# AI visibility: what is proven and what is speculation

Topic: [AI discoverability (GEO)](https://btlabs.dev/en/ai-discoverability)

> Four drawers for GEO advice: officially documented, scientifically studied, plausible but unproven, plainly false. With three questions to check any claim.

This query recently showed up in our Google search data: "many geo tips online contradict each other. what is scientifically or officially proven about ai visibility and what is speculation?"

Whether a person or a tool typed it, we cannot see. The question is still better than most guides on the subject. So here is an attempt at an honest answer — sorted by what actually stands behind each claim.

## What is officially documented

Here there is vendor documentation. That is not an opinion and not a study, but a statement from the people who run the systems.

**Structured data following schema.org is processed by search engines.** Google documents in [Search Central](https://developers.google.com/search/docs/appearance/structured-data/search-gallery) which types are read for which presentation format. It is readable and maintained by Google itself. What it does **not** say: that more markup automatically means better placement.

**Crawlers can be controlled via robots.txt, including those of AI providers.** OpenAI, Anthropic, Google and Perplexity publish their identifiers. That an exclusion is respected is a promise by the providers — technically it cannot be enforced.

**Excluding from training is not the same as excluding from reading.** Several providers now separate the two into distinct identifiers. You can therefore permit content to be read for an answer while forbidding its use in training.

## What has been studied scientifically

Here there are publications with a method. That does not make them the final word — it means someone measured and disclosed how.

**That content can be deliberately prepared for generative answers has been studied.** The [Princeton work on Generative Engine Optimization](https://arxiv.org/abs/2311.09735) tested different preparations against each other and measured differences in citation frequency. The strongest effect came from verifiable elements: attributed quotations, figures with a source, concrete details instead of general phrasing.

That is a solid direction. What it is not: a guarantee, nor a statement about your business in your niche.

**That a growing share of searches ends without a click has been measured repeatedly** — by different parties, with different numbers and different methods. The order of magnitude holds; the exact percentage does not. Anyone giving you a precise figure without naming the survey is passing off an estimate as a measurement.

## What is plausible but unproven

Most of what circulates as advice right now belongs here. That does not make it wrong — only unproven.

**The llms.txt file** is a proposal: provide a content index for language models. The idea is sensible. Whether large providers read and weight it is confirmed by none of them. We have one because it costs nothing — not because we know it works.

**Optimal text lengths, paragraph lengths, list formats.** Very precise recommendations circulate. We have found no solid source covering more than one model.

**That mentions carry more weight than links.** This is a widespread thesis and it is plausible — language models do not reason like a link graph. We have not measured it soundly, and we know of no survey that does so cleanly.

## What is simply false

**"We register your website with ChatGPT."** There is no registration. No provider runs a directory you sign up to.

**"We guarantee first place in AI answers."** There are no places. Two models answer the same question differently, and the same model answers differently on two days.

**"AI visibility shows up in visitor numbers."** If a model reads your page and builds an answer from it, the person asking never sees your address. The benefit arises, but it appears in no statistic.

## How to work out which drawer a piece of advice belongs in

Three questions that carry surprisingly far:

1. **Who is saying this, and what do they gain?** A provider selling a service is not a neutral source. That does not disqualify them — it just means a second source is needed.
2. **Is there a measurement behind it, and is it described?** "We found that" is not a method. How many cases, over what period, compared to what?
3. **What would have to happen for the claim to be false?** If no answer comes, it is not a statement about the world but an opinion.

## Our own rule

In our articles we state where a figure comes from, and when we cannot prove something we say so. That occasionally produces unsatisfying sentences like "we have not measured this". We prefer them to a number that sounds good and holds up to nothing.

If you want to know how to actually measure your own position rather than estimate it, that is [here](https://btlabs.dev/en/posts/what-you-measure-when-position-one-is-worth-nothing). And why a machine reads literally rather than by meaning is [here](https://btlabs.dev/en/posts/a-machine-reads-literally).

---
Source: https://btlabs.dev/en/posts/ai-visibility-proven-or-speculation
Last-Modified: 2026-10-16T07:00:01.277Z
Languages: [de](https://btlabs.dev/llms/de/posts/ki-sichtbarkeit-belegt-oder-spekulation) · [it](https://btlabs.dev/llms/it/posts/visibilita-ia-dimostrato-o-speculazione)
See also: [llms.txt](https://btlabs.dev/llms.txt) · [ai.txt (Policy)](https://btlabs.dev/ai.txt) · [identity.json](https://btlabs.dev/identity.json)