Methodology

Methodology

A repeatable audit, interpreted by a human.

The evidence collection is standardised. The questions, competitors, business facts and final priorities remain specific to the client.

The baseline

The standard 24-call baseline

The standard audit uses six buyer-style questions, two web-grounded AI providers—OpenAI and Anthropic—and two runs of every question on each provider.

6 questions × 2 providers × 2 runs = 24 measured responses

Perplexity can be added when its audience or citation behaviour is commercially relevant. Expanded work may use a 72-call profile, additional markets or deeper citation analysis.

Process

The delivery sequence

1

Scope and intake

Agree the business, website, market, products or services, competitors, decision-maker and whether implementation is included.

2

Canonical facts

Confirm the names, contact details, service area, leadership, services, prices, profiles and any real accreditations.

3

Query design

Write questions as prospective customers would ask them, without naming the client or forcing a desired answer.

4

Preflight

Run a small smoke test to confirm authentication, live web search, source capture, name matching and query scope.

5

Baseline

Run the approved profile unchanged. Every result is saved immediately with its provider, model, timestamp, response and sources.

6

Website review

Inspect initial HTML, indexability, metadata, schema, services, canonical facts, trust evidence, listings and stale information.

7

Analysis

Calculate recommendation frequency, competitor recurrence, provider differences, cited domains and factual errors.

8

Report and QA

Verify headline counts from the raw file, inspect the final PDF, remove unsupported claims and distinguish findings from hypotheses.

9

Results call

Confirm facts, explain priorities, assign ownership and scope implementation separately where needed.

10

Remeasurement

After changes have had time to propagate, rerun the same questions and providers and preserve the new dated snapshot.

Evidence

What is recorded

The exact buyer question
Provider and model
Timestamp and run number
Whether web search ran
The full response and cited sources
Client and competitor mentions
Errors, retries and query exceptions
Raw results are supplied as a technical appendix. They are not replaced by a score with no evidence trail.
Honesty

Limitations stated plainly

!AI responses vary between runs and change when models or indexes change.
!An API baseline is a controlled proxy for discovery, not a claim about every consumer session.
!A recommendation rate is not a revenue forecast.
!Correlation after implementation does not automatically prove causation.
!Technical improvements make a site easier to retrieve; third-party authority is usually also required.
!No provider offers a guaranteed position or permanent citation.

See the method applied to Deep Domain.

View the sample self-audit
Scroll to Top