The Future of AEO: What's Documented, What's Google's Plan, and What's Still Unresolved

This page isn't a forecast. Nobody has a verified model of where AI search is headed, and a page that pretended otherwise would fail the same evidence standard as everything else in this cluster. What follows sorts the available evidence into three honest buckets instead: things that are already happening and measured (a real, dated trend, not a guess), things a platform has explicitly said it's building (its own stated direction, not yet independently measured for citation impact), and things repeated often enough to sound settled that the evidence doesn't actually support (still unresolved, or contradicted outright). Keeping those three separate is the point of this page.

Documented: A Substantial Citation-Ranking Gap - Not Yet a Proven Widening Trend

This is the strongest, most concrete finding on this page, stated carefully. In July 2025, Ahrefs found that roughly 76% of URLs cited in Google's AI Overviews also appeared in the top 10 organic results for that query. By January 2026 - across 863,000 SERPs and 4 million AI Overview URLs - that same figure had fallen to 37.9%. That's a real, dated before-and-after comparison, but Ahrefs itself doesn't treat it as a clean trend line: the company discloses it also improved its citation-parsing methodology between the two studies, and its own explanation for the drop is hedged, not definitive - Ahrefs writes that the change "may be" partly explained by Gemini 3 leaning more on fan-out queries, and that it "could be" pulling more citations from related sub-query results, not that it has isolated that cause. Treat the 76%-to-38% comparison as a real before-and-after data point complicated by a methodology change, not as proof of a steadily widening trend. A separate, independent 2026 study (55,393 trending queries over 40 days) found a comparable pattern using a different measure: nearly 30% of AI Overview citations didn't appear anywhere in the co-displayed organic results at all - corroborating a substantial gap using its own methodology, not the same one Ahrefs used. Authoritas adds a related but distinct finding: the pages that do get cited in AI Overviews are unstable on their own terms, with roughly 70% changing over a 2-3 month tracking window and almost no correlation (0.09) to organic-ranking changes over the same period - that's evidence of volatility and weak correlation, not evidence that the citation-ranking gap itself is widening over time. Taken together, the honest summary is: a substantial disconnect between organic rank and AI citation is documented across multiple independently-run studies using different methods; whether that gap is steadily widening, rather than just consistently large, isn't established by any single comparison here.

Ahrefs ran a related, separate study in mid-2025 comparing citations across four AI assistants against Google's top 10 - and this is worth stating precisely, because an earlier draft of this page got the arithmetic wrong. Across five citation/reference series (Perplexity, two separate ChatGPT series for in-text citations and references, Gemini, and Copilot), the mean overlap with Google's top 10 was 11.9%, calculated as the average of all five values - including Perplexity's, not excluding it. Perplexity's 28.6% is the clear outlier in that average; the other four series ranged from 6.1% to 8.6%. Read correctly, that's still a real and useful finding - most AI-cited URLs across these tools don't rank in Google's top 10 for the same query, and Perplexity behaves meaningfully differently from the other three - but the 28.6% figure is part of the 11.9% average, not a number excluded from it. Roughly 80% of citations in that same dataset didn't rank anywhere in Google for the query at all. This is also the concrete evidence behind the standing caution that no single optimization recipe works identically across engines: the data justifies measuring Google AI Overviews, ChatGPT, Perplexity, and Gemini/Copilot separately and testing whether a given tactic actually transfers between them - it doesn't yet establish that a per-engine optimization program reliably beats a shared one, since no study here assigns sites to one approach or the other and compares outcomes.

Google Says Fan-Out Is Expanding; Semrush Observes a Broader AIO Query Mix

These are two different kinds of evidence, and worth keeping separate rather than blending into one "documented" claim. Google has stated directly that Search with Gemini 3 performs "more searches to uncover relevant web content" because the system "more intelligently understands your intent" - that's Google's own first-party description of expanded query fan-out, not an independent measurement of how much fan-out actually increased. Separately, and independently, Semrush's tracking of 10 million-plus keywords from January through November 2025 is a genuinely third-party measured trend: it shows AI Overviews moving well beyond their original informational-query base, with the share of AIO-triggering queries that were purely informational falling from 91.3% in January to 57.1% by October, while commercial queries rose from 8.15% to 18.57% and transactional from 1.98% to 13.94%. Navigational queries rose sharply too, from under 1% of AIO-triggering queries in January to 10.33% by October - worth flagging that Semrush's own page states that January starting point two different ways in two different places (0.84% in its key-takeaways summary, 0.74% in the body), a small but real internal inconsistency that doesn't change the direction of the trend.

