AI Mentions the Brands It Mentioned Before — A 35K-Observation Study and Korean SaaS Data

Key takeaway: The variable that best predicted whether a brand appears in an AI answer was not the sources behind this answer but the brand's mention history in earlier runs. A study published in September 2026 (Tannenbaum, arXiv:2609.23162, preprint) analyzed 34,960 unbranded prompt observations and reported mention rates of about 3% when the brand's own domain was absent from the sources, versus 49–58% when the own domain was cited. When RanketAI re-aggregated July and August measurement records for 12 Korean B2B SaaS companies, the direction was the same: 97% of brands mentioned in one round were mentioned again in the next. The author, however, is the CEO of a commercial measurement vendor and the data are operational records, so this post treats every figure as an association, not a causal effect.

---

Three-line summary

• According to the study, brand mention rates were about ten times higher when the brand's own domain appeared in the answer's source list (GPT 2.8% → 49.0%, Gemini 3.8% → 58.4%). When a search query containing the brand name also appeared, the rates were 91.4% and 100%.
• An even stronger predictor was whether the brand had been mentioned for the same question before. In our re-analysis of 12 Korean B2B SaaS brands, the previous round's mention status almost fully predicted the next round's, and 14 of 15 changes in own-domain citation left the mention status unchanged.
• In practice, track mention rate and own-domain citation rate as separate metrics, and decide the next action from their combination (four states).

What the study measured, and how

The study breaks the assumption "build a good page and AI will mention you" into stages. Between a user's question and the final recommendation, the AI writes search queries, chooses sources, and decides which brands go into the answer. The author turned the signals observable at those stages into variables and fitted a model that predicts whether a brand will be mentioned in the next run.

Item	Detail
Data	75 anonymized projects on a commercial measurement tool, 2,854 monitored prompts
Observations	34,960 unbranded prompt × engine observations
Engines	GPT models (with web search) and Gemini models, repeated runs from June to September 2026
Mention	The target brand name appears literally in the answer
Own-domain exposure	The company's registered domain URL is in the source list the engine recorded
Branded fan-out	The brand name appears in at least one search query the AI generated
Prior history	The share of earlier runs of the same brand, prompt, and engine in which the brand was mentioned

"Unbranded" means only prompts without the target brand name were counted. Why prompts that include the brand name almost always produce a mention in web-search-enabled AI — and therefore carry no measurement information — is covered in our brand-name prompt post.

Finding 1 — Mention rates differed tenfold depending on what appeared in the source list

The study's first table shows mention rates by the signals observed during answer generation.

Observed signal	GPT mention rate (n)	Gemini mention rate (n)
Neither	2.8% (15,524)	3.8% (13,801)
Own-domain citation only	49.0% (1,769)	58.4% (3,415)
Branded search query only	64.4% (59)	84.8% (33)
Both	91.4% (128)	100.0% (231)

Source: Tannenbaum, arXiv:2609.23162, Table 3. Numbers in parentheses are observations per cell.

Interpretation. When neither the own domain nor a search query with the brand name appears, the chance of the brand appearing in the answer is around 3%. When the own domain is confirmed as a source, it rises to roughly half. The author summarized the gap this way:

"Evidence exposure is associated with roughly an order-of-magnitude difference in brand-mention probability in both GPT and Gemini." — Tannenbaum, arXiv:2609.23162 (2026-09-19)

Looking only at repeated runs of the same prompt and comparing runs without a branded search query, own-domain exposure was still associated with a mean mention-rate increase of 40.2 points on GPT and 49.0 points on Gemini. Because this comparison removes differences between prompts, it is somewhat more robust than the table above.

The branded-search-query cells contain only 59 and 33 observations, so their magnitudes are not reliable. The direction matches earlier observations: AI includes brand names the user never typed in the queries it writes, and brands named in those queries had a 68.9% citation rate versus 2.1% for brands retrieved without their names. That overseas case and our measurement across 12 Korean categories are covered in the search-query brand injection post.

Finding 2 — The strongest predictor was prior mention history

The second finding is the core of the study. The author compared three prediction models on held-out data that was not used for fitting (AUC; closer to 1 is more accurate).

Prediction model	Information used	GPT AUC	Gemini AUC
Prior history only	Share of earlier runs that mentioned the brand for the same prompt	0.937	0.917
Live evidence only	This run's own-domain exposure and branded search query	0.880	0.840
Both	Prior history + live evidence	0.963	0.942

Interpretation. Whether this brand was already being mentioned for this question explained more than which sources this particular answer used. The combined model was the most accurate, which also means prior history and live evidence carry information that does not overlap.

How the two work together shows up in mention rates by prior-history band.

