ChatGPT, Claude, and Perplexity show you their sources. Gemini does not.

In a 46-prompt panel run across four AI engines, three of them returned citation URLs that identify specific pages: broadbandnow.com, facebook.com, ispreports.org, the brand's own site. Gemini returned 591 citation URLs too. But 590 of them point to an opaque Google redirect with no domain, no page title, and no way to know what is underneath.

What we found

We ran a full prompt panel for a brand in a locally-competitive services category: 46 buyer-intent prompts asked of four AI engines that use live web search (ChatGPT via the OpenAI Responses API, Claude, Gemini, and Perplexity). We extracted every citation URL each engine returned in its responses.

For ChatGPT, Claude, and Perplexity, the URLs were direct links to source pages: comparison aggregators, the brand's own website, directory listings, forums. You could open any of them and see what the engine read. For Gemini, every citation URL was a redirect through vertexaisearch.cloud.google.com, a Google intermediary that wraps each source in an opaque token string. The token does not decode to a readable URL and the redirect does not pass through to the original page in a way that reveals the domain.

The single non-redirect Gemini URL was a Google search results page, not a source page. In practice, Gemini returned zero auditable source domains across 46 responses.

What the other engines show

Because ChatGPT, Claude, and Perplexity return readable URLs, you can extract a ranked source list for each engine. Here is what the top domains looked like for this brand's category:

These tables drive the fix list. If the brand wants to improve its ChatGPT score, broadbandsearch.net and ispreports.org are the first stops. If Perplexity is the weak engine, the Perplexity source mix shows facebook.com as the dominant platform, not the comparison aggregators that dominate ChatGPT. The fix is different for each engine because the source mix is different.

With Gemini, there is no equivalent table to produce. Gemini mentioned the brand in 14 of 46 responses (30 percent mention rate, roughly on par with ChatGPT and Claude). So it is clearly reading pages relevant to the category. You just cannot see which ones from the citation data alone.

Why Gemini works this way

Gemini's web grounding routes source attribution through Google's Vertex AI Search infrastructure rather than passing original URLs directly to the API consumer. The design choice appears to serve Google's own product goals: the citation links work as intended inside the Gemini consumer interface, where clicking one triggers a Google Search for the source topic. As an API consumer trying to audit which pages shaped an answer, you get the redirect token but not the destination.

This is not a measurement artifact. Every other engine we tested passes the original source URL. Gemini's behavior is specific to how Google exposes grounding data through its API.

What this means for the fix strategy

For ChatGPT, Claude, and Perplexity, the per-engine source audit is the fix playbook. You look at which domains the engine cites most often for buyer-intent questions in your category, confirm whether you are present on those pages, and close the gap. Getting listed on a comparison aggregator that appears 24 times in Claude citations has a predictable effect on Claude visibility.

For Gemini, you cannot run that playbook from citation data. The alternatives:

• Broad entity presence. Since you cannot see which specific pages Gemini favors, distribute across the well-known sources in your category: major comparison sites, your Google Business Profile, directories with high overall authority. You are casting a wide net rather than targeting identified gaps.
• Consumer-surface testing. The Gemini consumer product (gemini.google.com) shows cited sources with working links. Running prompts there and reading the cited links manually gives a rough picture of which pages Gemini favors, though it diverges from the API surface and does not scale.
• Track mention rate separately. Because Gemini's source gaps are not auditable by the same method, Gemini visibility should be tracked as a mention-rate number on its own. A change in Gemini score tells you the direction; it cannot be explained by a source audit the way ChatGPT or Perplexity changes can.

What a full audit shows

A CitedWell audit reports the per-engine source tables for ChatGPT, Claude, and Perplexity: every domain cited, how many times, and whether the brand's own site was in the mix. For Gemini, the report shows mention rate and recommendation rate per prompt alongside the consumer-surface note on auditability. The fix plan distinguishes between the three engines where source gaps are traceable and Gemini, where the remediation path is broad entity coverage rather than targeted page placement.