Where Gemini's answer came from: a source-checking guide with a 100-answer audit

A Gemini answer can come from two places, and they look identical on screen: what the model absorbed during training, and pages fetched while your answer was being written. The only reliable way to tell them apart is the displayed source list attached to that specific answer. So the audit below counts one thing above all: how many answers carried at least one displayed source link, over the total answers tested, with the dates. We publish the full design, and we state plainly that as of 29 September 2026 we have not completed a 100-answer Gemini audit, so this page carries no Gemini percentage.

What it does carry is the design, the classification rules, and the equivalent counts we have actually taken on other assistants with their dates attached.

"Where the model learned" is not the same question as "what sourced this answer"

These two questions get merged constantly and the answers to them are unrelated.

Where the model learned is a question about training data. It is largely not public in any auditable form, it covers material from years rather than days, and you cannot inspect it from a chat window. If your business is described from that memory, the description reflects what was widely written about your category at some point in the past, which is why it can be confidently out of date.

What sourced this answer is a question about one request. It is partly visible: if the interface shows a search step and attaches links, those links are evidence that pages were fetched. This is the question you can audit, and it is the one that matters when you want to know whether your pages are reachable.

The practical test for which one you are looking at: ask for something that could only be known from a page published recently, and see whether the answer has a link on it. An answer with no links to a question that needs current information is almost certainly memory.

Google's documented information paths

Four paths that Google's own material describes, kept separate because they behave differently.

• Model knowledge from training. Fixed at training time, not inspectable from the app, and the reason an answer can be fluent and stale at once.
• Grounding with search, where it is used. The API documentation describes a grounding mode that runs queries and returns information about the sources used. This is the path that produces citations.
• Google's own indexes and product surfaces, which is why what your site exposes to a crawler matters even for an answer that is not a search result.
• What you put in the conversation. A document you upload, a URL you paste, or a previous turn in the same chat. This is a source too, and in a test it is a contaminant.

Read Google's own pages for each of these and note the date you read them. A quotation from developer documentation with no date on it is worth very little a quarter later.

Audit method

The 100-answer design

• Twenty-five questions, four runs each, which gives 100 answers. Four runs let you see per-question stability, which a single pass of 100 different questions cannot.
• Five questions in each of five bands: questions needing current information, questions about a company's facts, best or top questions, task questions, and questions about risks or problems in the category.
• A clean session every time. New chat, no uploads, no pasted links, no memory, no custom instructions.
• One surface at a time, with the model name as shown on screen recorded for each answer. App answers and API answers are counted separately and never added together.
• Nothing named. Your brand does not appear in any prompt.

The six fields recorded for each of the 100 answers

• Date, time and time zone.
• Surface and displayed model name.
• Whether a search or browsing step was visible while the answer was produced.
• The number of displayed source links, and every URL.
• Whether the answer contains a claim that could only come from a page published after the model's stated knowledge cutoff.
• Whether at least one displayed link, when opened, actually supports the specific claim it appears next to.

The sixth field is the one that separates a real audit from a link count. A link that does not support the sentence it is attached to is a decoration, and it is common enough that it needs its own column.

The three figures the audit produces

• Sourced answer rate: answers with at least one displayed source link, over 100, with the dates.
• Supported claim rate: answers where an opened link genuinely supported the claim, over the answers that had links.
• Your own domain rate: answers citing your domain, over 100.

Report all three as counts over denominators. A percentage from 100 observations is legitimate to write down, and it is still one sample of one prompt set on one set of days, which is why the dates travel with it everywhere.

Results

Gemini: not run, as of 29 September 2026. No sourced answer rate, no supported claim rate, no Gemini domain count. The section is empty on purpose and will be filled with the numerator, the denominator and the run dates when we have them.

