An AI visibility audit is a structured review of how your business appears in AI answers, separating page fit, source links and brand mentions. If another business appears while yours does not, that alone cannot explain why. The audit helps you identify which pages and information need checking.
Summary: Build a question set around customer needs. Record page fit, links to your domain and mentions of your brand separately. Repeat questions under consistent conditions, retain the evidence and use recurring gaps to prioritise checks. Measure visits, enquiries and sales separately.
What the research says, and what we recommend
Research suggests that citations and brand mentions can occur together, but they are distinct observations. A brand can appear without its website being cited. Your audit should therefore record source links and mentions separately, without treating their association as proof that adding a citation causes a mention.
Benjamin Tannenbaum’s single-author arXiv preprint, dated 19 September 2026, examines 34,960 GPT/Gemini prompt-engine-run observations from 75 anonymised projects during June-September 2026. Records naming the target brand in the user’s prompt were excluded. These are observations, not customers or purchases.
Within records without branded fan-out, meaning additional search queries derived from the user’s question:
- GPT: 438 of 15,524 observations without an own-domain source included a brand mention (2.8%), compared with 866 of 1,769 with an own-domain source (49.0%).
- Gemini: the corresponding figures were 524 of 13,801 (3.8%) and 1,995 of 3,415 (58.4%).
These are different observation groups, not before-and-after results. A mention is not necessarily a positive recommendation.
The author is founder and CEO of Aiso Boost Ltd., which develops commercial AI-search measurement software. The paper declares a financial conflict of interest. Raw data were not released, preventing independent refitting. Read our research review for context.
DijitalPi commentary: Use this observational, diagnostic study to separate your records. Page fit is your editorial assessment, not a hidden score assigned by an AI system.
Separate the technical prerequisite first
Check technical eligibility before interpreting content gaps. For a page to be eligible as a supporting link in Google’s AI Overviews or AI Mode, it must be indexed and eligible to appear in Search with a snippet. Eligibility does not guarantee that the page will appear.
Google Search Central states that there are no additional requirements or special optimisations. Do not generalise this rule to every AI service.
- Check that the page opens and its content is readable.
- Verify indexing when assessing Google eligibility.
- Record access problems separately from content gaps.
DijitalPi’s free PiScan checks whether AI assistants’ bots can crawl, read and quote your site. It is a technical access pre-check, not a visibility assessment.
Step 1: Split your question set into three groups
Separate questions about finding a service, comparing options and researching your business directly. This prevents answers prompted with your brand name from being mixed with unbranded discovery answers. As a practical starting point, DijitalPi recommends 15-30 questions, each tied to a customer need and recorded with fixed wording.
For an illustrative commercial water-filtration service company in Manchester, the groups could be:
- Discovery: “Who provides office water-filtration maintenance in Manchester?”
- Comparison: “Should an office choose an annual water-filtration maintenance contract or pay-per-call servicing?”
- Brand: “Which areas of Manchester does [Company name] cover?”
Draw wording from sales conversations, call notes and search-query records. If customers say “regular filter replacement”, include that phrasing.
Record the relevant service, location and expected landing page beside each question. Finding no suitable page is itself a finding. If you change a question, start a new series rather than overwriting its history.
Step 2: Record three stages separately for every question
Assess page fit, own-domain citations and brand mentions independently for each answer. Use yes, no or partial where appropriate, supported by evidence. These fields describe different things you can observe. Their presence together does not establish a sequence of internal decisions or explain why an answer was produced.
One question, three separate records
- 01Page fit
Does your selected page answer the customer’s question clearly and sufficiently?
- 02Source
Do the answer’s visible sources include a link to your own domain?
- 03Mention
Does your brand appear in the answer text, and in what context?
DijitalPi’s illustrative audit framework. These records describe observations, not the systems’ hidden decision sequence.
For page fit: Save the URL and relevant passage. A page explaining filtration technology may not answer a question about Manchester maintenance services.
For sources: Save and open the exact link. Distinguish your domain from another website discussing your business. If no link appears, record “no visible citation”; do not infer the information’s origin.
For mentions: Save the exact sentence. Classify its context as recommendation, comparison, neutral mention or incorrect information. A brand name in a source title is not a mention in the answer text.
Explain every “partial” assessment. Keep ambiguous brand references unconfirmed until checked.
Step 3: Ask the same question more than once
A single answer is insufficient for identifying a recurring pattern. DijitalPi recommends asking each question at least three times on different days, using new chats and the same language. Keep search conditions as consistent as possible, document changes and assess each AI service separately rather than combining its results.
