Brand Visibility in AI Search Has a Staircase, Not a Switch: New Data on Brand Mentions

If you are tracking brand visibility in AI search as a single number, you are measuring the scoreboard and ignoring the game. New research published on arXiv on 19 September 2026 by Benjamin Tannenbaum, covering 34,960 unbranded prompt-engine observations across GPT and Gemini, shows that getting recommended is not one event. It is a staircase, and almost everything depends on the first step.

Here is the practical version for a SaaS growth team. When neither the brand nor its own domain appeared in the observable live retrieval path, mention rates were 2.8% on GPT and 3.8% on Gemini. When the brand's own domain was cited, those rates rose to 49.0% and 58.4%. So the job this quarter is not winning the recommendation slot. It is getting into the retrieval path at all.

What the study measured, and over what period

The paper, From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search, analyses 2,854 distinct monitored prompts across 75 anonymised projects, with repeated GPT and Gemini runs from June to September 2026. All prompts were unbranded, meaning the prompt itself never named the target brand.

That design matters. Branded prompts tell you almost nothing, because the model already has the name. Unbranded prompt coverage tells you whether a buyer who does not yet know you will meet you.

The author is explicit that the model is predictive and observational, not a causal description of how proprietary engines work internally. We would hold that line too. The pattern is strong enough to plan around, but nobody outside OpenAI and Google can say why.

The three stages, and the size of each jump

The study separates two live retrieval signals: whether the brand's own domain appears in the retrieval path (own-domain exposure), and whether the engine issues a branded follow-up query (branded fan-out). Combining them produces three clearly different worlds.

Stage	Mention rate
No brand or domain in retrieval path (GPT)	2.8%
No brand or domain in retrieval path (Gemini)	3.8%
Own-domain citation, no branded fan-out (GPT)	49.0%
Own-domain citation, no branded fan-out (Gemini)	58.4%
Own-domain exposure plus branded fan-out (GPT)	91.4%
Own-domain exposure plus branded fan-out (Gemini)	100%

The first jump is the expensive one and the valuable one. Moving from invisible to cited is worth roughly 46 percentage points on GPT and 55 on Gemini. Moving from cited to fan-out adds another 42 points on GPT.

The effect also held within the same project, prompt and engine across repeated runs. Among prompt cells that varied in own-domain exposure while branded fan-out stayed absent, exposure was associated with a mean mention-rate increase of 40.2 points on GPT and 49.0 points on Gemini. That is not a comparison between strong brands and weak ones. It is the same prompt behaving differently depending on whether your page made it into retrieval.

Why early absence is more expensive than it looks

Prior visibility persisted independently of current signals. A previous non-mention combined with no current own-domain exposure produced next-run mention rates of 1.6% on GPT and 1.9% on Gemini. A previous mention plus current exposure produced 80.5% and 83.7%.

Read that as a compounding curve, not a snapshot. Invisibility is sticky. If your category prompts have been answering without you all year, the cost of entry goes up, not down.

Measuring brand visibility in AI search: what to track instead of citations

Most dashboards we see count domain citations. That is one of three stages, and on its own it undersells and overstates in different places. We would track brand visibility in AI search as a funnel with four numbers, reported monthly to the same committee that sees pipeline.

What to measure at each stage of AI-search visibility, and who should own it

Stage	Metric	Owner
Retrieval entry	Share of monitored unbranded prompts where your domain appears in the retrieval path	Head of SEO/GEO
Mention	Unbranded prompt coverage: share of prompts where the brand is named in the answer	Head of SEO/GEO
Recommendation	Share of mentions where the brand is presented as a recommended option	Content lead
Commercial effect	Branded search volume and direct demo requests from AI-referred sessions	Growth lead

The model in the paper reached AUC 0.963 on GPT and 0.942 on Gemini on the latest 30% holdout using prior history plus live retrieval signals. Prior history alone reached 0.937 and 0.917. Live signals alone reached 0.880 and 0.840. History and current retrieval each carry real information, which is exactly why a single monthly citation count is a weak instrument. If measurement design is the bottleneck, our measurement and attribution advisory is built for this problem.

Where GPT and Gemini diverge, and what that means for content

A separate 199-prompt page-corpus validation found that prompt-page match predicted Gemini exposure with an AUC of 0.641, against 0.545 for GPT. In plain terms, writing a page that closely matches the prompt is a more reliable route into Gemini retrieval than into GPT retrieval.

That places content relevance upstream of a larger engine-mediated exposure effect. You control relevance. You do not control selection. So build pages that answer specific unbranded buying questions, then accept that GPT will convert fewer of those into retrieval than Gemini does.

We would not spend budget chasing engine-specific tricks for Perplexity, Copilot or Google AI Overviews on the back of this paper. It covers GPT and Gemini only, from June to September 2026. Treat the staged pattern as a working model elsewhere and verify it with your own prompt set before you fund anything.

What to do about brand visibility in AI search this quarter

Brand visibility in AI search should be planned as a sequence, and the first action is to find out where you actually sit. Build a monitored set of 50 to 150 unbranded prompts that mirror how buyers describe your problem, not your product, and run them weekly on ChatGPT and Gemini. Then fix retrieval entry before you touch tone, framing or recommendation language, because a 2.8% base rate cannot be improved by better copy.

If you want an outside read on where your domain currently enters the retrieval path and what that means for your brand visibility in AI search, start with our free AI visibility audit, or see how we structure organic search, AEO and GEO as one programme. For a related engine-level shift worth reading next, see our analysis of Gemini reading your Google Business Profile.

Cover photo by Steve A Johnson on Unsplash