Worth noting precisely, because it cuts against a tidy "AI Overviews are strangling clicks" absolute: the same Semrush dataset found the zero-click rate for keywords that gained an AI Overview actually fell slightly over the tracked period (33.75% to 31.53%), not rose. That's a real observational finding, but it's observational - Semrush compared keyword-level behavior before and after AIO appearance without controlling for the query mix shifting at the same time, so it can't settle the causal question on its own. A more direct test exists and points the other way: a preregistered field experiment (N=1,100) that randomly assigned users to search with or without AI Overviews and AI Mode found that removing those features increased click-through to publisher sites. The two findings aren't necessarily in conflict - they use different designs, different denominators, and answer different questions - but the field experiment is the stronger design for a causal claim, and Semrush's result should be read as an observational data point, not a causal estimate, on its own. And AIO trigger rate itself wasn't a straight climb - it surged to 24.61% of queries in July 2025 before settling back to 15.69% by November, which Semrush reads as Google testing expansion and then recalibrating scope rather than an uninterrupted rollout. The safest reading of all of this together: broader query-type coverage is a documented Semrush measurement, expanded fan-out is Google's own stated direction, and neither one is a straight line - re-check these percentages periodically rather than treating any single month's snapshot as the new steady state.

Stated Direction, Not Yet Measured: Agents

At Google I/O 2026, Google described "information agents" that run in the background to monitor blogs, news sites, social posts, and real-time data sources for changes related to a user's stated interest, with a stated rollout beginning with Google AI Pro and Ultra subscribers "this summer," alongside expanded agentic booking capabilities on the same announced timeline. That's a real, specific, dated product announcement - not a rumor. It's not, on its own, evidence of an observed launch or of any citation-behavior finding: this page anchors the claim to what Google announced, not to an independently confirmed rollout status, since no source used here documents the feature's actual live availability or measures how content gets selected, weighted, or cited by it. No study in this cluster's source library measures that at all, for the simple reason that an agent monitoring sources continuously is a different surface than a single-prompt AI answer, and nothing reviewed here tests the continuous-monitoring case specifically. This connects directly to the field's own stated limits on measuring anything else here: a July 2026 single-author arXiv preprint reviewing 45 GEO studies - not a peer-reviewed systematic review, and its conclusion bounded to that reviewed corpus - found that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior," even for the existing single-prompt AI answers that have actually been studied. For what agentic browsing specifically means for citation strategy, that's covered in full on CoreAEX's dedicated agentic-browsers page rather than repeated here.

Stated Direction, Not Yet Measured: Multimodal Content

Google's Gemini 3 announcement describes "unparalleled multimodal understanding" and says that capability is being used to build more bespoke generative interfaces in Search. That's a real, confirmed statement about model capability and interface direction - it is not evidence that adding images or video to a specific publisher page raises that page's odds of being cited. What isn't established: a controlled measurement showing that adding images, video, or other non-text assets to a specific page increases that page's odds of being cited by an AI engine. No source in this cluster's library isolates that effect, and the general pattern documented elsewhere in this cluster - format-only changes tend to show weak or no isolated effect once you control for topical relevance - is a reasonable prior to apply here too, absent a study that tests it directly. The practical, honest version of this guidance: maintain accurate, well-described visual and video content because it serves users and supports accessibility and product clarity, which are legitimate reasons on their own - not because a specific citation lift from doing so has been measured.

