The Demand Side AI Visibility Misses — Which Sources Your Buyers Trust

Key takeaway: Most tools that measure brand visibility in AI answers check one direction only — did our content get cited. An arXiv preprint published on August 30, 2026 raises the other side: there is no demand-side data on what buyers ask and which sources they trust, so the two measurements cannot be joined. To close that gap the authors released a buyer persona corpus of just over a million records. This article takes the one axis you can apply immediately — which source types your buyers trust — and sets out the limits you should respect when citing this research.

---

Three-line summary

• The default in GEO measurement is supply-side. You can see whether your pages were cited, but not whether the prompt that produced the answer resembles a real purchase-evaluation question, nor which sources that buyer trusts.
• The new corpus records a preferred sources field for every persona. Joined against citation-provenance data, it makes a sharper question possible: are we being cited in the source types our customers actually trust?
• It is a synthetic corpus, and the authors themselves leave the empirical estimate of that join to future work. For now this is a measurement-design contribution, not a set of figures to quote.

Today's AI visibility measurement looks one way only

The question most AI visibility tools answer is a single one — did we appear in the AI answer. Mention or no mention, the cited URLs, how often competitors appear alongside us. Call this the supply side: it looks at how the content we produced was reflected in the answer.

The problem is what that view cannot settle. It cannot tell you how close your measurement prompts are to a real purchase-evaluation question, and it cannot tell you whether the sources cited in the answer belong to the source types your customers actually trust. The query-design half of that problem was covered in an earlier article, Why Brand-Name Prompts Can't Measure AI Visibility. This article takes the other half: source trust.

What the paper argues — there is no demand-side data to join

The preprint published on August 30, 2026 names the gap directly. Studying whether a brand enters the shortlist inside an AI answer, the authors argue, requires data from the demand side.

"Studying how brands enter or fail to enter that shortlist requires demand-side data: what buyers in a category ask, what information they need, and which sources they trust." — Demand-Side Measurement for Generative Engine Optimization, arXiv 2608.30023 (August 30, 2026)

The sentence that follows carries the weight. Large persona datasets already exist, but they were built for training-data diversity, so they carry neither a staged search-intent label nor a preferred-sources field — and therefore cannot be joined to supply-side measurement. The diagnosis is not that data is scarce, but that no data exists in a joinable shape.

PersonaGen-1M is the authors' attempt to close that gap. Its composition:

Item	Scale
Synthetic buyer personas	1,031,732
Industry labels	511
Market contexts	4
Structured behavioral attributes	19,416,821
Of which, search queries	5,160,046

It was built by taking roughly 40 million raw persona descriptions from four public datasets, deduplicating them, and enriching what remained to a fixed schema. The full corpus is shared on request for non-commercial research, and a stratified subset is published openly so the protocol and schema can be inspected directly.

The distribution of primary intent labels across personas looks like this:

Intent label	Share
Informational	78.3%
Commercial	17.4%
Transactional	4.3%

Do not read these shares as a measurement of market demand. They are the outcome of generating and labeling synthetic personas, so they do not mean that 78.3% of real buyers ask informational questions. What the table is good for is not the ratio but the structural point behind it: the commercial-evaluation queries that actually produce recommendations are a minority of all queries. Collect measurement prompts casually and most of them will be informational, leaving very few that test whether you enter the shortlist.

Preferred sources — the axis after query design

The part of this research you can apply immediately is the preferred-sources field attached to each persona. It names the source types that persona would trust, and the authors state that joining it against citation-provenance data is the corpus's primary intended use.

Where that join holds, the measurement question sharpens by one step. The old question was "were we cited." The question after the join is "are the source types we get cited in the same source types our customers trust?" The two lead to different work.

• If the answer to the first is no, the work is content and crawl accessibility.
• If the answer to the second is no, you are being cited — in the wrong channels. Writing more content does not fix that; the channels your customers trust have to be addressed directly.

That second situation is observable in practice. In a separate analysis of Perplexity's answer stream, the cited material shifted by question type — place data for local questions, video for how-to questions, primary announcements for news. The type of source being cited moves with the question, which makes where you sit among those types a separate thing to check.

Putting it to work — three questions

You do not need the paper's data to use its perspective. Three checks, in order.

Question to check	How	Common mistake
Do we enter the shortlist on purchase-evaluation questions in our category?	Measure repeatedly with category and comparison prompts that exclude your brand name	Reading results from prompts containing your brand name as performance
What source types does the AI rely on when it introduces us?	Group the answer's citation list by type — official docs, press, community, reviews, comparison content	Counting citations without looking at the type mix
Do those types match the ones our customers trust?	Organize the "where did you hear about us" answers from sales and support records into the same types, then compare	Assuming the channels easiest for us to produce are the channels customers trust

The third row is the point of this article. The first two can be measured with tools; the third needs your own customer data. Run the comparison and it is not unusual to find that the channels performing well on citations are not the channels customers name as trusted. That is where the need for the join becomes concrete.

RanketAI's brand visibility analysis covers the first two — whether your brand appears on prompts written from a position of not knowing it, and what sources the answer cited — measured repeatedly and shown as a trend. The third comparison is not something a tool can do for you; it has to be matched against your own sales and support records.

Three limits to respect when citing this research

The paper is interesting, and there are clear places where quoting it directly would be wrong.

1. It is synthetic data. These are generated personas, not observed buyer behavior. Quoting the intent distribution as a market statistic makes a claim the data does not support.
2. The central claim is not yet validated. The authors describe the join as the corpus's primary intended use and, in the same breath, write:

"that join is the primary intended use, and its controlled empirical estimate is future work." — arXiv 2608.30023 (August 30, 2026)

No controlled estimate of whether the join actually predicts anything has been presented yet.

1. It is a preprint without peer review. No author affiliation appears on the abstract page either. Borrowing the methodological idea and citing the results as evidence are two different things.

In short, the value here is in the problem definition, not the numbers. The claim that supply-side measurement is only half an interpretation without demand-side data stands independently of whether the corpus validates, and you can act on it with the material you already have.

FAQ

Related Reading

• Why Brand-Name Prompts Can't Measure AI Visibility — the half this article leaves out: standards for query design.
• Perplexity Never Skips the Web — an observation of how cited source types shift with question type.
• Ghost Citations: Why 62% of AI Citations Never Mention Your Brand — cases where the citation happens but the brand does not stick.