How AI engines answer marketing questions: a citation study
We asked a live AI engine eight marketing questions and checked every source it named. It performed a live search for two of them. Of the sources it listed for the rest, 21 did not exist. This is our own test, run on one engine on one day, and the method and its limits are below.
Method
Date. 14 September 2026. Engine. Gemini, gemini-3-flash-preview, with Google Search grounding enabled. Sample. One answer per question. Eight questions. US English. Check. Every URL the model named was requested and its HTTP status recorded. Working files: the listed sources, the status check per URL, the resolved grounding redirects, and the grounded page titles. Not sampled. ChatGPT, Perplexity and Google AI Mode.
That is a small sample and we are not going to pretend otherwise. One engine, one country, one day, one answer per prompt. Read it as a first cut that produced a clear enough signal to be worth reporting, not as a measurement of AI search in general.
Finding 1. The engine usually did not look anything up
Live retrieval fired for two of the eight questions.
The two were “is there a marketing encyclopedia or wiki” and “what is a retrievable brand”. Both are questions about current, contested, fast-moving topics.
The six where it did not search were basic principles of marketing, the history of marketing and advertising, neuromarketing studies, building a B2B brand for AI search, movies about marketing, and songs about advertising. Those were answered from the model’s training data, with no search performed.
Evergreen marketing knowledge comes out of training data. Only the fresh and contested topics triggered a search.
Show the data
| Question | live retrieval fired |
|---|---|
| Basic principles of marketing | no |
| History of marketing and advertising | no |
| What is neuromarketing, best studies | no |
| Build a B2B brand for AI search | no |
| Best movies and series about marketing | no |
| Songs about advertising and consumerism | no |
| Is there a marketing encyclopedia or wiki | yes, 5 searches |
| What is a retrievable brand | yes, 4 searches |
The split is not random. Evergreen knowledge gets answered from memory. Fresh and contested topics get looked up.
Finding 2. When it does not look, it invents the sources
This is the part that should change how people read AI citation reports.
Where retrieval did not fire, the engine still produced a list of sources. Those lists looked entirely ordinary: real domains, plausible paths, titles that fit the claim.
We checked all of them. Across the questions where sources were listed rather than grounded, 21 URLs returned 404 or failed to resolve, 17 existed, and 15 blocked automated checking so could not be confirmed either way.
Roughly two in five of the checkable sources were not real.
Roughly two in five of the checkable sources were not real. They had correct domains, sensible paths and titles that fitted the claim.
Show the data
| Item | URLs |
|---|---|
| Exist | 17 |
| Blocked the check, unverified | 15 |
| Do not exist | 21 |
The worst case was songs about advertising and consumerism: of eight sources listed, one existed and six did not. No page anywhere owns that answer, so the model filled the gap with plausible-looking addresses.
By contrast, the question where retrieval did fire properly, “what is a retrievable brand”, returned 16 grounded citations and all 16 existed.
So the pattern is clean. Grounded answers cite real pages. Ungrounded answers produce a bibliography that looks like research and is not.
Finding 3. Who owns the answers that are grounded
For “what is a retrievable brand”, the 16 real citations broke down as two platform vendors, two trade publications, two companies using the term as a product name, seven small agency or consultant blogs, and one Medium post.
Not a single established marketing institution. The term is being defined in public by whoever wrote a definition-first explainer with question headings and a date on it.
Every cited page shared that shape.
What this means if you are trying to be cited
Establish whether the engine retrieves for your category before spending anything. If it answers from memory, page-level work does not reach it, and the lever is third party mention over a much longer timescale.
Check citation reports before believing them. If a tool shows your brand cited in AI answers, click the URLs. Some proportion of them may not exist.
Definition-first pages with question headings and visible dates are what got cited in the one grounded answer we could analyse in depth. That is a sample of one answer and should be treated as a hypothesis rather than a rule.
Limitations, stated plainly
One engine. One day. One sample per question. No statistical claim is being made and none should be read in. Engines change retrieval behaviour without announcing it, and this test will need repeating before any of it can be called a trend.
We publish it because the direction was strong enough to be useful and because nobody else seems to be checking whether the citations are real.
If you run the company rather than the marketing
One instruction from this, and it costs nothing.
If anyone shows you a dashboard reporting that your brand is being cited in AI answers, ask them to open three of the citations in front of you. In our test roughly two in five of the checkable sources did not exist.
A citation report is not evidence until somebody has clicked the links. That is not a criticism of any particular tool, it is a property of how these systems produce source lists when they have not actually searched.
The second question worth asking: does the engine look anything up when our buyers ask their questions, or does it answer from memory? If it answers from memory, then any programme of website changes is addressed to a mechanism that is not running, and the honest route is slower and mostly involves other people writing about you.
Sources
Our own test. Run 14 September 2026 on Gemini gemini-3-flash-preview with Google Search grounding, US English, one sample per question. Raw files are held with the study: model-listed sources, HTTP status per URL, resolved grounding redirects, grounded page titles.