LoganAdams-ClearCited/Clear_Cited_AI_Visibility_Index

Clear Cited AI Visibility Index

Open, pre-registered measurement of which products generative AI engines name and cite when buyers ask for software recommendations.

Measured	2026-07-01
Engines	ChatGPT (OpenAI) - Perplexity - Claude (Anthropic) - Gemini (Google) - Grok (xAI)
Protocol	10 prompts x 10 runs per engine per category
Scope	9 categories - 90 ranked products
Licence	CC BY 4.0
Cite	https://doi.org/10.5281/zenodo.21612952

What this measures

Share of model is the proportion of qualifying AI answers that name a given product. It is a proportion over answers, not over mentions - an answer naming a product three times counts once.

A single AI answer is one draw from a distribution. The same prompt issued twice to the same engine can return different products in a different order. This dataset exists because a screenshot of one answer carries no information about the distribution it came from.

Files

Each category has a .csv (ranked summary) and a .json (full record).

Category file	Category
ai-observability-tools	ai observability tools
api-platforms	api platforms
cicd-platforms	cicd platforms
crm-software	crm software
databases	databases
feature-flags	feature flags
incident-management-platforms	incident management platforms
product-analytics-platforms	product analytics platforms
vector-databases	vector databases

CSV columns

Column	Meaning
rank	Position within the category, ordered by share of model. 1 = highest.
product	Product name as it appears in the frozen brand universe.
domain	The product's primary web domain.
share_of_model_pct	Percentage of qualifying AI answers in this category that name this product. The headline metric.
share_ci_low_pct	Lower bound of the 95% Wilson confidence interval on share of model.
share_ci_high_pct	Upper bound of the 95% Wilson confidence interval. Overlapping intervals between two products mean they are indistinguishable at this sample size.
appearance_rate_pct	Percentage of all answers - qualifying or not - in which the product appeared. Diverges from share of model when engines frequently decline to recommend.
citation_rate_pct	Percentage of qualifying answers that cite the product's own domain as a source. Naming and citing are different behaviours.
mentions	Raw count of answers naming this product. An answer naming it three times counts once.
answers_total	Total qualifying AI answers for the category. The denominator for share of model.

Overlapping intervals mean two products are indistinguishable at this sample size, whatever their point estimates. Rank order is reported; rank differences inside an overlap are not claimed as findings.

JSON

Strictly richer than the CSV. Adds per-engine breakdowns (by_engine), the full citation_sources block, captured_at_utc, per-engine model identifiers, and the pre-registration fields prompts_hash, brand_universe_hash and frozen_before_run.

Method

Ten buyer-intent prompts per category, each issued ten times to each of five engines. Prompts and the brand universe were hashed and frozen before execution; the hashes are in every JSON record. Runs are independent - no conversational context carries between them.

Products cannot pay to appear or to rank.

Google AI Overviews and Microsoft Copilot are measured separately as answer surfaces and are never summed into the five-engine share of model. A surface embedded in a search product draws on that product's retrieval index; summing correlated observations would inflate apparent coverage without adding independent information.

Full protocol: https://clearcited.com/methodology/ - and as a citable paper: https://doi.org/10.5281/zenodo.21614889

Limitations

• Point-in-time. One run captures one moment. Model versions change beneath the measurement, which is why every record is dated at capture.
• Prompt coverage. Ten prompts per category is a deliberate coverage/cost trade. Share of model is sensitive to phrasing; the prompt set is published so that sensitivity can be examined rather than assumed away.
• Run count. The ten-run minimum is a working figure, not an empirically derived one. No published study establishes where an AI-visibility ranking stabilises.
• Citation extraction. The extraction layer under-recovers on two of the five engines, and a large share of extracted citations resolve to platform redirect wrappers rather than sources. The citation columns are published as captured so the defect is checkable. Do not draw domain-level conclusions from them yet.
• Scope. English-language, US-market, B2B software categories only.
• Intervals describe sampling variation within one run of this protocol. They do not account for model drift between runs.

Conflict of interest

Clear Cited operates a commercial AI-search-optimization practice and therefore has a financial interest in the domain this dataset measures. The data, prompts, hashes and protocol are published in full so that any figure can be independently verified or contested.

Citation

Adams, L. (2026). Clear Cited AI Visibility Index [Data set]. Clear Cited. https://doi.org/10.5281/zenodo.21612952

That DOI is the CONCEPT DOI: it always resolves to the current version. Cite it rather than a version DOI so the citation stays alive across monthly refreshes.

Author: Logan Adams - ORCID 0009-0009-1235-9479

Source and attribution

The canonical, always-current data lives on the Index pages - please attribute Clear Cited and link back there: https://clearcited.com/ai-visibility-index/

• Knowledge-graph entity (Wikidata): https://www.wikidata.org/wiki/Q140443471
• Licence: CC BY 4.0. Free to reuse WITH attribution to Clear Cited and a link back to the canonical Index page.