Prior-history band	No own-domain exposure	Own-domain exposure
Low (earlier mention share below 0.1)	GPT 0.9% · Gemini 1.0%	GPT 25.1% · Gemini 23.8%
High (earlier mention share 0.5 or above)	GPT 32.0% · Gemini 28.7%	GPT 82.0% · Gemini 86.2%

Even for prompts where the brand had rarely been mentioned, own-domain exposure lifted the mention rate from around 1% to 24–25%. For prompts where the brand had often been mentioned, the rate was 82–86% with exposure and 29–32% without. A strong history does not hold the mention if the own domain drops out of this answer's sources.

In the conclusion, the author notes a second path: established brands can appear in answers even without their own domain being cited (in the original, "established brands can appear without their own domain being cited"), and page scoring or citation tracking alone cannot explain this path.

Re-analysis of 12 Korean B2B SaaS companies — same direction, stronger persistence

We checked our own records to see whether the same direction holds for Korean-language questions. RanketAI measured 12 Korean B2B SaaS companies under identical conditions for the July and August benchmarks. We re-aggregated the raw records from two rounds on July 23 and two rounds on August 12 into 144 cells of brand × AI service × round. The services measured were web-search-enabled ChatGPT, Gemini, and Perplexity.

First, how often own-domain citation and mention appear together:

AI service	Own domain cited → mentioned	Own domain not cited → mentioned
ChatGPT	7/8 (88%)	7/40 (18%)
Gemini	34/34 (100%)	6/14 (43%)
Perplexity	29/29 (100%)	12/19 (63%)
Total	70/71 (99%)	25/73 (34%)

Next, persistence between rounds. Across 108 round-to-round transitions for the same brand and service, we counted how well the previous round's mention status predicted the next.

Previous round	Own-domain citation this round	Mentioned this round
Mentioned	Yes	51/52 (98%)
Mentioned	No	17/18 (94%)
Not mentioned	Yes	2/2
Not mentioned	No	2/36 (6%)

Overall, 97% (68/70) of cells mentioned in the previous round were mentioned again, versus 11% (4/38) of cells that had not been mentioned. For the 20-day transition alone (last July round → first August round), the figures were 23/24 and 0/12.

The most notable result: own-domain citation status changed 15 times between rounds, and in 14 of those the mention status stayed the same. Brands that were already mentioned stayed mentioned even when their own domain dropped out of the sources, and brands that were not mentioned mostly stayed unmentioned even when their own domain was cited. This is the same structure as in our August re-measurement analysis, where mention rates held steady while own-domain citation rates moved sharply.

Read this re-analysis with three conditions in mind.

1. The unit of analysis is different. The study works at the single-prompt level; this re-analysis uses a service-level judgment across the whole question set. Grouping questions dampens round-to-round fluctuation, so persistence looks higher. That is one reason the high-prior, no-exposure cell is 29–32% in the study and 94% here.
2. The sample is small and uneven. All 12 companies are well known in their categories, and the "not mentioned before, cited this round" cell has only two cases. Read the direction, not the magnitudes.
3. Repeated measurements of the same cells are not independent observations. No significance testing was performed.

Why this happens — and why it should not be read as causation

"Get your own site cited and your mention rate rises tenfold" is not a conclusion this data supports. The author explicitly states that the paper does not claim inserting a company's URL into a source list would causally increase mentions tenfold. There are two reasons.

• The direction may be reversed. The AI may first decide which brands to include and then pull their official sites as sources to confirm them. In that case, the own-domain citation is a result of the mention, not its cause.
• A third factor is likely. Brands that are already widely discussed in a category tend to have official sites that surface well in search, and they also tend to appear in answers. The brand's existing standing raises both metrics at once. A study finding that large brands appeared in 73% of answers and small brands in 11% is covered in our brand tier analysis.

Even so, the association is valuable as a diagnostic. In the author's words, it is still "diagnostically important" even if it is not causal. Three practical takeaways follow.

1. A mention behaves less like a fresh event each run and more like an accumulated state. Once a brand starts being mentioned for a question, it tends to stay; where it is not mentioned, it tends to stay absent. That makes the first entry — turning an unmentioned question into a mentioned one — the hardest and most important step.
2. Own-domain citation and mention answer different questions. The author put it this way:

"In practical monitoring, a cited-source dashboard and a brand-mention dashboard answer different questions." — Tannenbaum, arXiv:2609.23162

1. Your own site is not the only path that brings your brand into an answer. The author describes own-domain citation as an incomplete measure of evidence exposure and notes that third-party pages can carry the brand into the answer and may matter more than owned pages (in the original, "A third-party page can also carry the brand into the answer."). Publishing more of the same pages does not always solve the problem.

Putting it into practice — decide the next action from four mention × citation states

Rather than merging mention rate and own-domain citation rate into a single score, split each question and AI service into states based on the combination of the two. The results screen of RanketAI's AI brand visibility analysis includes a "Brand mention × Source citation analysis" card built on the same criteria.