The equivalent work we have completed, on other assistants:

• Perplexity, September 2026, on aiknowsus.com. Across 24 batches and 72 conversations on our own domain, the phrase recording that we were not cited appears 161 times in the engines' own self audits of their answers. In one batch of six questions about our own category, with no brand named, the assistant reported afterwards that it had not cited or recommended aiknowsus.com in any of the six answers, so there was no position for it to hold.
• Perplexity, September 2026, on aiknowsus.com. The domains cited most often across the whole capture were the assistants' own documentation and the vendors' own websites, with a single well known review site appearing far down the list. That is a direct reading of what sources decide these answers, and it points at official pages and vendors' own pages rather than independent reviews.
• ChatGPT, 27 July 2026, on clawlaw.in. Asked 78 buyer questions with no brand named, ChatGPT then confirmed it had run no live web search for any of them. A sourced answer rate for that run would have been zero, and the answers still read as confident.
• Claude, 17 September 2026, on clawlaw.in. The assistant admitted it had used two specific arguments drawn from clawlaw.in pages and dropped the attribution in both places, which it described as a citation lapse rather than a ranking judgement. This is the case that shows why a link count understates how often your pages are actually used.

What the audit cannot infer

Six limits, and the first is the one most reports get wrong.

• A displayed link does not prove the answer was built from it. The model may have known the content already and cited a page that agrees. The link shows a page was available, not that it was decisive.
• No link does not prove nothing was fetched. Interfaces differ in what they show, and an absent citation may be an interface choice.
• The model's own account is not a log. Asked afterwards how many questions it had actually searched for, one assistant withdrew its own earlier statement, saying it could not honestly substantiate the claim that it had run a live search for each question, and in another batch that its claim to have searched all five was not adequately supported. Perplexity, September 2026, on aiknowsus.com.
• It cannot see training data. Nothing in this audit tells you what the model absorbed about you before the test.
• It is personal to your account. Location, sign-in state and history all shape what you get, so your rate is your rate.
• It does not transfer. A Gemini audit says nothing about ChatGPT, Claude, Copilot or AI Overviews.

Reproduce the audit

The whole thing runs on four files and about four hours of work spread over a fortnight.

• questions.csv: twenty-five rows, each with its band.
• log.csv: one hundred rows, carrying the six fields above.
• answers/: one saved file per answer, with the source list intact. Save these first and score later, because you cannot re-open an answer you did not keep.
• rules.txt: the classification rules written out, so the person scoring on day fourteen applies the same test as on day one.

Two habits make the difference between an audit and a pile of screenshots. Open the links. And write down the run in which an answer changed, rather than only the final state.

Common questions

Can I ask Gemini to tell me its sources after the fact?

You can, and a list produced on request is not the same as the list displayed with the original answer. Treat anything produced afterwards as a comment on the answer. We hold a dated case of an assistant withdrawing its own claim about having searched, which is the reason for the caution.

Why does my company appear with no link at all?

Usually because the mention came from memory rather than from a fetched page. That is worth knowing, because it means the description will not change when you publish something. It changes when the assistant has a reason to fetch, which is a different problem.

Is 100 answers enough?

It is enough to write a rate with its denominator, and it is not enough to claim a general truth about Gemini. It is a sample of twenty-five questions on a few days. Our own completed runs have been smaller than that, and we report them as counts for exactly this reason.

Should I count links to my competitors?

Yes, in a separate column, and it is often the most useful column in the file. It tells you which pages the assistant treats as authoritative in your category, and those pages are where your absence costs you most.

What if the answer quotes my page without linking it?

Record it as used without credit and keep it separate from cited. In our own programme an assistant admitted using two arguments from our pages and dropping the attribution in both places. Claude, 17 September 2026, on clawlaw.in. A business that only counts links cannot see that it has already won the reading and lost the link.

Primary-source references

For the documented paths: Google's Gemini app help centre, the Gemini API developer documentation on grounding and on the search tool, and Google Search Central's material on AI features and on site owner controls. Read them directly and date what you quote.

For our own observations: the clawlaw.in programme from July 2026, and the aiknowsus.com audit of September 2026 covering 24 batches and 72 conversations. Of the observations quoted on this page, the ones with a URL in them can be checked by a reader from outside; the counts taken across our capture files cannot be, until we publish the captures, and we would rather say that than imply independence we do not have.

Change log. First published 29 September 2026 with the Gemini results section empty and marked as empty. Disclosure: written by AI Knows Us, which sells AI visibility measurement.

What to do first

Pick five questions that need current information about your category, ask them on Gemini in five clean chats, and note for each one whether any source link appeared. Five answers will not give you a rate, and they will tell you within ten minutes whether you are being described from memory or from pages. That decides which piece of work is worth doing next.