METHODOLOGY

How the AI Recommendation Index is built

The AI Recommendation Index measures which businesses ChatGPT, Claude, Perplexity, and Gemini recommend when buyers ask for help choosing a service provider. Every ranking comes from real queries submitted to these platforms every month. No estimates, no models, no paid placement. This page explains exactly how.

Version 1.0 | Published March 2026 | Last updated April 2026

What the index measures

The AI Recommendation Index tracks how often a business appears in AI responses to buyer-intent queries. We ask ChatGPT, Claude, Perplexity, and Gemini who they recommend for a given industry in a given city, then we count which businesses get named, how often, and on which platforms.

Why two scores

Frequency alone does not distinguish a business recommended confidently across all four platforms from one that appears heavily on a single platform. A firm ranked first on ChatGPT alone may have a higher ARS than a firm mentioned across all four platforms at lower frequency, but the latter is a more durable AI recommendation signal. CIS catches that nuance.

How we collect the data

Every month, we run 60 distinct buyer-intent prompts across four AI platforms for every tracked city, giving 240 total responses per city per month. Queries use web-search-enabled API access to match consumer behaviour. No scraping of chat interfaces, no browser automation, no fake accounts.

Four AI platforms

Each platform uses a different training approach, different retrieval model, and different citation handling. Tracking all four gives a more accurate picture of which businesses are recommended consistently across AI systems, not just on one.

60 queries per city per month

For each industry and city combination, we submit 60 distinct buyer-intent prompts to each of the four platforms. Prompts use varied framings: "who should I hire", "best", "most recommended", and specific contextual variants. Each prompt runs on all four platforms, giving 240 total AI responses per city per month.

Prompts are locked at the start of each city's tracking program and never modified. This ensures month-on-month comparisons are genuine rather than artefacts of prompt changes. Prompts are published on request to journalists and researchers.

Controlling for response variance

AI platforms produce variable responses. The same prompt run twice can return different businesses in different orders. We control for variance three ways:

• Prompt set locked. Every city runs the same 60 prompts each month. Changes are logged in the update history section below.
• Platform parameters locked. Model version, temperature, and tool configuration are recorded per run. If OpenAI or Anthropic deprecates a model, we note the switch date.
• Monthly cadence. Running monthly rather than weekly smooths short-term noise. Real trends emerge over 3 to 6 month windows, matching how AI recommendation infrastructure actually evolves.

How the CIS is calculated

The Composite Index Score is built from four components, each weighted by how strongly it signals durable AI recommendation authority. A business can have a high ARS but a modest CIS if it only appears on one platform or always appears late in lists. CIS catches that nuance.

Worked example

How we parse AI responses

Every AI response is raw text. To turn text into rankings, we parse for business names, resolve them to real entities, and de-duplicate across platforms. The parser is conservative: when a name is ambiguous, we log it and exclude it rather than risk wrong attribution.

How parsing works

After each query runs, the raw response is stored verbatim. A parsing pass extracts candidate business names using pattern matching tuned to each industry (for example, accountant firm names often include "Partners", "Chartered", or "CPA"). Candidate names are then matched against a known-entity table built from Google Places, public business registries, and prior run history. Matched entities carry the rank position and citation context from the response into the intelligence_results table.

What we exclude

Accuracy tradeoffs

The conservative approach means true positives are high confidence, but we may under-count firms with common names. Over time we improve entity resolution as we see which responses correctly identify specific firms versus which are ambiguous. Parser changes are logged in the update history.

What the index covers

The Recommendation Index currently covers 10 professional service industries across 7 cities in 1 country. Coverage expands month by month as new cities complete their baseline runs.

Industries tracked

• Tax Accountants
• Mortgage Brokers
• Financial Advisers
• Family Lawyers
• Corporate Lawyers
• Personal Injury Lawyers
• Immigration Lawyers
• Psychologists
• Air Conditioning Installers
• IVF Clinics

We prioritise industries where AI recommendation has high buyer-decision weight: professional services where buyers typically research 3 to 5 options before choosing, and where AI systems are already being used as a research starting point.

Countries tracked

• Australia

How new markets are added

New cities and industries are added based on (1) buyer query volume, (2) data availability for parsing, and (3) specific requests from businesses or journalists. A new city takes approximately one month to go live: baseline run, parser tuning, and entity resolution. A new industry takes longer because we need to develop query patterns and entity recognition for that vertical.

See the live index at www.ai-seeyou.com/australia/ or pick a specific industry from the list above.

The research that shapes this index

The methodology draws on peer-reviewed GEO research, published AI platform documentation, and empirical citation analyses. Most GEO vendor claims have no evidence base. We only use techniques backed by published research or direct platform documentation.

Why monthly cadence

The Princeton paper confirmed AI recommendations shift as training data, retrieval, and underlying search indexes update. OpenAI updates its index frequently. Anthropic re-indexes Claude periodically. Monthly cadence is the minimum required to detect genuine change versus random response variance.

Using and citing the data

The AI Recommendation Index data is free to cite with attribution for editorial, journalistic, and academic use. Commercial use (reselling the data, embedding it in a competing product, training an AI model) requires a separate agreement. Contact dave@ai-seeyou.com for commercial terms.

Attribution format

Source: AI See You Recommendation Index, {month year}.
https://www.ai-seeyou.com/{industry-slug}/{country-slug}/{city-slug}/

For researchers

Academic researchers can request the full prompt set, parser configuration, and anonymised response corpus for peer review or replication studies. Subject to a data use agreement that prohibits re-identification of specific firms for commercial purposes. Email with institutional affiliation and intended use.

Version history

Methodology changes are logged here so journalists, researchers, and clients can track what changed and when. Material changes (for example, new AI platforms, changed prompt sets, scoring formula updates) trigger a major version bump.

Do AI engines recommend your business?

Get a free Recommendation Score in 30 seconds. See exactly where AI recommends you, where it doesn't, and what to fix.