The whole study, in 60 seconds.
If you only read one page of this report, read this one. Every number below is unpacked in the sections that follow.
- 11%
ChatGPT. Brand overlap between UI and API answers across 834 matched prompts. Same first brand: 13%.
- 12%
Gemini. Brand overlap across 895 matched prompts. Same first brand: 16%.
- 6%
Perplexity. Brand overlap across 836 matched prompts. Same first brand: 9%. The widest gap we measured.
Two channels, two different realities
If your AI visibility monitoring runs on the official APIs, it is measuring a distribution of brands that barely overlaps with what consumers are shown.
Across 2,565 matched prompt pairs, the average overlap between the brand set in a UI answer and the brand set in the matching API answer was 6 to 12 percent, depending on the engine. The first brand mentioned, the one that wins the recommendation, matched in only 9 to 16 percent of answers.
Brand overlap and same top brand, UI vs API
| Engine | Brand overlap | Same top brand |
|---|---|---|
| ChatGPT | 11% | 13% |
| Gemini | 12% | 16% |
| Perplexity | 6% | 9% |
Where the channels diverge
ChatGPT
ChatGPT: UI vs API
GPT-5.5 API vs chatgpt.com · 834 matched prompts
| Metric | UI | API |
|---|---|---|
| Brands per answer | 4.7 | 5.1 |
| Citations per answer | 15.2 | 4.4 |
| Zero-citation answers | 6% | 17% |
Top-25 domain overlap (UI ∩ API): 9%. Brand overlap 11%. Same top brand 13%.
- PlatformFans see it in the UI, the API top-25 misses it
- PlatformThe API surfaces it, the UI top-25 does not
ChatGPT
- Reuters
- Goal.com
- NerdWallet
- The Guardian
FS1ZelleFOX OneOptaOn Location
The starkest difference is sourcing: the real ChatGPT product reads 3.5x more sources per answer than its API, and leans on reddit.com (284 answers) and goal.com, while the API's most-cited domain after fifa.com is ftc.gov. Different reading lists, different recommendations.
Gemini
Gemini: UI vs API
Gemini API + Search grounding vs gemini.google.com · 895 matched prompts
| Metric | UI | API |
|---|---|---|
| Brands per answer | 7.3 | 3.2 |
| Citations per answer | 5.1 | 13.2 |
| Zero-citation answers | 24% | 8% |
Top-25 domain overlap (UI ∩ API): 52%. Brand overlap 12%. Same top brand 16%.
- PlatformFans see it in the UI, the API top-25 misses it
- PlatformThe API surfaces it, the UI top-25 does not
Gemini
- American Express
- Google Wallet
- Opta
YouTube TVThe GuardianFuboFS1
Gemini inverts the pattern: the consumer UI names 2.3x more brands per answer than the API, and one in four UI answers ships with no visible sources at all. Gemini's grounded API reads broadly (reddit.com appears behind 436 answers) but compresses to fewer named brands.
Perplexity
Perplexity: UI vs API
Sonar Pro API vs perplexity.ai · 836 matched prompts
| Metric | UI | API |
|---|---|---|
| Brands per answer | 2.4 | 3.8 |
| Citations per answer | 10.0 | 7.8 |
| Zero-citation answers | 0% | 0% |
Top-25 domain overlap (UI ∩ API): 22%. Brand overlap 6%. Same top brand 9%.
- PlatformFans see it in the UI, the API top-25 misses it
- PlatformThe API surfaces it, the UI top-25 does not
Perplexity
- Adidas
- Fanatics
- Puma
- Roku
- Booking.com
OptaYouTube TVUberAirbnbFubo
The closest product to its API in spirit, yet the widest brand gap of all three. The API's source mix is dominated by social (facebook.com behind 469 answers, instagram.com behind 401); the consumer product leans on youtube.com and editorial sources. Adidas, Fanatics and Puma reach fans through the UI and are invisible in the API's top brands.
Five takeaways for anyone measuring AI visibility
- API-based monitoring is a different instrument, not a cheaper one. With 6 to 12 percent brand overlap, an API readout is not a noisy version of the consumer answer. It is a different answer.
- The recommendation winner flips in roughly 6 of 7 answers. If your dashboard says you own a category, there is an 85 percent chance the interface your buyers use disagrees.
- Sourcing behavior, not model quality, drives the gap. ChatGPT's UI reads 3.5x more sources than its API. Gemini's API reads more but names fewer brands. The retrieval layer is the product decision.
- Reddit is the consumer layer's hidden referee. reddit.com is a top-3 cited domain in the UI channel of all three engines. Brands with no Reddit presence are negotiating with a judge they cannot see.
- Affiliates win the interface. NerdWallet, Goal.com and Booking.com appear in consumer answers where the APIs name none of them. The recommendation economy has already moved inside the answer.
Methodology and caveats
900 unbranded World Cup 2026 prompts across 30 commercial, predictive and informational categories.
API channel: official developer APIs with web search or grounding enabled (GPT-5.5, Gemini with Search grounding, Perplexity Sonar Pro), collected June 11, 2026. UI channel: the live consumer interfaces (chatgpt.com, gemini.google.com, perplexity.ai), collected June 12, 2026 via scraping infrastructure. Brand extraction by LLM structured extraction, identical for both channels, manually QA-sampled.
- Runs are one day apart; some divergence is news cycle, not channel. It cannot explain a 90 percent disagreement.
- Model identities differ by design: the UI serves whatever the consumer product runs. That difference is part of the finding, and is exactly why API proxies mislead.
- Extraction noise (~5 percent) applies to both channels equally and washes out of the comparison.
- All raw responses are archived and every number in this study is recomputable.
Written by Sariel Mazuz, Director of Product Marketing at Brandlight · LinkedIn




