Atlas research method

Methodology

Sampling, citation extraction, ranking calculations, and version controls for Viclaro Atlas.

Public Versioned prompt sets Immutable snapshots 95% confidence intervals
On this page

01

Scope

Atlas measures how often selected AI assistants name businesses in response to a fixed prompt set for a city and category. It does not measure service quality, market share, search traffic, or buyer conversion.

Observation
One assistant response
Scan run
Prompt × assistant × repetition
Snapshot
One published ranking dataset

02

Sampling

Prompts describe buyer situations, constraints, research questions, and selection decisions. Each prompt is sent independently to every assistant in scope and repeated according to the prompt-set protocol.

Crisis

Immediate event

Advice

Whether to act

Situation

Specific circumstances

Constraints

Budget or access

Learning

Process question

Decision

Who to consider

Validation

Check a named firm

Follow-up

Later conversation turn

Bucket names and counts are versioned. Selected prompt wording is published in dated vertical research reports so readers can inspect the context behind a recommendation; the complete operational panel remains restricted.

03

Ranking inclusion

Ranking layer

Prompts that request a business recommendation or evaluation. Citations from these responses determine rank and response mention rate.

Editorial layer

Advice, crisis, learning, or constraint prompts that do not request names. Samples may be published, but they are excluded from ranking calculations.

04

Entity extraction

  1. 1Parse business and professional names from each response.
  2. 2Match names to Atlas records using normalized names, aliases, domains, and canonical entity links.
  3. 3Create an unverified citation-only record when no existing entity matches.
  4. 4Store the citation with its response, assistant, prompt key, and snapshot source.

Directories and non-business reference sources are excluded from firm rankings.

05

Metrics

Mentions

Eligible responses that named the business.

Response mention rate

mentions ÷ eligible responses. Reported with a 95% Wilson confidence interval.

Mention share

business mentions ÷ all business mentions in the snapshot.

Assistant coverage

Number of tested assistants that named the business at least once.

Rank orders businesses within one snapshot. Comparisons across categories require the underlying rates and sample sizes.

06

Models and snapshots

Claude

anthropic/claude-sonnet-4.6

ChatGPT

openai/gpt-4o

Gemini

google/gemini-2.5-flash

Perplexity

perplexity/sonar

Snapshots retain their prompt-set version, model set, source runs, sample size, and publication date. Published snapshots are not rewritten. A model or prompt-set change starts a new comparison series.

Model change · August 3, 2026

Gemini changed from gemini-2.5-pro to gemini-2.5-flash; Perplexity changed from sonar-pro to sonar. Results on opposite sides of this change are not directly comparable.

07

Prompt sets

Active sets are shown with their published bucket counts. Expand a set for details.

No active prompt sets.

08

Access and corrections

Public

  • Rankings and confidence intervals
  • Per-assistant counts
  • Prompt-set versions
  • Professional citation tables

Research access

  • Response-level records
  • Per-prompt breakdowns
  • Historical exports
  • Custom scan scopes

For data access, factual corrections, or removal requests, email support@viclaro.app. Factual corrections include names, websites, entity matches, and category assignments. Ranking positions are not changed on request.

Atlas and per-brand audits

Atlas reports category-level results from its versioned prompt sets. A per-brand audit runs the same assistant panel — ChatGPT, Claude, Gemini, Perplexity — against a narrower deck of buyer prompts scoped to a single subject brand, and reports two measures side by side: the observed prompt-level citation rate (how often the panel names the brand across the deck) and a page-side citability prediction (a model-scored estimate of the ceiling on citation the audited page content alone could support). A divergence between the two measures is consistent with constraints beyond page quality — off-page authority, retrieval behavior, category positioning, brand familiarity, and sample variance are each candidate contributors. Draft answer blocks generated for uncited prompts are editorial starting points and require fact and compliance validation before publishing. Results across audits are comparable only when the prompt panel and model definitions match; the deck is held constant across successive audits of the same brand so that change between audits is attributable to shipped content or off-page shifts rather than measurement drift.