The practical brief
The finding
Across 34,960 unbranded GPT and Gemini observations, brand mentions were rare when neither the brand’s own domain nor a branded fan-out appeared in the observable live retrieval path, but much more common when own-domain exposure appeared, and highest when own-domain exposure and branded fan-out appeared together. A rolling holdout model predicted mentions best when it combined prior run history with these contemporaneous live signals. [c1] [c4] [c6] [c7] [c8]
Why it matters
For a business owner, that means a page-quality audit alone can miss the actual failure point. This paper treats visibility as stages: prompt-to-page match, evidence exposure, and answer selection, plus a separate prior-compatible path where a brand can appear without its own domain being cited. If you collapse those stages into one score, you can misdiagnose where budget should go.
| Observed live-evidence state | GPT | Gemini |
|---|---|---|
| Neither signal present | 2.8% (438 of 15,524 observations) | 3.8% (524 of 13,801 observations) [c1] |
| Own domain only | 49.0% (866 of 1,769 observations) | 58.4% (1,995 of 3,415 observations) [c1] |
| Branded fan-out only | 64.4% (38 of 59 observations) | 84.8% (28 of 33 observations) [c1] |
| Own domain and branded fan-out | 91.4% (117 of 128 observations) | 100.0% (231 of 231 observations) [c1] |
What to try
BLURSOR’s practical interpretation
Our interpretation: for high-value unbranded prompts, track at least three things separately by engine—prompt-to-page match, whether your own domain appears in the engine’s stored source list, and whether the answer mentions you. If match looks decent but own-domain exposure stays low, treat that as a diagnosis to investigate before commissioning another rewrite. The paper does not test which fix would raise exposure.
Study boundary
The study is observational and explicitly not causal. It was assembled from operational research datasets rather than one preregistered experiment; observed fan-out depends on instrumentation and hidden retrieval could weaken 'no fan-out' interpretations; own-domain citation is only a partial proxy because third-party pages may matter more; the fan-out cohort is small and not population-representative; one page-match analysis uses a later site crawl than the original visibility export; the data cover limited verticals and two engines observed in 2026; brand mention is not recommendation, rank, click, lead, or purchase; proprietary raw data were not released for independent refitting; and the author discloses being founder and CEO of Aiso Boost Ltd., which sells AI-search measurement software.
Decide whether the problem is content or exposure
If your brand is absent from AI answers, this paper suggests you should check measurement before assuming the page itself is the whole problem. The author frames visibility as a sequence: does a page match the request, does the engine expose supporting evidence, and then does the answer actually select the brand?
Two paper terms matter. 'Own-domain exposure' means the brand’s own domain appeared in the engine’s stored source list for that run. 'Branded fan-out' means an unbranded prompt was expanded into a search query containing the brand name. These are observed diagnostics from the run, not proven levers you can directly control.
The model result is also narrower than a forecasting score might imply. AUC is a discrimination metric, meaning how well a model ranks likely mentions above non-mentions. Here the strongest AUC came from a rolling diagnostic setup that used current-run signals, not an advance prediction for unseen prompts.
- A strong page can still lose if the engine never surfaces supporting evidence.
- A single 'AI visibility score' can hide whether the failure is match, exposure, or answer selection.
- The holdout test is diagnostic on monitored runs, not a pre-run guarantee.
Observed retrieval states aligned with large mention gaps
The clearest result is the paper’s four-state ladder for unbranded prompts. When neither own-domain exposure nor branded fan-out was observed, mentions were only 2.8% on GPT and 3.8% on Gemini. With own-domain exposure alone, rates rose to 49.0% and 58.4%. When both signals appeared, mention rates reached 91.4% on GPT and 100% on Gemini in this panel.
That pattern also persisted within repeated organization-prompt cells. When the same prompt for the same organization varied across runs, own-domain exposure still had a strong association with mentions. The paper is explicit that this remains observational, but it makes exposure hard to ignore in reporting.
- No observed live signal: mentions were uncommon on both engines.
- Own-domain exposure alone coincided with much higher mention rates.
- Exposure plus branded fan-out matched the highest rates in the dataset.
Past visibility and current evidence each carry signal
The paper does not reduce visibility to retrieval alone. It also models a prior-compatible path: brands can appear without their own domain being cited, potentially because of model familiarity, third-party evidence, or other unobserved information. The data cannot separate those mechanisms, so the paper keeps them bundled as prior rather than claiming memory as the cause.
