Gemini citation statistics: a broken chain, traced

Why this category is graded broken chain

Most broken grades on this site record a sourcing failure: a figure repeats across secondary coverage and dead-ends in a vendor blog with no method underneath. Gemini fails earlier than that. The measurement that a Gemini-specific figure would summarise has not been made public by anyone, so there is no chain to follow. The master index lists it alongside local search as one of two categories broken for structural rather than sourcing reasons.

That distinction matters practically. A partial-graded figure can be quoted with its qualifier attached. A structural gap cannot be quoted at all — there is nothing to attach a qualifier to. The honest output is a description of where each circulating claim stops.

Which Gemini figures circulate, and where each one stops

Read the middle column as the diagnosis. Four of these six are not weak Gemini research; they are strong research on a different surface, wearing the wrong label. That substitution is the same degradation pattern catalogued in the provenance audit, with an extra step: the qualifier that gets dropped here is the product name itself.

What are the four Google AI surfaces?

Four products, one shared brand, one shared model family — which is precisely why casual writing merges them. The two search-embedded surfaces have their own pages here, and the measured difference between their source sets is the subject of the source-delta study. Gemini in Chrome is thinner again: what a browser-resident assistant actually fetches is covered on the agentic browsers page.

The staggered rollout is a real reason the vocabulary stayed messy. Each surface picked up its own coverage, written by different people at different times, rarely cross-referenced. The same vocabulary problem one level up is GEO vs AEO vs LLMO.

Why does Gemini cite on some answers and not others?

One mechanism explains most of the confusion. Gemini can answer from what the model already learned, or it can search first and answer from what it found. The second path is grounding, and Google documents the behaviour. Fact When Gemini grounds, it has real sources to attach and it attaches them. When it does not, there is nothing to cite. Same product, same question, two entirely different outcomes for a publisher.

That binary has no analogue on an engine whose whole product is retrieval, which is one reason Perplexity statistics are so much easier to collect. It also makes "how often does Gemini cite sources?" a badly formed question: the answerable version is how often a given kind of query triggers grounding, and that is unmeasured at scale.

Hypothesis We expect grounding to track how time-sensitive and how specific a question is. A current price, a recent event or a named product plausibly triggers a search; a general conceptual question plausibly does not. That is reasoning from how the feature is described, not from measurement. The practical consequence, if it holds: on topics answered from model memory there is no citation to win, and your leverage concentrates on the queries that trigger a search at all.

Which layer of an answer did a statistic count?

Different studies count different parts of an answer and then report a single number. Being named in the composed text is the most valuable outcome and the least often measured; appearing in the source list is the easiest to count and worth less.

Independent write-ups are useful for structure and much weaker as sources of a single comparable number — a per-platform comparison of citation formats, Profound's platform citation patterns and a survey of how each engine sources information all describe format without fixing a population. When you read a Gemini statistic, ask which layer it counted before you ask how big it is.

Why are Google surfaces harder to study than the others?

Fact Google documents its crawlers and the robots tokens that govern them. Google-Extended controls AI use of your content and is separate from Search indexing; a plain-language account of the token is a reasonable entry point. What it does not offer is a per-surface switch. There is no token meaning "allow AI Overviews but not Gemini".

Non-Google engines usually appear in your logs under a distinctly named fetcher you can handle on its own terms. Three sources show those fetchers directly: Cloudflare's breakdown of who is crawling the web, the finding that Googlebot still tops AI-associated crawler traffic, and the working lists of AI crawlers. Our reading of the same material is on the bot user-agent registry and the crawler statistics page.

This is the single largest reason Google surfaces are under-researched: the log-level evidence that makes other engine research possible largely is not there.

It also makes the publisher decision coarse. Allow Google's AI use of your content and you are in scope for grounding across every Google AI surface; refuse it and you keep classic Search indexing. Selective participation is not on offer. For most publishers the exposure outweighs content that is already public. For a subscription archive or a licensed dataset the calculation is genuinely different, and different again for regulated and health-adjacent categories. Write down what you expect to gain and lose, date it, and revisit when the terms change.

What transfers from other engines, and what does not

In the absence of Gemini-specific evidence, people borrow. Some borrowing is defensible.

Probably transfers. Retrievability basics — crawlable, parseable without heavy client-side rendering, claims stated plainly. Whether parsing survives client-side rendering is measured in the JavaScript rendering study; the rest is the technical audit.

Possibly transfers. The finding that quotes, statistics and cited sources are associated with higher visibility in generated answers, from the original GEO paper and restated across the LLM-friendly-content literature. The mechanism is general enough that it plausibly carries. Plausibly is not demonstrated.

Probably does not transfer. Source-mix findings. A heavy skew measured on one engine is a property of that engine's retrieval and licensing — see the Reddit dependency page and the Wikipedia dependency page. Cross-engine work such as Ahrefs' overlap measurement, Semrush's AI Overviews study and Seer's SearchGPT-to-Bing comparison is sober and still measuring surfaces that are not built the same way.

Definitely does not transfer. Crawler-level tactics that depend on separately named fetchers. The control surface is different here, so the tactic has nowhere to attach.

How to measure Gemini yourself

Nobody is going to hand you a reliable Gemini benchmark soon. A small self-run panel is imperfect and still more informative than nothing. Fix a set of twenty to fifty questions in customer language and never change it between runs.

Record whether each answer was grounded. An ungrounded answer is a different category, not a loss, and mixing the two produces meaningless averages. Record naming and linking in separate columns, because they move independently. Repeat each query at least three times, since generated answers vary between runs. Note the surface explicitly. Date every run and keep the raw records — the summary is disposable, and the log is what answers the question you have not thought of yet.

Two misreadings are worth pre-empting. Appearing in AI Overviews does not mean you are winning in Gemini — correlated, plausibly; equivalent, no. And a small drop across a small panel sits well inside noise, given run-to-run variance; only a sustained move across many queries is signal, the same threshold problem covered in the citation half-life study. If you want the tracking automated, the Claude Code guide covers how.

The study that would close the gap

We state the design publicly so someone else can run it if we do not get there first. One query panel, four surfaces — the Gemini app, AI Overviews, AI Mode and a non-Google control — same day, same location, same conditions. At least three runs per query per surface, so variance is visible rather than averaged away. Recorded per answer: grounded or not, the full source list, whether each source was named in the text, and citation position.

Predictions get pre-registered before collection. Ours: Gemini grounds less often than AI Mode, source lists overlap substantially across Google surfaces but not with the control, and entity-type questions ground least. Raw data publishes under an open licence — if we cannot publish the file, we should not publish the conclusion. Flat results go in the null results registry.

Open question This is registered as a flagship candidate on the Citation Index roadmap. Until it or something like it exists, every cross-surface Gemini claim you read — including any we might be tempted to make — is an inference.

Verification status

This category is graded Broken chain for a structural reason: the measurement does not exist, rather than existing badly. Nearly every mechanism claim here is Hypothesis pending the study described above. This is deliberately one of the thinnest statistics pages on the site — padding it with borrowed AI Overviews numbers would misrepresent Gemini specifically, which is the exact error the trace table documents. Who grades, and on what basis, is set out in the method notes.