An AI visibility audit should answer three questions: where does the brand appear, why do those outcomes happen, and what should the team change next? If the audit ends with screenshots and a score, it is incomplete.
The checklist below is designed for a focused first audit. Use a defined time window, save the underlying answers, and write down assumptions. AI outputs vary, so the result is a baseline for decisions, not a permanent verdict on the brand.
Invite the people who will act on the findings before choosing the scope.
1. Define the commercial scope
Choose the product, audience, geography, language, and decision you are auditing. “Our AI visibility” is too broad. “Our visibility among US SaaS teams comparing AI recommendation monitoring tools” is testable. Record excluded markets and products so nobody generalizes the findings beyond the sample.
2. Build a buyer-intent prompt set
Use observed questions from sales, support, reviews, communities, interviews, and search data. Cover problem awareness, category discovery, use-case fit, comparison, validation, and final choice. Our guide to a buyer-intent prompt library provides the full method.
Mark which prompts should reasonably generate vendor recommendations. That becomes the denominator for recommendation rate.
3. Set a reproducible test protocol
Document the answer surfaces, dates, locations, account state where relevant, prompt wording, rerun policy, and how you handle errors. Save complete responses rather than manually copying brand names. No protocol removes model variation, but a consistent one makes comparisons less misleading.
4. Measure category mentions
Calculate the share of tracked answers that mention the brand, then inspect context. Separate positive, neutral, negative, and incidental appearances. Segment by intent and model. A brand can have a healthy overall mention rate while disappearing from every comparison prompt.
5. Measure recommendations and position
For recommendation-eligible prompts, record whether the brand was presented as a suitable option. Track top-three presence or stated order only when the answer actually provides an order. Note the reason given for inclusion or exclusion. Avoid assigning precise ranks to unordered prose.
6. Inspect citations and cited pages
Record visible citations, destination URLs, domains, ownership, and the claim each source supports. Verify that links resolve and that the cited page contains the relevant information. A citation is evidence exposure, not automatic endorsement.
OpenAI advises publishers to allow OAI-SearchBot when they want site content eligible for ChatGPT search summaries and citations. Microsoft now provides page-level AI citation activity while cautioning that counts do not indicate ranking or authority. Use platform reporting where available, but keep it separate from cross-surface observation.
7. Map recurring competitors
Count which competitors appear for which intent groups and how they are described. Look for displacement patterns: the competitor that repeatedly wins enterprise prompts may differ from the one winning entry-level prompts. Do not reduce every competitor appearance to one league table.
8. Compare the evidence footprint
For priority losses, compare owned pages, documentation, comparisons, proof, reviews, directories, partner pages, and relevant editorial coverage. Ask what concrete claim supports the competitor’s recommendation and whether equivalent evidence exists for your brand.
Use the framework in why AI systems use some sources to assess relevance, access, specificity, consistency, and freshness.
9. Audit owned content clarity
Can a careful reader quickly determine what you do, who it is for, supported use cases, key differences, pricing approach, integrations, proof, and limitations? Check the homepage, product pages, comparisons, docs, help content, about page, and key proof pages.
Do not rewrite for a mythical robot dialect. Google’s current guidance says useful, original, people-first content and sound SEO foundations remain central to its generative Search experiences. Clarity helps because readers need it too.
10. Check technical eligibility
Confirm important pages return successful responses, expose meaningful main content, use intentional robots directives, have accurate canonicals, appear in sitemaps where appropriate, and are internally linked. Review JavaScript rendering, duplicate URLs, redirects, and stale pages. Check the crawler controls for each platform you care about instead of assuming Googlebot rules cover every system.
The Swep documentation can help connect monitored prompts and sources to your ongoing workflow, but Search Console, Bing Webmaster Tools, server logs, and URL inspection remain important first-party checks.
Test key pages as both a crawler and a reader. Confirm that consent layers, bot protection, authentication, or client-side failures do not hide the main answer. Look at the rendered page rather than assuming a successful status code means the content is usable. Where platform controls differ, keep a small crawler-access register with the responsible owner and intended policy. That prevents an accidental security or infrastructure change from quietly removing important pages from discovery.
11. Prioritize gaps by impact and confidence
Score each finding on commercial value, frequency, evidence strength, effort, and ownership. A repeated loss on an important comparison prompt with a clear missing proof page should outrank a one-off mention change on an informational query.
Label observations, hypotheses, and confirmed defects separately. “The competitor was cited in four answers” is an observation. “Their comparison page caused it” is a hypothesis until stronger evidence exists.
12. Create a rerun and ownership plan
Every priority action needs an owner, due date, affected prompt cluster, proposed asset, and measurement date. Preserve the original baseline and rerun the stable prompt set after changes have been accessible long enough to be discovered. Do not promise an immediate response from systems you do not control.
A simple audit output
- Scope and methodology
- Prompt inventory and eligibility rules
- Mention, citation, and recommendation findings by segment
- Answer and citation evidence
- Competitor and source-gap analysis
- Technical findings
- Prioritized action register
- Rerun date and success criteria
This is enough for leadership to understand the stakes and for content, SEO, product marketing, and engineering to know what to do.
What not to conclude
Do not claim market share from a small prompt set. Do not treat one model as “AI.” Do not infer endorsement from a citation. Do not claim a page caused a change because it was published before the change. Do not hide prompts that produced uncomfortable results.
Also do not turn the audit into a mandate to publish hundreds of pages. Google warns against scaled, low-value content and query-variation pages. One complete, evidence-rich page can resolve several related gaps.
From audit to operating rhythm
A first audit creates the map. The durable value comes from repeating a smaller core workflow: monitor priority prompts, inspect meaningful changes, diagnose the evidence, ship one credible improvement, and measure again.
Swep brings that loop into one place by connecting prompt-level answers, competitors, citations, and action plans. The audit is not successful because the score looks good. It is successful when the team can defend what it measured and knows what to improve next.