritual. Google says existing SEO best practices remain relevant for AI has and that no additional technical requirements are needed to appear in AI Overviews or AI Mode. The sensible response is to protect fundamentals while adding a direct measurement layer for cited visibility.

Start with a prompt universe tied to commercial reality. For a SaaS company, that may include category queries, implementation questions, alternatives, integrations, and comparison prompts. For a local business, include service-and-location questions, urgent need queries, qualification questions, and cost questions where factual support is available. Write each prompt as a real buyer would ask it, then keep the set stable enough to compare over time.

Next, record the answer surface. Capture whether an AI answer appeared, which domains were cited, which of your pages were cited, the language and locale, the date, and the prompt used. This lets a team calculate Citation Share: the percentage of relevant AI answers in a category that cite the brand. Citation Count per day measures volume. Answer Presence measures breadth across the prompt universe. Share of Voice compares you with named competitors.

Then use editorial work to close verified gaps. Publish the source-backed explanation your audience needs. Improve existing pages that already have topical authority but do not answer the question directly. Add comparison pages only where the comparison is fair and supportable. Build related pages so that a buyer can move from a broad question to a decision without finding contradictions.

Finally, report the two surfaces together. A strong weekly view includes organic visibility, Citation Share, Answer Presence, and the specific prompt groups where competitors appear. This avoids the vanity metric of a single favorable answer and makes progress visible when rankings and citations move in different directions.

1. Keep SEO fundamentals and technical eligibility intact.
2. Define a stable prompt universe from buyer questions.
3. Capture cited domains, page URLs, dates, locales, and answer presence.
4. Use Citation Share alongside rankings to assess visibility.
5. Prioritize pages that can close a documented question or evidence gap.

What should content teams publish differently?

Content teams should publish pages that earn use as support, not pages that merely repeat a target phrase. The most defensible format answers the question early, distinguishes fact from opinion, identifies the source behind every statistic, and explains where the answer does not apply. That gives a reader something they can verify and gives an answer engine a clearer source to cite.

A direct answer does not mean a thin answer. It means the opening resolves the question before the page expands into method, trade-offs, examples, and sources. A buyer asking whether rankings predict AI citations needs the answer first: not reliably. They then need the why, the audit design, the limitations, and an action plan. That sequence respects the question without hiding the evidence.

Original reporting and maintained reference material have an advantage because they can contribute facts that no generic summary has. If you publish an original dataset, explain the collection method, date range, definitions, exclusions, and update cadence. If you publish a comparison, show the criteria, disclose unknowns, and link to first-party materials. If you publish a how-to, make every step testable.

The same rule applies to freshness. Do not refresh a date without checking the claims beneath it. Review changes in product documentation, regulations, pricing, source data, and terminology. Remove claims that can no longer be supported. A citation is a trust signal only when the cited page still deserves trust.

Editorial scale can help coverage, but volume cannot replace evidence. Publishing more pages that say the same thing creates clutter. Publishing clear, distinct, sourced answers across the questions a market actually asks creates a library that can be found, read, and cited.

1. Put the direct answer before the explanation.
2. Use dated primary sources for factual claims.
3. State methodology for original data and comparisons.
4. Review factual pages for substantive freshness.
5. Create distinct coverage instead of keyword variations with the same answer.

How should leaders report the gap to the business?

Leaders should report the gap as a visibility measurement problem, not as a claim that every AI answer must cite the brand. The business question is whether relevant answers cite the company often enough to remain considered. The answer comes from a defined sample, consistent collection, and comparison with competitors.

A useful dashboard separates volume from breadth. Citation Count per day shows how often the brand is cited. Answer Presence shows how widely it appears across the chosen question universe. Citation Share shows the proportion of relevant answers that cite the brand. Share of Voice reveals whether named competitors dominate the same prompts. Together, these measures clarify whether a brand has a handful of citations or genuine category visibility.

Pair those measures with conventional search indicators. A page that rises in rankings but remains absent from cited answers needs a different investigation from a page that gains citations without a rank increase. The first may need a better answer format, stronger evidence, or broader topical coverage. The second may be useful evidence that the brand is reaching buyers through a surface the rank report undervalued.

Do not turn a limited observation into a causal promise. AI answer outputs vary. Google says the links and responses in its AI has can vary, and its systems decide when AI Overviews are additive to classic Search. Report confidence honestly, retain the source material, and compare like with like.

The audit's core lesson is blunt: there is no page two in an AI answer. If the answer chooses a short set of sources, being merely present somewhere in the rankings may not be enough. The remedy is not manipulation. It is better measurement, durable editorial coverage, and a clear view of whether your work is actually cited.

1. Report Citation Share, Citation Count per day, Answer Presence, and Share of Voice separately.
2. Compare citation results with organic performance for the same prompt groups.
3. Preserve dates, locales, prompts, cited URLs, and competitor evidence.
4. Describe variation and limitations instead of promising a citation outcome.

Sources

Source: Springer Nature, Information Systems Frontiers

The peer-reviewed study introduces Citation-Ranking Divergence and examines how Google AI Overview citations differ from traditional rankings. Springer Nature, Information Systems Frontiers, 2026-09-10

Source: Scienmag

The news report describes the study's two-wave audit design, including 1,500 base queries, 3,000 wave-specific observations, and five language-region locales before eligibility filtering. Scienmag, 2026-09-12

Source: Google Search Central

Google states that existing SEO best practices remain relevant for AI features, that no additional technical requirements are needed, and that AI has responses and links can vary. Google Search Central, 2025-12-10