Ask ChatGPT or Perplexity “what are the best tools for this job” without naming your company. Does your brand show up? A large-scale measurement published in June 2026 gives an uncomfortable answer: across 102 tracked brands, niche and emerging brands appeared in just 11.4% of unbranded answers (source: Generative Engine Optimization at Scale, Pratyush Kumar, June 2026).
This article unpacks two public studies and one set of platform data to answer four questions: where the real baseline sits, how far apart the engines are, which prompt types actually reveal the gap, and which sources AI engines cite. By the end you will have a repeatable way to build your own baseline instead of guessing that you are “probably fine” in AI search.
The baseline is a staircase, not a slope
AI visibility does not fall on a smooth curve. Across 102 brands, the three tiers of brand maturity sit roughly 30 percentage points apart from each other.
Between March and May 2026, the study ran 3,508 tracked jobs across five engines — ChatGPT, Gemini, Perplexity, Claude and Grok — collecting 102,025 prompt responses and 149,912 source citations. Grouping brands into three tiers and measuring how often each appeared in unbranded answers produced this ladder:
| Brand tier | Appearance rate (unbranded prompts) | Sample | 95% CI |
|---|---|---|---|
| Tier 1 — global household names | 72.9% | 11 | 60.1% – 84.2% |
| Tier 2 — established mid-market | 43.6% | 36 | 36.4% – 50.9% |
| Tier 3 — niche and emerging | 11.4% | 55 | 4.2% – 20.3% |
All pairwise differences were statistically significant, with Cohen's d between 1.31 and 2.34. The practical takeaway for most companies: the number to benchmark against is 11.4%, not 72.9%. Moving from 11% toward 40% is a workable goal. Benchmarking against tier-one brands mostly produces the conclusion that nothing can be done.
Five engines, four-fold spread
The gap between engines is wider than most teams assume. On unbranded prompts the same cohort scored 51.5% on the strongest engine and 12.0% on the weakest.
| Engine | Branded prompts | Unbranded prompts | All prompts |
|---|---|---|---|
| ChatGPT | 94.2% | 22.1% | 28.6% |
| Gemini | 94.0% | 18.7% | 26.6% |
| Perplexity | 98.5% | 23.9% | 30.9% |
| Claude | 100.0% | 51.5% | 54.6% |
| Grok | 100.0% | 12.0% | 24.1% |
One caveat the authors raise themselves: the Claude column is inflated because its tracked cohort was smaller and skewed toward larger brands. Read it as a sampling artifact, not as evidence that one engine favors small brands. That habit — check the sample before reading the number — is worth applying to every AI visibility leaderboard you encounter.
The operational implication is blunt: checking one engine tells you nothing about the others. If you found yourself in Perplexity, that says little about Gemini. For why engines diverge in the first place, see how AI search picks its sources.
How you ask decides the score
Prompt type predicts visibility better than brand strength does. The same brands scored 97.0% on branded research prompts and 22.9% on discovery prompts.
| Prompt category | Average appearance rate | What it tells you |
|---|---|---|
| Brand research (names the brand) | 97.0% | Near-certain mention; least informative |
| Comparison (A vs B) | 74.6% | You are already in the consideration set |
| Discovery (what tools exist) | 22.9% | Closest to real acquisition behavior |
| Expert (what would a specialist pick) | 15.3% | Tests depth of content |
| Use case (what to use for X) | 11.4% | Tests content-to-scenario fit |
| Problem/solution (how do I fix X) | 10.8% | Hardest, and the most valuable |
This explains why a first self-check so often produces a falsely reassuring result: the team typed “tell me about [our brand]” and the model obliged. The real test is asking, without naming anyone, which providers do this work — and seeing whether you make the list. That distinction is the most commonly missed part of how GEO differs from SEO.
Where the citations actually go
Citation sources are highly concentrated. Of 149,912 citations, roughly 78% pointed to corporate websites, and the single most-cited content format was the ranked “best-of” listicle at about 21% of all citations.
