AI Visibility Audit: How to Establish an Honest Baseline
An AI visibility audit should help you decide what to investigate next. A percentage alone is not enough: you need to know which questions were asked, which model answered, and whether it searched the web.
This guide explains how to build a baseline you can repeat. It includes a first-party example from Organiq’s own research, and a worksheet you can use without treating an AI answer as an authoritative fact.
Decide what you are measuring
Separate three observations. A brand mention means the answer names your business. A domain citation means the answer contains a source link to your website. A referral visit means somebody followed a link and arrived on your site. These observations answer different questions and should have separate columns.
An answer can name your business without linking to it. A link can point to an informational page without recommending your product. A referral visit does not establish that the visitor purchased anything. Keep those distinctions visible when you share a report with your team.
Also distinguish a model API from the consumer assistant. An API response is an observation of that API, model version, and configuration. It is not automatically an observation of a person using ChatGPT or Claude with different settings and context.
Build questions without seeding the answer
Start with the business profile: the customer, the problem, the product category, and the decision being made. Create questions for discovery, comparison, and purchase consideration. For an early-stage software business, examples might include:
- “Which tools help a small team research customer search demand?”
- “How can a startup track whether AI assistants mention its product?”
- “What should I check before buying a content automation tool?”
These are generated examples, not measured customer demand. Do not insert your brand into discovery questions. “Does Organiq help with AI visibility?” can test the model’s description of Organiq, but it cannot fairly measure whether the model brings up Organiq unprompted.
Keep branded questions in a separate research category if you need them. Check exact duplicates and the presence of your brand before execution. Organiq’s self-test initially exposed this bias, and the clean replacement run excluded brand-seeded questions.
Record an auditable row for every answer
Use a worksheet with these fields:
| Field | What to record |
|---|---|
| Question | The exact text sent to the model |
| Audience and intent | Who could ask it and what decision they are making |
| Provider and model | The executed service and model version |
| Date and configuration | Execution time and relevant settings |
| Web search | Observed true/false, or unknown |
| Answer | The actual response, preserved for inspection |
| Brand mention | Whether the answer names the business |
| Citations | The actual source links returned |
| Search queries | Only queries exposed by the provider |
Do not reconstruct hidden searches from the wording of an answer. Do not call a link a citation unless it is actually present in the returned evidence. DataForSEO documents response fields and availability in its LLM response API reference.
Organiq’s first-party baseline
On October 13, 2026, Organiq ran 100 unique, unbranded generated questions against its Cloudflare development website. The tests used DataForSEO API responses from gpt-4.1-nano, with web search disabled.
The clean sample contained zero unsolicited Organiq mentions. A Wilson 95% interval for this sample is approximately 0–3.7%. This interval describes sampling uncertainty; it does not make the question set representative of every potential customer.
Citation share and hidden search queries were not measured by this configuration. Google traffic and indexing were also unknown. The site was a new development subdomain, rather than an established production domain. The published baseline summary records these limitations alongside the result.
This is useful evidence about one controlled API sample. It is not proof that consumer assistants never mention Organiq, and it is not a prediction of future search traffic.
Calculate rates against the eligible sample
For brand mention rate, divide answers containing the brand by eligible unbranded answers. Report the numerator and denominator, not just a percentage. If a prompt was branded, duplicated, failed, or otherwise excluded, preserve the exclusion reason.
For citation rate, use a sample that can actually produce observed citations. A non-search configuration with no citations does not establish poor search visibility. If you compare models, report each model’s configuration and sample size separately.
When comparing competitors, use the same question set and execution method. A brand appearing in an answer is an observed textual mention; it does not validate every claim the answer makes about that company.
Turn the audit into a next action
Read the underlying questions before choosing content. Identify a customer problem your website explains poorly, then inspect relevant sources and competing pages. Decide whether to improve an existing page or create a distinct resource. Avoid generating a separate article for every slight phrasing variation.
Record what you change and why. Repeat a comparable sample later, preserving the earlier answers and configuration. A change in mention rate can coincide with your work without proving that the work caused it.
Organiq stores dated research and requires a price quote before paid execution. You can review the website preview, or connect the same workspace through the MCP guide. Start with a defensible baseline, then choose one useful improvement.
Want an evidence-based strategy for your website? Analyze your website.