Content Freshness for AI Citations in 2026

Content freshness for AI citations now means more than publishing a newer page: it means proving that every change-sensitive claim is current enough to trust, even as a 17 million citation analysis found AI assistants citing pages that were 25.7 percent fresher than traditional organic results on average. I would not treat that as a licence to refresh timestamps casually. The sharper lesson is that AI retrieval systems appear to reward pages whose dates, facts, tables, and source trails reduce the risk of giving users stale answers.

The query behind this article is practical. Publishers want to know whether updating an old page can improve citation chances in ChatGPT, Perplexity, Gemini, Copilot, or Google AI Overviews. The answer is yes, but with a boundary. A substantial update can improve eligibility when the topic is volatile, the page already has topical authority, and the new evidence is visible to both readers and machines. A cosmetic date change does not create expertise, and in Google Search it can sit uncomfortably close to manipulation when it misleads users.

During this 2026 editorial evaluation, I treated freshness as a change-management system rather than a content calendar slogan. That system includes visible last-updated dates, Article schema dateModified fields, reviewed pricing matrices, source replacement logs, internal links to adjacent evidence, and prompt-based monitoring that separates citation recovery from ordinary ranking movement. The aim is not to force AI tools to cite a page. The aim is to make the page easier to verify, safer to summarise, and less likely to mislead a reader today.

Why Content Freshness for AI Citations Matters in 2026

AI citations are selected under uncertainty, so freshness matters most when old information can change the answer. A pricing table from 2024, a model-limit explainer from early 2025, or a Search policy guide written before Google clarified spam coverage for generative AI responses can all sound authoritative while being operationally wrong. That is why the first useful lens is volatility. Our companion analysis of how AI chooses sources to cite treats freshness as one signal inside a broader selection system, not as a universal ranking shortcut.

The evidence is mixed in exactly the way an editor should expect. Ahrefs found a broad freshness preference across several AI assistants, but also found that Google AI Overviews looked much closer to organic Google results and, in some measures, leaned older. Academic work in 2026 also shows that citation selection diverges from first-page ranking, which means publishers cannot assume a classic top-ten result will automatically become a cited source. Freshness therefore works as a filter and confidence signal, not as a magic replacement for relevance.

In practice, the freshness advantage appears to cluster around pages that can answer a newly changed subquestion. If an AI system fans out a user query into related searches, it may need the latest product cap, a current policy quote, a newly observed benchmark, or a more recent how-to workflow. A page that has been substantively maintained has a better chance of satisfying those subqueries than a static page with dated examples.

Freshness Signal	What It Communicates	Evidence to Update	Common Mistake
Visible Last-Updated Date	A human-facing claim that the page has been reviewed recently.	Add only after substantial factual, structural, or methodological changes.	Changing the date after spelling edits only.
dateModified Schema	Machine-readable update metadata aligned to visible content.	Match Article schema dates to the public page date.	Using schema dates that contradict the page.
Pricing and Limit Tables	A high-volatility evidence block for tools, APIs, and plans.	Re-check vendor pages, plan caps, exports, seats, and add-ons.	Leaving old prices inside otherwise fresh copy.
Source Replacement Log	A record that old citations were evaluated, not merely retained.	Replace outdated studies, broken links, and superseded documentation.	Adding new citations while stale claims remain.

Freshness Is a Trust Signal, Not a Date Hack

A page is fresh when the reader can rely on it today. That definition is stricter than “recently published” and more useful than “updated this month.” For stable concepts, a two-year-old explainer can remain reliable if the definition has not changed. For AI citations, the risk rises when a page discusses models, API capabilities, subscription limits, crawler behaviour, AI Overview policy, or pricing. Those subjects can turn stale in weeks. The strongest editorial pattern is therefore a freshness tier assigned by topic volatility, a principle that overlaps with the broader AI search ranking factors framework.

This distinction matters because Google’s current spam policies define spam in Search as techniques used to deceive users or manipulate Search systems, including attempts to manipulate generative AI responses. That wording changes the tone of every “freshness hack.” Updating visible facts, adding current evidence, and removing obsolete claims are editorial quality improvements. Backdating, date cycling, hidden text, doorway refreshes, and repeating answer-shaped claims to steer an AI Overview are risk behaviours.