In repeated runs, prior history and current exposure complemented each other. In low-prior GPT states, mention rates were 0.9% without own-domain exposure and 25.1% with it. In high-prior states, they were 32.0% and 82.0%. Gemini showed the same structure. That means established brand familiarity does not make current exposure irrelevant.
- Low prior is not the same as no opportunity.
- High prior does not remove the value of current observed exposure.
- Full models beat prior-only and live-only versions on the latest 30% holdout.
What this study can guide, and what it cannot
Our interpretation is to use this paper as a diagnostic framework, not as proof that any specific optimization will cause citations or recommendations. The page-match result itself was engine-specific: prompt-page match predicted Gemini exposure more clearly than GPT, which is one reason blended reporting can mislead.
The limitations are substantial and decision-relevant. The datasets came from operational research rather than one preregistered experiment. Observed fan-out depends on instrumentation, so hidden retrieval could make 'no fan-out' states less clean than they look. Own-domain citation is an incomplete exposure proxy because third-party pages may matter more. One page-match analysis used a later site crawl than the visibility export, so it is retrospective rather than contemporaneous.
The study is also bounded in scope. It covers limited verticals, two engines, and systems observed in 2026 rather than stable engine behavior. Brand mention is not the same as recommendation quality, rank, traffic, leads, or sales. And outside researchers cannot independently reconstruct the models from raw client-level data because those records were not released. The author further discloses being founder and CEO of Aiso Boost Ltd., a seller of AI-search measurement software.
- Measure match, own-domain exposure, and mention separately by engine.
- Treat low exposure as a diagnosis to investigate, not proof of a specific remedy.
- Do not read these results as guaranteed business impact.
The source and evidence
From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search
Study limitations and disclosures
- "First, the datasets were collected for operational research rather than designed as one preregistered experiment." (Section 12)
- "Second, the real-prompt fan-out cohort is small. Only 20 of 80 unique prompts trigger observed fan-out, and the intent mix is not population-representative." (Section 12)
- "Third, fan-out visibility depends on instrumentation. A prompt with no recorded search_query is treated as not triggering an observed fan-out. Hidden retrieval actions would weaken that interpretation." (Section 12)
- "Fourth, the Organization A page-match crawl is temporally later than the engine visibility export. The analysis should be read as a retrospective coverage association." (Section 12)
- "Fifth, own-domain citation is an incomplete measure of evidence exposure. Third-party pages may mention the brand and can be more important than owned pages." (Section 12)
- "Finally, all model and search behavior is time-dependent. The paper is a dated measurement of systems observed in 2026, not a claim about permanent engine internals." (Section 12)
- "The proprietary raw Aiso conversation corpus and client-level monitoring records are not included because privacy and commercial data-governance constraints do not permit redistribution." (Appendix A Reproducibility Notes)
Evidence behind this briefing
[c1] For unbranded prompts in the large validation panel, brand mention rates rose sharply across live-evidence states, with the highest rates when both own-domain exposure and branded fan-out were present.
Section 7.1, Table 3 · Read the study
[c2] Within repeated prompt cells, own-domain exposure remained strongly associated with mention odds even after holding organization and prompt fixed, but the result is still observational.
Section 7.2 · Read the study
[c3] Historical prior visibility and current exposure jointly stratified next-run mention rates, showing that prior and live evidence were complementary rather than interchangeable.
Section 7.3 · Read the study
[c4] The chronological holdout models predicted mention better when combining prior history with live evidence than using either alone.
Abstract · Read the study
[c5] Prompt-page match was more predictive for Gemini than GPT, indicating engine-specific differences in how upstream relevance translates into exposure.
Abstract · Read the study
[c6] The fitted visibility model used a chronological logistic specification with prior-run history, current own-domain exposure, branded fan-out, and request-intent controls, and was explicitly framed as predictive rather than causal.
Section 6, Equation 6 · Read the study
[c7] The paper states that the fitted equation should not be interpreted causally or as a description of engine internals.
Abstract · Read the study
[c8] The author discloses a financial conflict of interest because the author leads a company that sells AI-search measurement software.
Data Governance and Conflict of Interest · Read the study