Six independent studies published between May and July 2026 looked at different corners of AI search — Google AI Overviews, cross-engine mention tracking, content freshness, click behavior, a core algorithm update, and the academic literature on generative engine optimization itself. None of them set out to answer the same question. Read together, they answer it anyway: the AI citation data enterprises use to run GEO programs is coming apart at the seams, faster than most reporting stacks have adjusted for. Citation no longer predicts recommendation. AI citation no longer moves with organic SEO visibility. "Fresh" content is not what most teams think it is. And some of the causal claims behind practitioner advice, including our own past claims, don't hold up against the research literature as cleanly as they get repeated in decks.
This piece is not a seventh study. It is a synthesis: we pulled six external, independently published pieces of research from the same ten-week window, laid their methodologies and findings next to each other, and looked for where they agree, where they contradict, and what the combination means for running a generative engine optimization program in the second half of 2026. Every statistic below belongs to the researcher or outlet that published it, named at first use and again in the citation list at the end. Where we extend our own prior work — the anatomy of an AI citation and our enterprise GEO readiness framework — we say so explicitly.
Why AI citation no longer means recommendation
Lily Ray, VP of SEO & AI Search at Amsive, tracked 100 B2B "best [category]" queries through Google AI Overviews at three checkpoints between April and June 2026, pulling citation data through Ahrefs Brand Radar. She published the analysis on her Substack on June 17, 2026, and Search Engine Land covered it the next day. The pattern: self-promotional "best X" listicles — pages where a company ranks itself against competitors on its own site — got cited 323 times across 80 AI Overviews she tracked. In 224 of those 323 citations, 69%, the brand whose own listicle got cited was excluded from the actual recommendation, with a named competitor recommended instead. Across all 100 tracked prompts, 74% showed some version of that split: the AI Overview cited one source and recommended a different vendor. Ray's framing is blunt: a citation is not a recommendation.
The mechanism is not mysterious once you separate what an AI Overview is actually doing into two steps. Retrieval finds pages that plausibly answer the query — a structured, on-topic "best X" listicle qualifies easily. That is the citation. Recommendation is a second, independent judgment: the AI Overview weighs what multiple sources say about a category rather than repeating what one self-interested page claims about itself, and a page where the publisher ranks itself first is structurally the weakest source in the room for that second step, even while it is a fine source for the first one.
This lines up with, and sharpens, our own prior finding that comparison content earns 32.5% of the citations we tracked across engines — more than any other format. That number was always a citation share, describing what gets pulled into an answer as a source. It was never a claim about which source the engine ultimately recommends. Ray's data supplies the piece that figure always needed: a comparison page can win the citation and still lose the recommendation to a name inside its own text.
“A citation is not a recommendation. — Lily Ray, VP of SEO & AI Search, Amsive, June 17, 2026”
AI citation data vs. SEO visibility: the platforms are diverging
Kevin Indig's Growth Intelligence Brief #21, published on Growth Memo on July 10, 2026, found something that had not happened before: for the five largest platforms by AI mention volume, June 2026 was the first month organic SEO visibility and AI Overview citations moved in opposite directions at the same time. LinkedIn's organic visibility index jumped 43.3%, from 289.4 to 414.7. Reddit rose 18.1%, reversing a prior decline. Instagram climbed 21.4% and Facebook 20.1%. On the losing side, Barnes & Noble fell 55.4%, Alibaba fell 40.6%, and Shutterfly fell 34.4%. None of those swings tracked cleanly with how often the same brands got mentioned inside AI answers over the same month.
Organic SEO visibility index change, June 2026 (Kevin Indig, Growth Memo)
The bigger warning sits underneath the headline platform moves. Total AI mention volume across Indig's tracked platforms sat almost flat, near 6.0 million mentions a week, for eight straight weeks. Read as one number, June looked uneventful. It was not. Split by engine, AI Overviews mentions fell 3.2%, AI Mode mentions rose 22.5%, and ChatGPT mentions fell 28.4% over the same stretch. A flat aggregate hid a real redistribution of where AI attention was actually landing.
| ENGINE | MENTION VOLUME CHANGE, SAME 8-WEEK WINDOW |
|---|---|
| AI Overviews | -3.2% |
| AI Mode | +22.5% |
| ChatGPT | -28.4% |
The practical read: a brand's blended "AI mentions" count can hold steady while its exposure to ChatGPT's specific user base collapses, or while AI Mode quietly becomes the channel doing the work. Our reporting and analytics work treats these as separate lines for exactly this reason, and it is the same failure mode we flagged when we found that a mention-rate dashboard that blends engines into one number was quietly lying to the teams reading it. A single visibility score cannot tell you which engine moved, and by the time a client asks why traffic shifted, the channel-mix answer needs to already be on the slide.