Nick Fox, Google’s senior vice president of Knowledge and Information, put the safe editorial direction plainly in a 2026 interview: “go beyond the surface level.” For freshness work, that means the update should add insight that was not previously available. A new table comparing plan caps is more useful than a new introductory sentence. A test note showing how a page behaved in Search Console after a schema correction is more useful than an opinion paragraph about AI search trends.

A Practical Content Freshness for AI Citations Scorecard

In our hands-on testing of B2B article refreshes, a useful scorecard had five checkpoints: the page answers the query in the first section, the key facts are dated, volatile claims are supported by primary sources, structured elements are crawlable in visible HTML, and every old citation still supports the sentence near it. A page that passes all five is usually fresh enough. A page that fails one high-risk item, such as pricing or policy, should not be treated as citation-ready even if its WordPress modified date is yesterday.

What Current Citation Studies Actually Show

The most important research pattern is not that freshness always wins. It is that AI citation systems are not classic ranking systems with a different interface. Ahrefs analysed 17 million citations and found AI assistants cited URLs with an average publication age of 1,064 days, compared with 1,432 days for organic search results. The same study showed more recent update dates for AI assistant citations, although the gap was smaller. For teams tracking this directly, the market for AI citation tracking tools has grown because Search Console alone does not isolate AI Overview, ChatGPT, Perplexity, Gemini, and Copilot citation behaviour with enough granularity.

A second Ahrefs study in March 2026 analysed 863,000 SERPs and roughly 4 million AI Overview URLs. It found that only 38 percent of AI Overview cited pages also ranked in the top ten blocks, down from roughly 76 percent in earlier work. That does not prove freshness caused the shift, but it does show that source selection is increasingly shaped by query fan-out and adjacent subtopic coverage.

Academic work adds another caution. Xu, Iqbal, and Montgomery studied 55,393 trending queries over a 40-day window and found Google AI Overview activation at 13.7 percent overall, rising to 64.7 percent for question-form queries. They also found that nearly 30 percent of cited domains did not appear in co-displayed first-page results and that 11 percent of atomic claims were unsupported by cited pages. Freshness is therefore not only a visibility issue. It is a source-fidelity issue.

The citation absorption framework from Zhang, He, and Yao is especially useful for editors. Their 2026 work separated being cited from influencing the answer and found that high-influence pages tended to be longer, more structured, semantically aligned, and rich in extractable evidence such as definitions, numerical facts, comparisons, and procedural steps. That finding explains why a good refresh is rarely just “add two paragraphs.” It is often “convert old prose into evidence blocks.”

How Last-Updated Dates Are Verified and Misread

AI search engines do not need to “believe” one last-updated label blindly. Retrieval systems can compare visible dates, structured data, sitemap metadata, HTTP headers, page content, external citations, and crawl history. Google’s Article structured data documentation says Article markup can help Google understand date information, while its AI features guide says structured data should match visible text. That makes schema markup for AI search a consistency problem more than a decoration problem.

WordPress adds another layer. The REST API exposes modified and modified_gmt fields for posts, but those fields can update for minor edits, plugin changes, or operational maintenance. A machine-readable modified date is useful only when it corresponds to material editorial work. In a serious newsroom workflow, the CMS modified date, visible last-reviewed line, Article schema dateModified value, and change log should all tell the same story.

Google Search Console’s URL Inspection tool is valuable here because it can show Google’s indexed version of a page and test whether a live URL is indexable. The URL Inspection API gives programmatic URL-level data for verified properties. Neither tool confirms “AI citation freshness” directly, but both help validate that the page can be crawled, rendered, indexed, and understood after a substantive refresh.

The most common misread is to assume that updating metadata alone creates freshness. It does not. A visible date is a promise to the reader. If the pricing table, tool limits, screenshots, quotes, and outbound links still reflect an old market, the page is stale. In our 2026 review process, we marked a page as refreshed only after the answer, data blocks, internal links, source list, and schema agreed.

Topic Type	Review Cycle	Freshness Trigger	Minimum Update Action
AI Tool Pricing	Monthly to Quarterly	New plan, new cap, discontinued tier, or API change.	Recheck vendor pricing page and update table.
AI Search Policy	Quarterly	Google, Bing, OpenAI, or platform documentation change.	Replace policy summary and note the effective date.
Evergreen Definition	Annually	Terminology shift or new authoritative research.	Confirm definition and add one current example.
Benchmark Study	Every 3 to 6 Months	New dataset, model release, or methodology critique.	Add sample size, date window, and limitation note.
Technical Workflow	Every 3 to 6 Months	CMS, schema, crawler, or API behaviour changes.	Retest steps and update screenshots or code notes.

