If You Measure Gemini Citations, You're Probably Misfiling the Security Sources
· Published August 11, 2026 · 6 min read
Everyone measuring AI visibility now reports which sources an engine cites. It is the most useful number in the category, because it tells a brand who is actually deciding its reputation.
For Gemini, that number is quietly wrong, and it is wrong in a specific direction.
We found this while building the citation pipeline for a study on crypto safety questions. It is not a rounding error and it is not a Gemini bug. It is an artefact of how almost everyone works around a Gemini design choice, and it systematically undercounts security and audit firms.
TL;DR
- Gemini does not cite publishers directly. It cites through a Vertex AI redirect that hides the real domain.
- The common workaround reads the publisher from the citation title. Across 429 unique Gemini citations we hold, that title was a bare domain 429 times out of 429, so the workaround appears to work perfectly.
- It does not. The title is correct at registrable-domain level but drops the subdomain.
- On multi-tenant platforms the subdomain is the publisher. Following the redirects recovered certik.medium.com, slowmist.medium.com, peckshield.medium.com, quillaudits.medium.com and blockcrunch.substack.com.
- Four of those five are security audit firms, filed as generic
medium.comand therefore bucketed as publisher or media rather than security or audit. - Only 4% of citations need a network call to fix. The rest resolve free.
What Gemini actually returns
Ask Gemini a question with search enabled and the answer arrives with citations. The citations do not point at the publisher. They point at vertexaisearch.cloud.google.com/grounding-api-redirect/... followed by an opaque token.
Take that at face value and every Gemini citation belongs to Google. Nobody does that, because the resulting chart is obviously absurd. So tools reach for the citation title instead, which usually contains the publisher’s domain, and use it as the source.
That is a reasonable workaround. It is also where the error enters.
The failure everybody expects, and the one that actually happens
Our own research brief predicted the wrong failure. It assumed Gemini would sometimes return a human-readable page headline instead of a domain, leaving those citations unclassifiable, and it made solving that a hard prerequisite before any Gemini number could be reported.
We measured it. Across 429 unique Gemini citations in our stored scans, the title was a bare domain every single time. The failure the brief was designed around does not occur in any sample we hold.
The real defect is subtler and easier to miss precisely because the output looks clean. The title gives you the registrable domain. It does not give you the subdomain.
For most publishers that distinction is harmless. developers.uniswap.org arriving as uniswap.org is fine, because it is still Uniswap. en.wikipedia.org arriving as wikipedia.org is fine.
For anything published on a platform, it is not fine at all.
Where it breaks
We followed the redirects on every platform-hosted citation we hold and read the final host. Here is what the title path was throwing away:
| What the title said | What the redirect resolves to | Who that actually is |
|---|---|---|
medium.com | certik.medium.com | CertiK, smart contract audit firm |
medium.com | slowmist.medium.com | SlowMist, blockchain security firm |
medium.com | peckshield.medium.com | PeckShield, blockchain security firm |
medium.com | quillaudits.medium.com | QuillAudits, smart contract audit firm |
substack.com | blockcrunch.substack.com | Blockcrunch, crypto media |
Four of the five distinct platform publishers we recovered are security or audit firms.
Under the title-only approach, all four collapse into medium.com. Any classification scheme then files them as publisher or media, because that is what medium.com is. The security and audit bucket loses them.
Why this specific error matters more than it looks
If you are measuring which sources shape a brand’s reputation, this is not a cosmetic problem. It moves citations out of the one bucket that carries the most weight on the highest-stakes question a buyer asks.
Consider the query “is this exchange safe”. The answer is decided by a mix of community posts, review platforms, media explainers and security assessments. A brand reading that breakdown wants to know whether auditors are speaking for them. If every auditor publishing on Medium is silently reclassified as a media outlet, the report says security firms are near-absent from the conversation when some of them are in fact present.
It would not look like a bug. It would look like a finding. That is what makes it worth publishing.
The fix
The correction is small, and cheaper than the problem suggests.
Resolve the redirect, but only where it matters. Following every Vertex redirect would mean one extra network request per citation. Following it only for known multi-tenant platforms costs almost nothing: in our 429-citation sample, 16 citations, or 4%, needed a network call. The other 96% resolve for free from the title.
Keep the platform list explicit. Medium, Substack, WordPress, Blogspot, Tumblr, GitHub Pages, GitBook, Read the Docs, Notion, Wix, Ghost, Zendesk, Webflow, Netlify, Vercel and similar. On those hosts the subdomain is the publisher and the title is never trusted.
Never guess. A citation that cannot be resolved is recorded as unresolved. It is not silently dropped, and it is certainly not assigned to google.com. An unresolvable citation is data about coverage, not a gap to be papered over.
Record how each one resolved. Every citation in our results file carries whether it came from the title, from a redirect, or not at all. If someone disputes a classification, they can check the specific path rather than argue about totals.
How we verified it
We did not want to trade one assumption for another, so we checked the title path against ground truth.
We took 25 stored Gemini citations at random and followed each redirect to its final host. All 25 redirects were reachable, which also settles a practical worry: these links do not expire on the timescale a study runs over. Twenty matched the title exactly. The five that disagreed are the ones described above, and every disagreement was subdomain-level rather than a different publisher.
So the title is not unreliable. It is precise about the wrong unit.
What this does and does not claim
This is a finding about the DataForSEO route to Gemini, which is how a large share of the AI visibility tooling market gets its data. We have not tested the direct Google API, and the behaviour may differ.
Our sample is 429 unique citations from our own scans, which skew crypto and fintech because that is the category we work in. Crypto is unusually heavy in Medium-published security research, so the 4% figure is likely higher here than it would be in, say, retail. The direction of the error would be the same; the size would not.
We are not claiming anyone else’s numbers are wrong. We are saying that if a tool resolves Gemini citations from the title alone, and does not special-case platform hosts, then its security and audit share is understated by construction. That is checkable, and we would rather it be checked.
What to do about it
If you buy AI visibility reporting, ask your vendor one question: how do you resolve Gemini’s Vertex redirects, and do you preserve the subdomain? If the answer is that they read the domain from the citation title, their security and audit numbers are low.
If you build your own measurement, the fix is under an hour. Resolve platform hosts through the redirect, leave everything else on the free path, and mark unresolved as unresolved.
If you publish research on a platform rather than your own domain, this cuts the other way and is worth knowing: on Gemini, your byline may be getting credited to the platform instead of to you.
