Methodology

How we test whether AI recommends a brand

This is the public methodology behind RecommendByAI. We use a fixed prompt set, compare results across the major assistants, score visibility and framing, then inspect the evidence likely shaping the answers.

Answer first

What is the RecommendByAI methodology?

A repeatable way to test whether AI assistants mention, describe and recommend a brand: fix the prompts, run them across the main systems, score the results, then inspect the evidence behind the answers.

The point is not to guess how models work in theory. The point is to observe how they behave in category-level recommendation prompts, compare that behavior across assistants, and turn it into actions a company can take.

Why publish it

We want the method to be inspectable, not mystical

Product rule

If recommendbyai.com is not itself becoming more recommendable, then the methodology is not finished.

This page exists because AI Visibility is still an emerging category. We would rather show how we think, what we measure and where the method has limits than hide behind vague claims about proprietary scoring.

Coverage

Which systems we review

The assistant set matters because buyers do not ask only one model, and different systems can frame the same category differently.

ChatGPT

We test recommendation-shaped prompts, category questions and competitor comparisons in the assistant most buyers already use.

Claude

We compare whether Anthropic's assistant frames the same category differently, names different competitors or cites different evidence.

Gemini

We include Google's assistant because category understanding and recommendation behavior can diverge from the rest of the field.

Perplexity

Perplexity is especially useful because citations are visible, which makes source analysis easier and more auditable.

Google AI answers

Where relevant, we also review AI-generated answers in search because buyers increasingly meet category guidance there first.

Prompt design

How we build the prompt set

A good methodology starts with realistic questions, not generic keyword variations.

Category-definition prompts

Questions like 'What is AI Visibility?' or 'What is LLM visibility?' that show how a model defines the space and which sources it seems to rely on.

Pain and diagnosis prompts

Questions like 'Why does ChatGPT recommend our competitor instead of us?' that reveal how the assistant explains absence, gaps and category ownership.

Method and solution prompts

Questions like 'How do I get recommended by AI?' that test whether the model recognizes a practical process and which sources it trusts for that advice.

Commercial prompts

Questions like 'AI Visibility audit' or 'who offers AI Visibility audits?' that matter most if the goal is eventually to be recommended as a provider.

We keep the set stable enough to compare over time, but varied enough to reflect how a buyer actually thinks: broad category questions, pain prompts, solution prompts and commercial-intent prompts all reveal different recommendation behavior.

Measurements

What we actually score

The method combines a small metric set with a human read of the answers.

Mention rate

How often the brand appears at all across the fixed prompt set.

Share of voice

How often the brand is mentioned relative to named competitors in the same answer set.

Accuracy

Whether the assistant describes the product, category, audience and limitations correctly.

Intent coverage

Whether the brand appears for the specific recommendation jobs it wants to own, not just for navigational brand queries.

Framing

What role the assistant assigns: default pick, niche option, enterprise choice, budget choice, or afterthought.

Process

How a methodology run works, step by step

  1. 01

    Fix the category and competitor set

    We start with the company's target use case, a small list of real competitors and the recommendation intent it most wants to win.

  2. 02

    Build a stable prompt set

    We design prompts across the four main buckets so repeated runs stay comparable instead of turning into anecdotal tests.

  3. 03

    Run prompts across multiple systems

    We test the same logic across ChatGPT, Claude, Gemini, Perplexity and, where relevant, Google AI answers.

  4. 04

    Record answers and named brands

    We note who is mentioned, in what order, with what reasoning, and whether a source is visible or inferable.

  5. 05

    Score the result and read the framing

    Numbers alone are not enough, so we combine metrics with a qualitative read of how the model is positioning the brand.

  6. 06

    Trace likely evidence sources

    We inspect documentation, comparisons, directories, reviews and community discussions that appear to be shaping the answer.

  7. 07

    Turn findings into actions

    The output is a short, prioritised set of actions aimed at improving legibility, evidence quality and recommendation fit.

Limits

What this methodology can and cannot prove

  • -Models change over time, so this method measures a moving system rather than a fixed ranking environment.
  • -Two runs of the same prompt can vary, which is why repeated tests and stable prompt wording matter.
  • -No outside observer can see full model internals; the method measures output behavior and likely evidence, not hidden weights.
  • -A higher mention rate does not automatically mean a better outcome if the brand is framed wrongly.

In other words, this method is built to improve decision quality, not to promise control over a closed system. That is also why we pair it with a public learning loop and a manual audit rather than a synthetic score.

Next step

What this page should lead to

To see how this methodology is applied to RecommendByAI itself, open the public experiment log. If you want the operational measurement view, continue to how to measure AI Visibility. If you want this methodology applied to your category, request the AI Visibility Audit.

Apply the methodology to your market

We use this same process to test who gets recommended in your category, why competitors are named instead of you, and what evidence is shaping those answers.