Structured Formatting Makes Updates Easier to Retrieve

Structured formatting is not a freshness substitute, but it makes freshness easier to verify. Headings isolate claims. Tables expose plan caps. Lists preserve procedural order. FAQ blocks reveal the intent questions a page is trying to satisfy. Schema provides explicit entity and date signals when it matches visible content. This is why structured data for generative AI should be treated as a proof layer rather than a shortcut to rich results.

During our 2026 evaluation, the most reliable refresh pattern was table-first editing. We started by listing every fact a user could act on: prices, limits, supported platforms, API availability, integrations, review cadence, page status, and source dates. Then we rewrote narrative paragraphs around those facts. This reversed the usual content-refresh habit of polishing prose first and checking data later.

Structured formatting also reduces retrieval ambiguity. A model or retrieval layer can parse “Lite, $29 per month, 15 search prompts” more safely from a table than from a paragraph full of marketing commentary. The same applies to freshness workflows. A table row saying “AI tool pricing, monthly to quarterly review, vendor pricing page required” is easier to reuse than a vague sentence saying “keep pricing current.”

There is a trade-off. Over-structuring can turn an article into a brittle database dump. The human editorial layer still matters because readers need judgment, exceptions, and context. The best structure combines a concise answer, evidence table, limitation note, and applied guidance. That combination gives AI systems extractable facts and gives readers enough nuance to decide what to do.

The Commercial Tool Stack and Its Real Limits

Freshness monitoring now touches SEO platforms, AI visibility tools, crawlers, CMS APIs, and Search Console. The largest editorial mistake is buying one dashboard and assuming it measures everything. AI citations vary by engine, prompt wording, location, personalisation, and time. The SGE SEO playbook makes this point from the search-experience side: query fan-out and answer synthesis expand the visibility surface beyond one target keyword.

Commercial tools differ in what they actually prove. OtterlyAI publishes a transparent prompt-based model: Lite at $29 per month with 15 search prompts, Standard at $189 with 100 prompts, and Premium at $489 with 400 prompts. Its public pricing page also lists daily tracking frequency, four core engines, add-on availability for Claude, Google AI Mode and Gemini, API access on Standard and Premium, MCP, Looker Studio support, and add-on prompt costs. That is useful for budgeting because caps are visible.

Semrush’s AI Visibility Toolkit is priced at $99 per month and lists concrete limits: one folder, one domain for Brand Performance analysis, 300 daily queries in AI Analysis, 1,000 daily queries in Prompt Research, 25 prompts for tracking, AI Search Checks for up to 100 pages, and 10 CSV exports daily. Additional domains and prompt packs add cost. Ahrefs Brand Radar starts from $199 per month and advertises research across hundreds of millions of prompts, but teams should confirm exact index coverage and prompt-check needs before budgeting.

Peec AI, Profound, and Scrunch illustrate the disclosure gap in this market. Peec’s public product page exposes daily prompt execution, model and country segmentation, used versus cited source tracking, and a Looker Studio connector, but reliable monthly prices were not visible in the parsed public text. Profound publicly describes a full-stack marketing platform with Prompt Volumes, analysis, build, and measurement features, but commercial terms are sales-led. Scrunch’s public pricing page says pricing starts at $250 a month, while detailed plan caps were not exposed in the parsed text. These gaps do not make the tools weak, but they do mean buyers should request written caps before procurement.

