Capability gap · from our probe corpus

How does Gemini grounding pick citations?

We can't see Google's internal ranking logic — no outside party can. What we can report, from our own dated probe runs and from building the matching logic ourselves, is a mechanical detail that matters if you're trying to measure Gemini citations at all: the links Gemini's grounding API returns aren't the real destination URLs — they're redirect links that have to be resolved before you know who was actually cited.

The redirect-URI finding

Gemini's grounded responses include a groundingMetadata block with citation chunks, and the URL in each chunk points to vertexaisearch.cloud.google.com/grounding-api-redirect/... — a Google-hosted redirect, not the cited page's own domain. Reading that URL directly and extracting a domain from it (a "title heuristic" — guessing the domain from the page title text instead) produces the wrong answer or no answer at all. Our probe engine resolves every one of these redirect links by following it to its real destination before recording a citation — a deliberate design choice, not a default in most naive implementations of this API.

This matters for anyone building or evaluating an AI-visibility tool: a tool that reads Gemini's raw grounding URLs without resolving redirects will systematically undercount or misattribute Gemini citations. It's the kind of gap that's invisible unless you go looking for it.

What our probe runs show, once resolved

Across the ten intents we probed on 2026-07-20 (CRM and AI-marketing-tools categories), the Gemini grounded lane cited between 1 and 9 distinct resolved domains per answer. For "best free crm for startups," Gemini cited five domains — claritysoft.com, mo.agency, hubspot.com, nutshell.com, and zeeg.me — a mix that included one recognizable vendor (hubspot.com) alongside four smaller or less-known domains. See the full CRM benchmark's raw evidence for the per-intent breakdown.

Vendor sites and independent content both appear

For "best crm for small business," Gemini cited pcmag.com (an independent review outlet), zoho.com, salesforce.com, hubspot.com (vendor domains), and uschamber.com (a nonprofit trade association's resource content) in the same answer — spanning vendor, media, and institutional sources without an obvious preference for one type over another in this sample.

What this means for you

If you're trying to measure whether Gemini cites your site — whether by hand or with a third-party tool — confirm the tool actually resolves grounding redirect URIs to real domains rather than reading the raw link or guessing from the answer's title text. A tool that skips this step can report you as uncited when you weren't, or misattribute a citation to the wrong domain entirely.

Frequently asked

Is this how Google officially documents Gemini grounding citations?

The redirect-URI behavior is visible in the API's own response structure (the groundingMetadata field), not a secret — but the implication that naive readers of it will misattribute citations is our own finding from building against it, not something Google calls out prominently.

Does Gemini cite the same sources every time for the same query?

Not reliably — like the other lanes, answers are generated fresh per query from a live grounded search, so repeated asks can return a different citation set.

How is this different from Perplexity's citations?

Structurally different end-to-end — see how Perplexity chooses its sources, which returns direct citation links rather than a redirect layer, and showed a different domain-count range in our runs.

Related

Want the raw per-intent data behind this page? See the full benchmark hub, or check your own domain's overall AI citation status with the AI visibility checker.