What 2026 AI citation data says about content freshness
Seer Interactive Insights, the research arm of Wil Reynolds' agency, published a study on July 24, 2026 that complicates the entire "freshness" conversation. Measured by the visible update date, 75% of LLM-cited pages had been updated within the last year, and 88% within two years — the numbers most teams would expect if the theory is "engines only cite new content." But Seer also checked the original publish date behind each page, and that number tells a different story: only 42% of cited pages were actually published recently. More than a quarter of pages that look freshly updated are two or more years old, refreshed rather than new. Engines are mostly citing old pages kept current, not rewarding a stream of net-new output.
Share of cited content updated within the past year, by engine (Seer Interactive)
Freshness requirements are not uniform across content types either. Seer found marketplaces need the most constant updating, 78% of cited marketplace pages updated within a year, followed closely by comparison and review content at 77%. Reference content sat at 74%, brand and corporate pages at 72%, blogs and guides at 67%, and news and editorial content trailed at 45% — the one category where engines seem comfortable citing something that has aged in place.
| CONTENT TYPE | SHARE OF CITED PAGES UPDATED WITHIN A YEAR |
|---|---|
| Marketplaces | 78% |
| Comparison / reviews | 77% |
| Reference | 74% |
| Brand / corporate | 72% |
| Blogs / guides | 67% |
| News / editorial | 45% |
The strategic implication cuts against a common instinct. Teams under pressure to feed the AI engines often default to publishing more net-new content. Seer's data says the higher-leverage move, for most content types, is auditing and updating what you already have — checking facts, updating numbers, refreshing examples — rather than adding volume, a job that is closer to on-page structure work than to a content calendar. That is a coverage problem in the terms of our enterprise GEO readiness framework: the framework already treats coverage as a scored dimension, and this summer's evidence says the freshness half of coverage needs its own tracked number, separate from raw publish velocity.
Zero-click search keeps compounding, independent of AI citation
Rand Fishkin's SparkToro published "In 2026, Less than One Third of Google Searches Still Send a Click" on June 8-9, 2026, and the trend line is the starkest in this review. 68.01% of Google searches produced zero clicks in the first four months of 2026, up from 60.45% in 2024 and roughly 49% in 2019. AI Overviews now appear on more than 20% of all searches, and their presence reduces click-through by nearly 60%. AI Mode, despite the attention it gets in trade press, routed only 0.34% of searches between January and April 2026. Total clicks of any kind, across all of Google, fell 22.9% between 2024 and 2026.
Share of Google searches ending with zero clicks, by year (Rand Fishkin, SparkToro)
Two things about this data matter for how you read every other study in this piece. First, the zero-click trend predates and outsizes the AI Overview story: it was already climbing from 49% to 60% before AI Overviews existed at scale, so AI is accelerating a structural shift, not creating it from nothing. That is consistent with what we found when we checked Google's own claim about AI search clicks against Ahrefs' click-through data earlier this summer. Second, AI Mode's 0.34% share means the citation-versus-recommendation fight Ray documented, and the platform-level divergence Indig documented, are both happening on a small and fast-growing slice of total search behavior, while the much larger zero-click shift is happening on ordinary Google results pages that never involve a generative answer at all. Generative engine optimization programs built only around AI Mode and AI Overviews are optimizing for the visible minority of the problem.