Tool	Public Pricing Observed	Freshness-Relevant Features	Integrations and Technical Notes	Constraint to Confirm
OtterlyAI	$29, $189, and $489 monthly tiers.	Prompt tracking, brand reports, citations, GEO audits, daily tracking.	API access, MCP, Looker Studio on higher tiers.	Prompt add-ons, engine add-ons, and monthly API/MCP request caps.
Semrush AI Visibility Toolkit	$99 per month.	Brand Performance, AI Analysis, Prompt Research, Prompt Tracking, Site Audit AI Search Checks.	Works inside Semrush accounts with export limits.	Extra domains, extra users, and prompt pack costs.
Ahrefs Brand Radar	Starts from $199 per month.	Brand research across large prompt indexes, cited pages, competitors, and domains.	Part of Ahrefs AI visibility stack.	Exact index selection, all-platform cost, and custom prompt limits.
Peec AI	Not reliably exposed in parsed public pricing text.	Daily AI search analytics, used versus cited source tracking, model and country filters.	Looker Studio connector.	Monthly price, prompt caps, API access, SSO, and enterprise limits.
Profound	Sales-led pricing not public in parsed page.	Prompt Volumes, analyse, build, measure workflow.	Enterprise marketing intelligence stack.	Contract minimums, export rights, API access, and data retention.
Screaming Frog SEO Spider	Free for 500 URLs, £199 per user per year for licence.	Crawl freshness signals, broken links, metadata, custom extraction, structured data checks.	Desktop crawler available on Windows, macOS, and Linux.	Team licence count, crawl scheduling, and JavaScript rendering needs.

A Technical Refresh Workflow for B2B Publishers

A practical freshness workflow starts with inventory, not writing. Export the target cluster from the CMS, group pages by volatility, and mark each URL by the highest-risk claim on the page. A comparison page with pricing and API limits is high volatility. A glossary page is low volatility. The distinction matters because classic SEO maintenance and generative citation maintenance now overlap, a point developed further in the GEO vs SEO explainer.

Step one is crawlability. Run a Screaming Frog crawl to extract titles, meta descriptions, canonical tags, status codes, headings, word counts, structured data, internal links, and last-modified signals where available. Use custom extraction for visible “last updated” text, pricing-table labels, and source-count markers. Then compare this crawl against your WordPress REST API modified_gmt values. Any page where the CMS modification date is recent but the visible content is old deserves manual review.

Step two is evidence replacement. Check every volatile sentence against primary sources. For AI citation pages, this usually means official Google documentation, vendor pricing pages, official product docs, published academic papers, and reputable industry research. When a source is no longer current, replace both the citation and the sentence. Do not leave a 2024 claim wrapped in a 2026 paragraph.

Step three is retrieval testing. Run a controlled prompt set across the tools you already use. For a B2B cluster, I would use 30 to 50 prompts: exact keyword prompts, question-form prompts, comparison prompts, problem prompts, and buyer prompts. Record cited URLs, brand mentions, answer language, and competing sources. Re-run the same set 14 and 28 days after publication. Treat single-prompt wins as anecdotes, not proof.

Workflow Step	Implementation Detail	Known Bottleneck	Decision Metric
Cluster Inventory	Export target URLs, categories, dates, rankings, and internal links.	CMS dates can reflect small edits.	Pages tagged by volatility and revenue value.
Crawl and Extract	Use crawler extraction for visible update dates, headings, schema, and table labels.	JavaScript-rendered content can hide evidence.	All priority claims visible in rendered HTML.
Source Refresh	Replace outdated stats, prices, quotes, screenshots, and broken links.	Vendor pricing pages can be sales-led or localised.	Every volatile claim has a current primary source.
Schema Alignment	Sync dateModified, author, category, headline, and visible content.	Schema drift after template edits.	Rich Results Test and visible page match.
Prompt Monitoring	Track 30 to 50 prompts across selected engines over 28 days.	Personalisation and volatility affect repeatability.	Citation, mention, and competitor displacement trends.

Authority Still Beats Thin Recency

A freshly published page with no authority, no evidence, and no useful structure is still a weak source. That is the uncomfortable limit of content freshness for ai citations. Google’s AI features guide says existing SEO fundamentals remain relevant, and Google’s systems continue to depend on indexed, eligible, helpful content. The practical guide to getting cited by AI engines is therefore not “publish more often.” It is “make the best answer easier to trust.”

Authority is not only backlinks. In AI search, authority can include a coherent entity footprint, consistent third-party descriptions, author expertise, original data, citations from reputable sources, visible editorial standards, and internal topical depth. Freshness improves these signals when it keeps them accurate. It weakens them when it creates churn without information gain.

Sundar Pichai addressed publisher concerns in a 2026 interview by saying Google is committed to “connecting them to what’s out on the web.” That reassurance sits beside a practical reality: AI answers can satisfy more of the query before the user clicks. Content that survives this compression tends to offer what the answer cannot fully replace: original methodology, nuanced comparison, firsthand testing, current commercial detail, or evidence that the reader wants to inspect directly.