Save:
- Exact question wording and question group.
- Service, date and model name if displayed.
- Language, location, session and personalisation settings.
- Whether search was enabled.
- Answer text, source links and assessment notes.
Record unavailable details as “not shown”. Do not add your brand name between attempts.
Our Turkey AI agency visibility research uses Turkish questions in new chats under consistent language and search conditions. It assesses mentions, explicit recommendations and source links separately, aiming for three measurement days per question. No results or ranking have been published because human review remains incomplete.
Repeated answers are not independent customers.
Step 4: Read the results in one table
Give each question-and-service pairing its own row. Keep page fit, citation counts and mention counts in separate columns, with a short interpretation alongside them. The illustrative entries below show how one business can require different checks across questions. They contain no customer results, measured performance or comparative success claims.
| Question group | Example question | AI service | Page fits? | Own-domain source (3 attempts) | Brand mentioned (3 attempts) | Interpretation |
|---|---|---|---|---|---|---|
| Discovery | Who provides industrial water analysis in Manchester? | ChatGPT | No | 0/3 | 0/3 | The selected page covers office equipment maintenance only. |
| Discovery | Who provides office water-filtration maintenance in Manchester? | ChatGPT | Yes | 0/3 | 0/3 | A suitable page exists but appears in neither sources nor answer text. |
| Comparison | Should an office choose an annual water-filtration maintenance contract or pay-per-call servicing? | Gemini | Yes | 2/3 | 0/3 | The page is cited; the brand is absent from the answer text. |
| Discovery | Who provides regular filter replacement for Manchester workplaces? | ChatGPT | Yes | 0/3 | 2/3 | The brand appears without an own-domain citation. |
| Brand | Which areas of Manchester does [Company name] cover? | Gemini | Partially | 1/3 | 3/3 | The brand appears, but service-area information is incorrect. |
| Discovery | Who provides office water-filtration maintenance in Manchester? | Gemini | Yes | 2/3 | 2/3 | The same question produces different observations from ChatGPT. |
“2/3” means the condition appeared in two of three recorded attempts. It is neither market share nor an overall visibility rate.
Keep each original response behind the summary. Explain partial assessments and exclude ambiguous mentions from confirmed counts. Keep branded questions separate from discovery questions: frequent mentions cannot compensate for incorrect service-area information.
What does each gap in the table point to?
Gaps indicate hypotheses to investigate, not established causes. Begin with the match between question and page, then examine technical access, visible sources and clarity of business information. Repeat the same question series after changes. A different answer does not prove that your edit alone caused the difference you observed.
| Gap | Likely area to investigate | First check |
|---|---|---|
| Page does not fit | Content may not answer the question | Select the right page and address the actual service scope. |
| Page fits, no citation | Access, indexing, answer clarity or other sources | Check technical status, the direct answer and visible third-party sources. |
| Citation, no mention | Brand or service association may be unclear | Check how the brand is identified on-page and elsewhere. |
| Mention without own-domain citation | Information origin is unresolved | Verify the stated service, location and business details. |
| Mention contains incorrect information | Outdated or conflicting details | Compare your page with visible external sources. |
| Services differ on the same question | Sources or answer conditions may differ | Repeat attempts while keeping service records separate. |
Write finishable tasks: “update the service area”, “identify the relevant existing page” or “verify the cited information”. Assign an owner and the question to recheck.
How often should you repeat the audit?
DijitalPi recommends a monthly audit using a stable question set. Consider an additional round after substantial page, pricing or service changes, or a relevant AI-service update. Preserve previous records, note change dates and start a new series whenever question wording changes so that comparisons remain interpretable over time.
Compare recurring gaps with isolated differences. Record completed actions and unresolved hypotheses.
Measure visits, enquiries and sales separately. Citation and mention counts cannot replace them; our review of AI recommendations and website visits explores this distinction.
DijitalPi’s SEO services include a GEO/AEO layer that builds representative question sets and tracks which answers mention your brand over time.
Sources
The research figures come from the arXiv preprint below; technical eligibility statements concern Google’s documentation. Question counts, repetition schedules and the audit framework are DijitalPi’s practical recommendations. The illustrative business and table are teaching examples, not customer findings, measured performance or evidence of a ranking outcome.
- Benjamin Tannenbaum, From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search, arXiv preprint, version 1, 19 September 2026.
- Google Search Central, AI features and your website, last updated 10 December 2025.