State	Mention	Own-domain citation	Meaning	Next action
Fully visible	Yes	Yes	Appears in the answer, backed by your own page	Re-measure to confirm it holds; refresh figures and dates on the supporting page regularly
Used	Yes	No	Appears, but backed by third-party or competitor documents	Strengthen your own supporting page — a comparison table using the same criteria as competitors
Ghost Citation	No	Yes	Your page is used as evidence, but your name is missing from the answer	State the brand name and product category together in the page body
Not visible	No	No	No signal from you in this answer's sources	Secure an exposure path — the four-step procedure below

The reasoning behind each action:

• Fully visible: in our re-analysis, 98% (49/50) of cells in this state kept the mention in the next round, and 86% (43/50) stayed fully visible. The job here is maintenance, not new work. How citations shrink for pages that are not refreshed is covered in our citation decay post.
• Used: the mention holds, but third-party documents decide how your brand is described. In the cold-start experiment, an own-domain listicle went from 4 to 49 mentions after a competitor comparison section was added (cold-start experiment analysis). A table comparing vendors on the same criteria worked as a supporting page better than a "we're number one" list.
• Ghost Citation: your page's content is used, but your brand does not make it into the answer. This state was rare in our re-analysis, but for a brand entering a category for the first time it may be the step just before a mention. Causes and checks are covered in our ghost citations post.
• Not visible: in the study, the mention rate was around 3% when there was no signal from the brand in the sources, and in our re-analysis 92% (34/37) of cells in this state were still not visible in the next round. This state rarely changes on its own. The paper notes that in this situation, adding more content on the same domain may not remove the bottleneck, and distribution-side work such as indexing, crawlability, and third-party coverage may be needed.

Four steps to secure an exposure path from the "not visible" state

A recommendation like "increase third-party exposure" only becomes actionable when it is broken into steps.

1. Identify — For your target questions, find the sources that cite competitors but not you, separately for each AI service. Measuring the same questions side by side with a tool such as competitor comparison produces the list. At the same time, use a page structure check to confirm that AI crawlers can read your pages. If they cannot, your pages cannot enter the source list.
2. Classify — Sort the sources into three groups.

Group	Examples	How to act
Directly publishable (UGC)	Naver Blog, Brunch, WikiDocs	Create an account and publish posts rewritten for each channel. No cost, can start right away
Press and contributed articles	Business press, startup media, industry blogs	Pitch press releases or contributed articles when you have a launch or research result. Takes time and relationships
Structurally closed	Other companies' blogs, public institutions	Exclude from the target list

When RanketAI classified the 15 sources that cited only competitors in one comparison measurement, the shares were 20%, 33%, and 47%. This was a single category and a single measurement, so the ratios do not generalize. The criteria and examples are covered in our source gap analysis.

1. Execute — Start with the no-cost, directly publishable channels, and bundle press and contributed articles around results worth announcing.
2. Re-measure — Re-run the same questions under the same conditions to confirm whether the state has moved out of "not visible." Do not judge from a single result; compare against your same-day re-measurement range. How many runs you need is covered in our measurement statistics post.

Four lines to hold when citing this study

Use the study's figures as evidence of direction, not as evidence of magnitude.

1. There is a conflict of interest. The author is the founder and CEO of a commercial AI-search measurement software company and states a financial conflict of interest in the paper. The data also come from that tool's customer projects.
2. These are operational data, not an experiment.

"The datasets were collected for operational research rather than designed as one preregistered experiment." — Tannenbaum, arXiv:2609.23162, limitations

1. Some cells are small, and the market mix is not disclosed. The branded-search-query cells contain only dozens of observations, and in the real-prompt cohort only 20 of 80 unique prompts triggered an observed search query. The projects' industry and language mix is only partly described, so there is no evidence that the results carry over unchanged to Korean-language questions or Korea's domestic AI services.
2. A mention is not a recommendation, click, or purchase. The author notes that a brand mention is not equivalent to a positive recommendation, rank, click, lead, or purchase, and says the paper should be read as a dated measurement of systems observed in 2026. The models measured may also differ from each service's current default model.

FAQ

Related reading

• AI Picks Its Shortlist Before It Searches — 72 Queries Across 12 Korean Categories — the study's "branded search query" signal observed in Korean-language questions.
• Monthly Change Was Smaller Than Same-Day Measurement Noise — a monthly record in which mention rate and own-domain citation rate moved independently.
• Brand Tier Decides How Often AI Names You — 73%, 44%, 11% — the effect of brand standing, which connects to prior history.
• The Source Gap — Finding the Third-Party Sites That Cite Only Your Competitors — criteria for classifying and acting on third-party sources.
• From Zero to 298 AI Mentions in 30 Days — a first-entry case for a brand with no prior history.