An Active Research Area: Formal Citation-Quality Standards

Today, "does an AI engine cite my content" is mostly a binary check - is the domain named as a source, yes or no. That's starting to change on the research side, if not yet in any production ranking system. A 2025 ACL paper introducing "CiteEval" argues that a binary or simple entailment check is "a suboptimal proxy for citation evaluation" and proposes a graded 1-5 evaluation instead, considering source credibility, redundancy, missing evidence, and overall citation quality - evaluating not just whether a source was cited, but whether the citation actually earns its place. This is worth watching specifically because it signals where citation-quality thinking is heading in the research community, not because any AI search engine has confirmed it uses this or a similar framework in production today - that distinction matters, and the paper itself doesn't claim otherwise. On the measurement side that already exists for site owners, Google Search Console now provides a first-party report for link impressions in AI Overviews and AI Mode - a real production feature, described precisely rather than as an unbounded superlative: it rolled out to a limited set of properties in June 2026 before reaching general availability, and the dedicated view reports impressions by URL, country, device, and date only - no queries, no clicks, no position, and no detail on whether or how a specific page was named or credited in an answer. That's covered in full, with its current limits, on CoreAEX's measurement page. And on llms.txt specifically: it's still a third-party proposal, not an adopted standard on any major engine, a status covered in full (and not repeated here) on CoreAEX's dedicated llms.txt page.

Unresolved: Publisher Licensing Deals Are Growing, But They're Not Evidence of Paid Citation Preference

Publisher-AI licensing activity is real and growing, and it kept expanding through the time this page was last checked. OpenAI's April 2025 partnership with The Washington Post surfaces the paper's summaries and direct quotes inside ChatGPT with "clear attribution" and links back to full articles - one of more than 20 publisher partnerships OpenAI had described at the time, spanning over 160 outlets; OpenAI announced another regional media partnership in May 2026, with Brazil's Grupo Folha and Grupo UOL, giving both outlets access to OpenAI's enterprise and API tools in exchange for their journalism appearing in ChatGPT. Perplexity's publisher economics evolved too: its original Publishers Program (expanded December 2024) shared ad revenue and per-article performance data with outlets including the LA Times, The Independent, and Adweek, without disclosing exact financial terms; in August 2025 Perplexity added Comet Plus, a $5/month subscription product that allocates 80% of subscription revenue to participating publishers based on how much traffic their content drives, backed by an initial $42.5 million pool for early partners. Worth keeping in the picture rather than editing out for a cleaner narrative: this expansion is happening alongside active litigation, not instead of it. The New York Times moved past an earlier cease-and-desist letter and filed a federal copyright infringement lawsuit against Perplexity in December 2025, and Dow Jones has a separate active suit against the company - not every major publisher is opting in, and two of the largest are actively contesting how their content gets used in court. None of this - the growing deals or the litigation - is evidence that a licensing or partnership relationship increases a publisher's odds of being cited over a non-partner's content for a given query; the available reporting documents access, attribution format, and revenue terms, not a measured citation-rate or ranking effect. That's an unresolved question, not a disproven one - CoreAEX isn't aware of a study that tests it directly, and this page won't claim one exists.

Don't Conflate: Voice Search, Featured Snippets, and AI Citations Are Different Evidence Bases

These three surfaces get talked about as one continuous trend toward "the answer, not the link" - and conceptually, they do share that goal. But they're not measured the same way, and treating findings from one as evidence about another overstates what's actually known. Google's own featured-snippets documentation defines the format and its opt-out controls (nosnippet, data-nosnippet, max-snippet); checked directly, that documentation draws no connection anywhere to AI-generated answers or citations. That absence is a real documentation boundary worth noting - it isn't affirmative proof that featured-snippet mechanics and AI-citation mechanics are unrelated, only that Google's own reference for one doesn't claim to describe the other. Voice assistants share the underlying goal of direct-answer delivery, but a voice response, a featured snippet, and a generative AI citation are three different retrieval and display mechanisms with three different (and not fully overlapping) evidence bases behind them. If a claim about "voice search optimization" is being used to justify an AEO tactic, or the reverse, check whether the source actually studied the surface being discussed - a lot of pre-generative-AI featured-snippet research predates the systems this cluster is about by several years and doesn't automatically transfer.

The Honest Summary

The strongest current evidence shows a substantial but unstable disconnect between same-query organic rankings and AI citations, documented across several independent datasets using different methods. Cross-engine citation sets differ enough that measuring each engine separately, and testing whether a given tactic actually transfers, is justified by the data - but the evidence does not yet prove that gap is steadily widening over time, or that a per-engine optimization program reliably outperforms a shared one; both of those are real, live questions, not settled ones. Google has announced broader fan-out, agentic continuous-monitoring search, and deeper multimodal integration - those are product directions the company has stated publicly, not measured publisher-citation effects, and this page treats them accordingly. Publisher licensing is expanding - new deals, new revenue models - at the same time as active litigation against the largest AI search challenger from two major publishers; neither the growing deals nor the lawsuits demonstrate a paid-citation preference either way. A few other commonly-repeated ideas - that one format or schema choice reliably drives citations across engines, that voice-search research directly predicts AI-citation behavior - remain unresolved or unsupported by anything in this research library. The practical move isn't to treat any of the unresolved claims as settled. It's to keep measuring what's actually happening on the engines that matter for a given business, using the tools and caveats covered on CoreAEX's measurement page, and to treat this page's own claims the same way - worth re-checking against newer data before repeating them a year from now.