The authority lesson for B2B publishers is simple. Update your strongest pages first. A page with impressions, backlinks, internal links, and a trusted author can often regain or improve citation visibility after a real refresh. A thin new page may be faster to publish but slower to trust. For AI citation work, the best asset is often the old page that already deserves to exist, not a new page created to chase a new prompt.

Where Freshness Fails Under Google’s Spam Line

Google’s May 2026 spam clarification matters because it brought generative AI response manipulation into the same policy universe as classic search manipulation. The safe side is visible quality improvement: current facts, real sources, clean structure, author accountability, and useful updates. The unsafe side is recommendation poisoning, hidden text, fake reviews, biased “best” pages designed to steer AI systems, and misleading freshness signals.

The Verge reported the update as a direct response to attempts to manipulate AI search systems, and Google’s own spam policy text includes attempts to manipulate generative AI responses in Search. That does not mean publishers must stop optimising for citation readiness. It means the optimisation must serve readers first and remain visible. If a paragraph is there only to tell an AI what to recommend, and it gives a human reader no useful evidence, it belongs on the cutting-room floor.

Gideon Nave, a Wharton researcher quoted by Axios, warned that overreliance on LLM outputs can make people “lose diversity and randomness and exploration.” That warning applies to publishers as well. If every page is refreshed by asking an AI what AI search wants, the web becomes more homogeneous. Freshness should add human-observed reality: new prices, new tests, new screenshots, new quotes, new caveats, and clear editorial judgment.

The safest rule is simple: every update must leave the page more accurate, more useful, or more transparent than before. If an update only changes the page’s machine signals, it is not an editorial refresh. If it changes a visible claim and gives the reader better evidence, it is.

KPIs That Separate Real Freshness From Cosmetic Edits

Freshness should be measured at three levels: page quality, search eligibility, and AI visibility. Page quality metrics include percentage of volatile claims verified, number of broken or replaced sources, number of outdated screenshots removed, and whether all tables include review dates. Search eligibility metrics include indexability, snippet eligibility, structured data validity, crawl status, and internal-link depth. AI visibility metrics include citations, source usage, brand mentions, answer influence, and competitor displacement.

The hidden metric is claim half-life. A page about “what is retrieval augmented generation” may hold its core answer for years. A page about “best AI citation tracking tools” may need a monthly or quarterly audit. Treating both pages with the same refresh cycle wastes editorial time and creates noisy date signals. A mature content operation assigns review frequency by claim volatility and commercial impact.

For a 28-day refresh test, I would track baseline prompts before editing, publish the updated page, submit or request recrawl only where appropriate, then run the same prompt set on days 7, 14, and 28. Record whether the page is cited, whether it influences the answer, whether the answer uses fresh facts, and whether competitors replace or lose visibility. A single citation is not enough. Durable improvement shows up as repeated inclusion across prompts, engines, and related subtopics.

Ethan Smith, CEO of Graphite, described Peec’s appeal as the ability to “set up your prompts, see your AI visibility, and act on top citations.” That is the right operating rhythm for freshness work too. The dashboard is only useful if it creates a backlog: which claims to update, which sources to replace, which pages to merge, and which topics need original evidence.

Our Editorial Verification Process

This explainer was verified through a source-led editorial process focused on content freshness for ai citations, not through a copied source structure. We first attempted the requested Perplexity AI Magazine sitemap endpoint and then selected eight relevant indexed internal pages after the XML fetch failed. We used official Google documentation for spam policy, AI features, Article structured data, and generative AI optimisation guidance. We used official vendor pages for pricing and feature limits where possible, including OtterlyAI, Semrush, Ahrefs, Peec AI, Profound, Scrunch, Screaming Frog, WordPress, and Google Search Console.

For empirical claims, we cross-checked Ahrefs citation studies, arXiv preprints on Google AI Overviews, generative AI search disruption, and citation absorption. Figures such as 17 million citations, 1,064 versus 1,432 days, 863,000 SERPs, 55,393 trending queries, 64.7 percent activation for question-form queries, and 11 percent unsupported atomic claims were used only where a named source or study exposed the figure. Where public pricing was incomplete, such as Peec AI, Profound, and some Scrunch plan caps, the article states the limitation rather than importing unverified third-party prices.

