Being cited by AI and being named by AI turn out to be different, weakly-correlated outcomes. We reviewed the study that split them apart, and the 2026 Google core updates that connect classic search rankings to AI answers.
Home › Research › Ghost citations
Whether the AI search landscape had shifted enough in 2026 to change what an honest AI visibility assessment should measure. Two Google core updates landed six weeks apart, Google published new guidance on AI search optimisation, and a new study claimed that citation and brand mention, the two things everyone bundles as "AI visibility", are not the same event.
We reviewed the Semrush and Kevin Indig "Ghost Citations" study (9 June 2026: 3,981 domain appearances from 115 prompts across ChatGPT, Gemini, Google AI Overviews and AI Mode in 14 countries, with fully disclosed methodology), the March and May 2026 core update analyses from three independent named analysts (Barry Schwartz's timeline reporting, Lily Ray's winners-and-losers analysis, Glenn Gabe's case studies), and the cross-confirmed Grokipedia case study in which two analysts using different tools observed the same Google-to-AI citation cascade. We also verified that Google's Quality Rater Guidelines have not changed since September 2025, which matters for what did not need re-examining.
Citation and mention must be measured as separate outcomes, per engine, with query phrasing controlled, or the measurement is noise. And because AI engines partly ground on Google's index, classic organic rankings remain the upstream supply line for AI visibility rather than a legacy metric. We revised our own assessment methodology on both counts.
The Ghost Citations study is the first widely-published dataset to split "AI cited your page" from "AI said your name". Across 3,981 domain appearances: 61.7% were citation-only, a source link with no brand name in the answer text; 25.1% were mention-only, named with no link; and just 13.2% were both.
This distinction is not pedantry. A ghost citation gives you a footnote most users never click. A mention without a link builds brand recognition but sends no traffic and no attributable evidence. The two outcomes have different values, different drivers, and, as the per-engine data shows, they do not even come from the same engines.
| Engine | Citation (links you) | Mention (names you) |
|---|---|---|
| ChatGPT | 87% | 20.7% |
| Gemini | 21.4% | 83.7% |
ChatGPT is a citation engine that rarely names; Gemini is a mention engine that rarely links. Google's AI Overviews and AI Mode sit between. The practical consequence: a visibility report built on one engine's behaviour says almost nothing about the others, and any tool selling a single blended "AI visibility score" across engines is averaging two opposite behaviours into a number that describes neither.
Query phrasing alone swings brand-mention rates 30-50x: short conversational queries produce dramatically more mentions than long structured queries on the same topic, and comparative phrasing ("best", "vs", "recommend") produces 2.4x more mentions than informational phrasing. Measurement that does not control query phrasing is not measurement.
In June 2026, two analysts using different tools independently documented the same phenomenon on the same domain: when a site drops in Google organic rankings, its citations in ChatGPT, Gemini and AI Overviews drop days later, because those platforms partly ground their answers via Google's index. Glenn Gabe tracked it through SEO visibility data; Malte Landwehr of Peec AI confirmed the same pattern through citation tracking. When the site recovered in Google, the AI citations surged back.
It is a single-domain case study, and we grade it accordingly. But it is the cleanest direct observation yet of something structurally important: traditional organic visibility is upstream of AI visibility. The businesses being told to abandon "old SEO" for "AI optimisation" are being told to cut the supply line that feeds the thing they are buying.
The March 2026 core update rewarded first-party "Information Gain" content, original substance a reader cannot get elsewhere, while UGC and aggregator platforms took significant organic losses. The May 2026 update was much larger, hit YMYL categories hardest, and arrived just six weeks after March, an unusually fast cadence. Alongside them, Google published new documentation naming scaled AI-generated content and AI-driven translation abuse as tactics it intends to police in AI search.
Two details matter for small businesses. First, the crackdown on "gaming AI search" is signalled but not yet fully enforced: spot-checks of known gaming sites after May showed mixed results, so the window where spam tactics appear to work is exactly the window in which to not adopt them. Second, Google's Quality Rater Guidelines, the ground truth for content quality assessment, have not changed since September 2025. The quality bar is stable; the enforcement around it is tightening.
Our free AI Visibility Snapshot distinguishes the outcomes this research shows are different: whether AI systems can read you, cite you, and name you. One clear page, within 24 hours.
Get My Free Snapshot →Part of the AI Visible Research: Our Findings series. Related: Brand Mentions Are the New Backlinks. Questions or challenges to the method are welcome via our enquiry form.