← Back to the AEO fundamentals pillar
How to measure whether AI engines are actually citing your content →
Does llms.txt actually do anything? →
Will agentic AI browsers replace crawling? →

---

Sources: The citation/ranking-overlap figures are from two Ahrefs studies: "38% of AI Overview Citations Pull From the Top 10" (863K SERPs, 4M AIO URLs, published March 2026, using January 2026 data, reporting the change from Ahrefs' own July 2025 ~76% figure, and disclosing an improved citation-parsing methodology between the two studies - Ahrefs' own explanation for the drop uses hedged language, "may be" and "could be," not a definitive attribution) and "Only 12% of AI-Cited URLs Rank in Google's Top 10" (15,000 prompts, July 2025 data, published August 2025) - the 11.9% headline average is calculated across five citation/reference series and includes Perplexity's 28.6%, not excluding it. An independent 2026 study using a different method reaches a comparable conclusion: Xu, Iqbal & Montgomery, "Measuring Google AI Overviews" (arXiv, May 2026; 55,393 trending queries over 40 days) found nearly 30% of AI Overview citations absent from the co-displayed organic results. AI Overview page-level volatility (roughly 70% of ranking pages changing over 2-3 months, near-zero correlation with organic-ranking changes) is from Authoritas' SERP and AI Overview volatility research (11,203 keywords, desktop only, August 2024-January 2025) - a volatility/correlation finding, not itself a widening-trend finding. AI Overview trigger-rate and query-type-mix trends, and the zero-click-rate figures, are from Semrush's AI Overviews impact report (10M+ keywords, January-November 2025) - note that page states the January navigational-query baseline as both 0.74% and 0.84% in different sections. The zero-click observation is paired with a stronger causal design reaching a different conclusion: Wang, Gleason, Bart, Wilson & Metaxa (arXiv preprint, August 2026; preregistered field experiment, N=1,100) found that removing AI Overviews and AI Mode increased publisher click-through. Google's own stated product direction is from Google's I/O 2026 Search announcement (information agents, agentic booking, May 2026) and Google's Search with Gemini 3 announcement (query fan-out, multimodal understanding, November 2025) - both first-party product descriptions, not independent measurement. The field's stated causal limits ("no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior") are from Olivier Martinez's critical survey - a single-author, non-peer-reviewed arXiv preprint reviewing 45 studies, conclusion bounded to that reviewed set. The graded-citation-quality research direction is from "CiteEval: Principle-Driven Citation Evaluation for Source Attribution" (ACL 2025, peer-reviewed) - a proposed evaluation framework, not a documented production standard on any engine. Google Search Console's generative-AI performance report scope and staged rollout (limited release in June 2026, broader availability following; impressions by URL/country/device/date only, no queries, clicks, position, or citation detail) are from Google's own announcement (title confirmed: "Introducing Search Generative AI performance reports in Search Console," developers.google.com, June 2026) cross-checked against third-party reporting after the primary post's full body text could not be retrieved directly. Featured-snippet definitions and controls, and the absence of any stated connection to AI-generated answers, are from Google Search Central's featured snippets documentation. The OpenAI/Washington Post partnership details are from OpenAI's own partnership announcement (April 2025); the OpenAI/Grupo Folha/Grupo UOL partnership is from OpenAI's own announcement (May 2026). Perplexity's original Publishers Program is from TechCrunch's reporting (December 2024); its Comet Plus revenue-sharing model (80% of subscription revenue to publishers, $42.5M initial pool) is from Axios' reporting (August 2025). The New York Times' copyright infringement lawsuit against Perplexity, filed December 5, 2025, is reported by multiple outlets including CNBC and Axios - superseding an earlier cease-and-desist letter.