Corporate websites here include both a brand's own pages and pages published by other companies in the same category — meaning competitors' comparison pages are also feeding the answers about you. Among non-corporate sources, YouTube led, followed by Reddit, editorial media and Wikipedia (source: Generative Engine Optimization at Scale).
So which writing choices earn citations? GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) ran controlled experiments on GEO-bench, a set of 10,000 queries spanning roughly 25 domains, and reported:
- Adding authoritative quotations: about 27% improvement in visibility
- Adding statistics: about 26%
- Citing sources explicitly: about 25%
- Best combinations lifted visibility by up to about 40%; on Perplexity in production, subjective metrics improved by about 37%
- Classic SEO keyword stuffing underperformed in generative engines
These findings and the citation distribution point the same way: engines prefer passages that carry a source, a number and a clear structure. The mechanics are covered in writing content AI engines will quote.
The Chinese-language gap you have to measure yourself
Both studies cover English-language engines. No public benchmark of comparable scale exists yet for Chinese AI search, so that side has to be measured first-hand.
Scale is not the obstacle. Per QuestMobile's Q1 2026 review of AI applications, Doubao, Tongyi Qianwen and DeepSeek reported roughly 340 million, 170 million and 130 million monthly active users respectively (source: QuestMobile Q1 2026 AI application insights). The audience is there; what is missing is a cross-brand benchmark comparable to the 102-brand study.
Until one exists, borrow the method and generate your own numbers: build the same six prompt categories, separate branded from unbranded prompts, and log which sources each answer cites. Platform-specific behavior is covered in optimizing for DeepSeek.
Turning a benchmark into your own baseline
A benchmark is only a reference point. What changes decisions is your own baseline, re-measured on a fixed schedule. Four steps build one.
- Fix a prompt set. Write 5–10 prompts per category above, with at least 70% of them unbranded. Once set, freeze it — changing prompts destroys comparability.
- Fix engines and cadence. Cover at least three engines, keep language versions separate, and run at the same time each week.
- Record three things every run: whether you were mentioned, which sources the answer cited, and whether the framing was positive or neutral. The second is often more actionable than the first — it tells you exactly which page to go build.
- Change one variable at a time and wait one to two weeks before re-measuring. Note that the study found sentiment framing flips about 6.7× more often than mention itself, so never read a single-run sentiment change as a trend.
This can be run by hand or handed to a tool. AiVisible does exactly this: it runs a fixed prompt set across multiple platforms on a schedule and records mention rate, cited sources and competitor comparisons as a comparable time series. To see where you currently stand, run a free brand AI visibility check; the full manual procedure is in how to run an AI visibility audit.
One closing note. What makes 11.4% useful is not that it is low — it is that it is measurable. In AI search, the distance between an invisible brand and a cited one is usually less about budget than about whether anyone is watching the curve at all.
FAQ
What counts as a good AI visibility score?
There is no single pass mark; it depends on brand tier. Across 102 brands measured on unbranded prompts, global names averaged 72.9%, mid-market brands 43.6%, and niche or emerging brands 11.4%. Most small and mid-sized companies should benchmark against 11.4%, not the top tier.
Why does AI seem to know my brand well when I check it myself?
Almost always because the prompt named the brand. Branded research prompts returned a 97.0% mention rate, while unbranded discovery prompts returned 22.9%. A branded self-check passes by default and reveals very little.
How many engines should I monitor?
At least three, with language versions tracked separately. The same brands scored between 12.0% and 51.5% on unbranded prompts depending on the engine, so a result from one engine does not generalize to the others.
Which content format gets cited most by AI engines?
Of 149,912 citations analyzed, roughly 78% went to corporate websites, and the single most-cited format was the ranked best-of listicle at about 21% of all citations. Among non-corporate sources, YouTube led, followed by Reddit, editorial media and Wikipedia.
How often should I re-measure?
Run the same frozen prompt set weekly at a consistent time, and wait one to two weeks after a content change before reading results. Sentiment framing was found to flip about 6.7 times more often than mention itself, so a single-run sentiment swing is not a trend.