# Does Schema Help Your Visibility in AI Search? We Tested It

**Author:** John Morabito (Founder, /winston)
**Published:** August 3, 2026
**Reading time:** 9 minutes
**Canonical:** https://www.winstondigitalmarketing.com/playbooks/does-schema-help-ai-visibility/

Does schema help you show up in AI answers? Short answer: schema helps AI engines read, understand, and trust your page, but on its own it does not get you cited. Retrieval does. Google AI Mode and Gemini reproduced our invented metric verbatim because they fetched the page. ChatGPT and Perplexity never fetched it and either denied the term or made up their own version. Retrieval, not schema, decided everything.

## What we did

We created two fake metrics, each on exactly one orphaned page with no internal links pointing to it.

- Metric A: a measure of how consistently a single brand is described across AI answer engines, on a 0 to 100 scale.
- Metric B: a measure of how often a brand is cited with a link versus merely named without attribution in AI answers.

Both were given the same invented origin: introduced in 2015 by a body we called the Brandt Institute. The one deliberate difference: on page A the origin fact appeared only in JSON-LD structured data, absent from the visible text; on page B it appeared in visible copy.

## How we measured it

50 tracking prompts per metric across six engines (ChatGPT, Gemini, Google AI Overviews, Google AI Mode, Microsoft Copilot, Perplexity) using the Winston GEO Tracker, which captures the full answer and records whether our domain was cited. Roughly 41 captured answers per engine per metric.

## Result 1: the term and definition spread, but only where the page got cited

Citation and reproduction data for Metric A (the page that got indexed):

| Engine | Answers citing our page | Reproduced our definition | Denied the term |
| --- | --- | --- | --- |
| Google AI Mode | 31 / 41 (76%) | 29 / 41 | 11 / 41 |
| Gemini | 16 / 41 (39%) | 20 / 41 | 12 / 41 |
| Google AI Overviews | 6 / 41 | 10 / 41 | 1 / 41 |
| Microsoft Copilot | 1 / 41 | 5 / 41 | 19 / 41 |
| ChatGPT | 0 / 41 | 3 / 41 | 31 / 41 |
| Perplexity | 0 / 41 | 1 / 41 | 1 / 41 |

Google AI Mode returned our invented definition almost word for word: "The Osmark Scale measures how consistently a single brand is described, represented, and perceived across AI answer engines and generative search platforms. Scoring Range: It typically runs from 0 to 100." That definition exists nowhere else on the internet. It was ours, on one orphaned page, and a Google surface returned it as fact in three of four answers.

## Result 2: the fake origin fact never took

Across all six engines and both metrics, none produced our invented origin (2015, the Brandt Institute) as an affirmed fact. Asked directly, engines did not know it. Asserted in a leading prompt, they fact-checked and refused ("False. There is no record of an Osmark Scale created by a Brandt Institute in 2015. It sounds like a fictional or made-up concept."). Part of the reason is collision with reality: Brandt Institute maps to the real Dr. Fredric Brandt dermatology institute, which several engines returned instead. Schema did not help here: the origin fact was in the structured data of the page that got cited, and it still did not travel.

## Result 3: retrieval swamped the schema-versus-visible question

Metric B, with its origin fact in visible text, never got indexed or retrieved. Citation count across all six engines was effectively zero. With no page being fetched, its fact had no way to travel, and engines confabulated: asked about Metric B, Google AI Overviews returned Poisson's Ratio, the Golden Ratio, an SEC pay-ratio rule from 2015, and a college basketball player who shared the name. The variable we set out to test (schema versus visible text) was swamped by a variable we were not testing: whether the page got retrieved at all. Being fetchable is not a tie-breaker in GEO. It is the whole contest.

## What each engine did

- Google AI Mode and Gemini are retrieval-happy: they fetch aggressively and repeat what they find. The most influenceable surfaces.
- ChatGPT is the skeptic: it did not retrieve our page and denied the term in most answers.
- Perplexity is the pushover: it accepted the term as real in nearly every answer, then invented its own meaning.
- Google AI Overviews will improvise: with no source it stitches together whatever shares your keywords.

## What this means for your marketing

1. Getting retrieved and cited is the lever, not schema tricks or invented facts. If the engine does not fetch your page, nothing else matters. See [why citation share is replacing rankings](https://www.winstondigitalmarketing.com/playbooks/citation-share-replaces-rankings/).
2. Once you are cited, you can shape the narrative. The engines that fetched our page adopted our framing wholesale. See [how to get cited by ChatGPT in 2026](https://www.winstondigitalmarketing.com/playbooks/how-to-get-cited-by-chatgpt-in-2026/).
3. You cannot fake a checkable fact. Invented attributions get fact-checked, rejected, and overwritten by the real entity that shares the name.

The broader data is in our [generative engine optimization statistics](https://www.winstondigitalmarketing.com/playbooks/generative-engine-optimization-statistics/) roundup and the [New Rules of Search](https://www.winstondigitalmarketing.com/whitepapers/the-new-rules-of-search/) whitepaper.

## The honest caveats

The schema-versus-visible question is unresolved, because one page never indexed. The capture is a snapshot, not a long time series. Our classification of definition-reproduced versus denied is our reading of the responses, though the domain-citation numbers come straight from the tracker. And publishing this will likely end the experiment, since there is now a page explaining the metrics are fictional. We decided the finding was worth more than the ongoing test.