A defensible AI visibility audit uses a documented query library of 50 to 200 queries, runs it on at least four AI engines, has trained reviewers code every output, extracts every cited source, and records the method so the next audit is comparable. Audits commissioned across the industry vary widely in rigor, and the differences in method produce very different findings on the same brand. This guide shows what to ask for and what to reject.
What are the components of a defensible AI audit?
A defensible AI audit has five components: a query library, cross-platform execution, human review, source extraction and documented methodology. Each one answers a specific weakness in weaker audits.
| Component | The standard | Why it matters |
|---|
| Defined query library | 50 to 200 documented queries: branded, category, competitive and risk queries | Re-running the same queries each quarter is what produces trend data |
| Cross-platform execution | At least ChatGPT, Claude, Perplexity and Google AI Overviews, ideally Gemini | Differences across platforms show where the brand is strong and where it has gaps |
| Human review of outputs | Trained reviewers code outputs on a defined rubric | Automated sentiment scoring misses context and subtle errors in AI outputs |
| Source extraction | Capture every cited publication, article and recurring author | It is the most actionable data the audit produces |
| Methodology documentation | Query library, timestamps, platforms, review rubric and score calculation all recorded | The next audit can then run on the same basis |
The four query categories in a useful library are branded queries (the brand name alone, or with a product or category), category queries ("who are the leading companies in Y"), competitive queries (brand X versus brand Y) and risk queries (questions that touch past controversies, reputational issues or sensitive topics).
What should an AI audit deliver?
A defensible audit delivers five things the communications leader can defend to a CFO, which means transparent method, a defensible sample size and recommendations that translate into specific work.
- Presence rate by platform: the share of relevant queries that surface the brand at all.
- A sentiment and accuracy summary: how the brand is framed when mentioned, and where the factual errors are.
- A cited-source map: which publications and articles appear, ranked by frequency.
- A competitive comparison: how the brand does against named competitors on the same query set.
- A recommendations document that ranks specific, addressable gaps.
What are the red flags of a weak AI audit?
Five patterns should trigger skepticism, and each one removes information a buyer needs.
| Red flag | What is wrong with it |
|---|
| A single-score output | A vendor that gives a brand a "65 out of 100" AI visibility score without the underlying queries, method and rubric is selling a marketing product, not a measurement |
| Synthetic benchmarks | Query sets that are not specific to the brand's category or business give precise but rarely meaningful output |
| No platform comparison | A one-platform audit shows one platform's view, and brand performance varies across platforms |
| No source extraction | An audit that records what the AI said but not what sources it drew on omits the most actionable data |
| No follow-up cadence | A single snapshot has limited value, and strategic insight lives in trends across several quarters |
How much work is a baseline audit?
For one brand at moderate complexity, a baseline audit with 75 to 150 queries across four to five platforms, with full human review, source extraction and a recommendations document, typically takes two to four weeks of work. Ask every vendor for a written scope that lists the query count, platforms, review process and deliverables, and compare scopes before comparing prices.
Several agencies offer dedicated audit practices, including 5W, which shares common ownership with Everything-PR. Prospective clients should ask each one about the query library, platforms covered, human review process and reproducibility before commissioning.
What can an AI audit not do?
An audit documents the current state, and it does not change what AI engines say about the brand. The work that shifts the state, which is earned media, owned content and entity hygiene, is separate and longer-term.
An audit does not predict performance, because AI products change quickly and answers shift week to week. It captures a moment, supplemented by trend data over several moments.
An audit also does not remove hallucination risk. Even brands with strong AI visibility programs find themselves misdescribed by a model from time to time. Monitoring catches it, and audits document it.
What an audit does is establish a defensible baseline, find the highest-leverage gaps and provide the measurement system that later work builds on. For brands serious about AI visibility, the audit is the place to start. See the 5-step AI visibility audit for the method.