Core updates now reward source type and market fit, not one ranking factor
Aleyda Solis of Orainti published her analysis of Google's May 2026 core update on June 3, 2026, using Sistrix domain-level visibility data across the US and UK for the May 21-June 2 rollout window. The headline pattern: canonical reference brands gained (+24% UK, +10% US) while reference aggregators and tools fell (-29% UK, -13% US). Forums and Q&A content fell too (-24% UK, -12% US) — Reddit alone lost roughly 408 visibility points in the UK index and 361 in the US. E-commerce split by market and format: UK-local entities rose while US .com marketplaces operating inside the UK index fell hard, with amazon.co.uk up 21.3% against walmart.com down 59.5% in that same UK dataset. Health content split hardest of all: webmd.com rose 8.8% in the UK while goodrx.com fell 80% in the same market.
| SOURCE TYPE | UK VISIBILITY CHANGE | US VISIBILITY CHANGE |
|---|---|---|
| Canonical reference brands | +24% | +10% |
| Reference aggregators / tools | -29% | -13% |
| Forums & Q&A (incl. Reddit) | -24% (~-408 pts) | -12% (~-361 pts) |
| Local e-commerce vs. US marketplace | amazon.co.uk +21.3% / walmart.com -59.5% | — |
| Health: brand vs. aggregator | webmd.com +8.8% / goodrx.com -80% | — |
Solis's own thesis is the throughline that matters most: visibility shifted toward whichever source type best matched intent, market, and format, not toward one universal winner. A brand publisher won in the reference vertical. A local retailer won in e-commerce inside the UK index while a US-based marketplace lost ground in that same market. A branded health information site gained while a price-aggregator health site lost 80% in the same country. Four different verticals, four different winners, one update — a pattern that matches what we have seen in topical authority patterns across AI search, where authority concentrates by vertical rather than by domain size alone. Anyone reporting a single "we won or lost the core update" verdict to a client this summer is compressing four separate stories, in four separate markets, into one sentence that cannot be true for all of them at once.
What the academic AI citation research actually supports
Olivier Martinez's "Optimizing Visibility in Generative Engines: A Critical Survey", posted to arXiv on July 15, 2026, reviewed 45 GEO studies published between November 2023 and July 2026 — the broadest look at the underlying evidence base in this review. His conclusion is a useful check on the rest of the industry's confidence, including ours. The evidence that documents already sitting inside an LLM's context window can have their citation rates causally altered by GEO-style tactics is strong; multiple studies show it, and show it consistently. The evidence that those same tactics improve real-world organic discoverability, or that citations translate into actual traffic and conversions, is weak. The widely repeated "40% visibility gain" statistic that shows up across GEO sales decks and blog posts traces back to one narrow, single-metric result from one study, not a generalizable finding, no matter how many places have repeated it since.
Four metrics, not one: extending our GEO readiness framework
Put the six studies side by side and one throughline survives contact with all of them: the AI-era metrics that used to move as one number are separating, and they are separating faster than most enterprise GEO reporting has adjusted for. Citation does not guarantee recommendation. Citation volume does not move with organic SEO visibility, and both move differently by engine. A page's update timestamp does not tell you whether new information is actually needed, or just whether someone edited a paragraph. Zero-click behavior compounds regardless of how well any of the above is optimized. And core-update winners now split by source type and market rather than answering to one ranking factor. None of this is an argument against GEO work. It is an argument against measuring it with one blended score.
This extends, rather than replaces, our enterprise GEO readiness framework, which already scores accessibility, structure, authority, and coverage. What this summer's data adds is evidence that even within coverage, a single dimension, the underlying metrics are splitting into pieces that move independently. A page can be structurally sound, technically accessible, and authoritative, and still fail one of these four numbers without the other three catching it.
Start with an audit, not a rebuild. Pull your twenty highest-value pages and check four things against them this quarter: citation frequency across engines, whether the AI's actual recommendation names you when it cites you, whether the publish date matches the update date, and what share of the relevant queries convert to zero clicks regardless of your ranking. Where the numbers split, and on this data they will, report them separately in the next board deck instead of averaging them into a score that hides which one is actually moving. We are running this exact four-metric audit as the opening diagnostic on new GEO engagements this quarter, backed by our AI citation tracking build, the same way a DevSecOps platform like Arnica needed its citation and structure numbers separated before its GEO program could compound. The six studies here agree on very little in their particulars. They agree completely that one number can no longer carry the whole story.
See where you are cited today
A free snapshot audit of your rankings and AI citations before we ever talk.
Josh leads work at the intersection of SEO and generative engines at Something Inc., helping B2B brands get ranked and cited across every major AI engine.