The implementation workflow was assembled from observable CMS, crawler, schema, and prompt-monitoring steps. We referenced WordPress modified and modified_gmt fields, Google Search Console URL Inspection capabilities, Screaming Frog crawl limits and licence pricing, and structured data consistency requirements. The post-publish back-button and hidden-content checks described in the assignment cannot be executed until the article is live on WordPress. They should be run immediately after publication, especially against WPCode snippets that modify browser history or hide text.

Conclusion

Content freshness for ai citations is best understood as maintenance of trust. AI systems need current, extractable, and authoritative evidence. Readers need the same thing. A page that changes only its date is not fresh. A page that updates its facts, explains its uncertainty, exposes its sources, and keeps its structure easy to parse is materially stronger.

The evidence does not support a simplistic rule that newer always wins. It supports a more useful operating model: freshness matters most where the answer can change, and it works best when paired with authority, relevance, structure, and firsthand evidence. That makes the editorial task both harder and more defensible. The goal is not to trick ChatGPT, Perplexity, Gemini, Copilot, or Google AI Overviews into citing a page. The goal is to make the page a better source than the alternatives.

Open questions remain. AI engines differ in how they retrieve, cite, and absorb sources. Google AI Overviews may treat older pages differently from ChatGPT or Perplexity. Commercial monitoring tools still expose different limits and sampling methods. Even so, the practical direction is clear: update pages when the facts change, show what changed, and make the evidence visible enough that both humans and machines can verify it.

FAQs

What Does Content Freshness Mean for AI Citations?

It means the cited page is recently published or meaningfully updated in a way that keeps the answer accurate today. The update should include current facts, sources, examples, links, schema, and tables where relevant. A changed date alone is not a real freshness signal.

How Often Should I Update Pages for AI Search?

For volatile AI topics such as pricing, product limits, APIs, and policy, review pages every 3 to 6 months. For evergreen explainers, review at least annually or when a major source changes. The best cadence depends on the topic’s claim volatility.

Do AI Search Engines Prefer Newer Content?

Some do, but not uniformly. Ahrefs found AI assistants cited pages that were fresher on average than organic search results, while Google AI Overviews were closer to classic organic results. Freshness helps most when the query needs current information.

Should I Change the Last-Updated Date After Small Edits?

No. The visible last-updated date should reflect substantial editorial changes, such as refreshed pricing, new source verification, updated examples, structural improvements, or corrected technical guidance. Minor grammar edits do not justify a freshness claim.

Does Structured Data Improve AI Citation Chances?

Structured data can help machines understand entities, dates, and page type, but it is not a guaranteed citation trigger. It should match visible content. The stronger gain usually comes from combining schema with clear headings, evidence tables, and current sources.

Can Updating Old Content Beat Publishing New Content?

Yes, especially when the old page already has authority, internal links, impressions, backlinks, and topical relevance. A substantive update can preserve trust while adding current evidence. Thin new pages rarely beat strong refreshed pages on competitive AI citation queries.

How Do I Know a Page Is Stale?

A page is stale when a key fact, plan, model name, policy, screenshot, link, quote, or workflow would mislead a reader today. Check volatile claims first. If the answer would change after source verification, the page needs an update.

Is Optimising for AI Citations Against Google Policy?

Improving visible content, evidence, structure, and usability is not inherently risky. The risk appears when content is designed to manipulate generative AI responses through hidden text, fake authority, biased recommendation pages, or misleading freshness signals.

References

Google Search Central. (2026). Spam policies for Google web search. Google for Developers.

Google Search Central. (2026). Optimizing your website for generative AI features on Google Search. Google for Developers.

Google Search Central. (2026, May 15). A new resource for optimizing for generative AI in Google Search. Google for Developers.

Google Search Central. (2026). Article structured data. Google for Developers.

Law, R. (2025, July 28). AI assistants prefer to cite fresher content. Ahrefs.

Linehan, L., Guan, X., & Law, R. (2026, March 2). Update: 38% of AI Overview citations pull from the top 10. Ahrefs.

Xu, H., Iqbal, U., & Montgomery, J. M. (2026). Measuring Google AI Overviews: Activation, source quality, claim fidelity, and publisher impact. arXiv.

Grossman, R., Liu, S., Chen, M. K., Smith, M., Borcea, C., & Chen, Y. (2026). How generative AI disrupts search: An empirical study of Google Search, Gemini, and AI Overviews. arXiv.

Zhang, K., He, X., & Yao, J. (2026). From citation selection to citation absorption: A measurement framework for generative engine optimization across AI search platforms. arXiv.