Insight
Run fixed prompts, save every answer and citation, find claim gaps, and assign each content fix to a named owner for follow-up.
An AI visibility audit is a structured check of how AI search platforms mention a brand, describe it, and cite sources (Ahrefs). Start with fixed buyer prompts. For each response, save the exact prompt, complete answer, visible citations, and run conditions. Review each checkable claim against approved sources, record what the answer gets wrong or leaves unsupported, then assign one content fix to a named owner. Supporting pages still need to be indexed and eligible for Google Search with a snippet, and Google requires no special AI file or schema (Google Search Central). Keep mentions, citations, and claim accuracy as separate observations. Do not turn them into a universal score or claim that a content edit caused a later answer. Each capture records what appeared for that prompt under those conditions at that time.
An AI visibility audit examines brand mentions, factual accuracy, and cited sources across AI search platforms (Ahrefs). Its scope runs from a saved buyer question to a captured answer, then to a checked claim and an owned correction. The output is an evidence ledger plus decisions, not a standalone number.
Use this distinction:
For a boutique consulting firm, the audit should answer bounded questions. Did the firm appear? What exact claim appeared? Which source was visible? Did the claim match an approved source? What should change, and who owns it?
Jungle Roots' July 3, 2026 run shows why the rows matter. Across 18 fixed Perplexity prompts run logged out, headless, and with fresh search, Jungle Roots was cited in 2 and not cited in 16 (dated proof run). That result applies only to Perplexity, that prompt set, and that date. It does not show why either outcome occurred or predict a later result.
If you want the ledger, claim gaps, and owners reviewed together, get the operations audit.
AI visibility means that a brand appears in an AI-generated answer for a defined question and context. The observation is bounded by the prompt, assistant, mode, language, market context, session state, and capture time. It is not a permanent property of the brand.
Visibility is only one field in the audit. Keep these observations separate:
This separation matters. A firm can appear without a citation. A cited page can sit beside an inaccurate claim. A correct description can still omit the evidence a buyer needs. Ahrefs includes mentions, frequency, accuracy, and sources in its audit definition (Ahrefs). This publication's rubric stores those dimensions as distinct fields so a reviewer can inspect what happened.
Treat each answer as an observation. Treat each citation as evidence to inspect, not proof that every nearby sentence is correct.
Check AI visibility by running a saved prompt set under recorded conditions and logging each response in a query-level ledger. Start with buyer decisions, not a loose list of keywords. The prompts need to expose how an assistant understands your category, fit, evidence, alternatives, and branded facts.
This publication's prompt rubric uses these question families:
Write each prompt as a buyer would ask it. Add a market, company type, or constraint only when it changes the decision. Once the set is ready, freeze the wording. A rewritten prompt becomes a new observation and should not replace the old row.
For each run:
ChatGPT search responses can include inline citations and a Sources panel. Record the URL shown by the interface and the claim it appears to support (OpenAI). Do not save only the publication name. The URL and nearby sentence make later review possible.
The ledger should preserve enough context for another person to repeat the prompt and challenge the interpretation. One row represents one captured response.
| Field | What to record | Why it matters |
|---|---|---|
| Observation ID | A stable identifier | Connects findings, captures, and tasks |
| Capture time and timezone | When the answer appeared | Bounds the observation |
| Assistant and mode | Product and visible search mode | Prevents unlike runs from being merged |
| Session state | Signed-in state and new or continuing chat | Preserves a material condition |
| Market and language | The context used for the response | Keeps conclusions inside scope |
| Prompt family | Category, problem, comparison, validation, or branded fact check | Shows the buyer decision being tested |
| Exact prompt | Verbatim input | Makes the observation repeatable |
| Firm mentioned | Yes or no | Records visibility without inference |
| Own-domain citation | Exact URL or none | Separates a mention from first-party support |
| Other citations | Every visible third-party URL | Shows which sources appeared |
| Citation-to-claim note | The nearby claim for each citation | Makes source review possible |
| Claim check | Accurate, inaccurate, partly accurate, or unverifiable | Keeps factual review separate from tone |
| Claim reference | Approved page or named internal owner | Grounds the judgment |
| Missing evidence | Information required by the prompt but absent | Identifies the content gap |
| Next fix and owner | One action and one accountable person | Converts the finding into work |
Attach the raw response and screenshot to the row. The table is the working index. The capture is the underlying evidence.
There is no universal score in this article's rubric because one blended number hides the decision you need to make. A mention rate, an own-domain citation count, a tone label, and an accuracy judgment answer different questions. Combining them requires weights, and the packet provides no approved scoring model.
Report the fields separately instead:
If leaders need a compact view, show counts beside the defined audit scope. Name the assistants, prompt cohort, market, language, conditions, and capture dates. The result then says what the audit observed. It does not pretend to estimate every answer that any buyer might receive.
A score can describe a bounded dataset. It cannot replace the rows that explain why a content change is needed.
ChatGPT can help inspect supplied material and organize findings, but its answer is not a substitute for captured evidence or direct technical checks. ChatGPT search can search the web and show citations when those features are present (OpenAI). That makes it one possible observation environment for an AI visibility audit.
Keep the roles clear. Use the assistant to produce the answer being audited or to help classify saved text. Use approved pages to check brand claims. Use direct site checks for crawl access, internal discoverability, textual content, page experience, and matching structured data. Google says its established SEO practices apply to AI Overviews and AI Mode, with no extra technical requirements beyond eligibility for Search with a snippet (Google Search Central). Google also states that meeting its requirements does not guarantee crawling, indexing, or serving (Google Search Central).
Do not ask ChatGPT to produce a score and accept it as the audit. Ask it a saved buyer question, preserve the output, inspect the citations, and verify each brand claim against its source of truth. The audit remains an evidence process owned by your team.
Find claim-level gaps by splitting the answer into checkable statements and matching each statement to visible evidence and an approved reference. Review one answer at a time. Do not infer why the assistant wrote it.
Use this sequence:
A source problem exists when a cited page carries an incorrect or stale statement. A content problem exists when your approved page lacks a clear, checkable answer. If the approved sources conflict, mark the claim unverifiable and assign a person to resolve the source of truth. Do not guess which version is current.
The claim-level view also prevents a common reporting error. Favorable language is not proof. If an assistant recommends the firm but shows no checkable evidence, record positive framing and weak evidence as two separate findings.
Route each material gap to one owner, one page or source, and one explicit correction. The action should name the finding that triggered it. "Improve AI visibility" is not a task. "Add the missing service eligibility statement to the approved services page" is.
| Observed pattern | Claim-level interpretation | Next content fix | Owner |
|---|---|---|---|
| Firm appears, own site is cited, claims match | Evidence exists for this prompt | Preserve the cited answer and keep the row for comparison | Content owner |
| Firm appears, no own-site citation | The answer names the firm without first-party support | Inspect the cited sources and clarify the relevant approved page | Content owner |
| Firm is absent, other firms are cited | A prompt-level visibility gap exists | Compare the cited claims with the firm's actual offer, then fill a relevant evidence gap | Strategy owner and content owner |
| Firm appears with an inaccurate claim | The answer creates a factual issue | Correct the approved page and identify cited pages that carry the error | Subject owner |
| Favorable framing lacks evidence | Positive wording has no visible proof source | Add or clarify a checkable case study, credential, or named method if one exists | Consulting lead |
| Approved sources conflict | Accuracy cannot be determined | Resolve the source of truth before changing public copy | Subject owner |
The owner must be a person or a role with authority over the source, not a broad department. Keep the triggering observation ID on the task. At review, compare the new capture with the saved prompt and conditions. A later citation is a new observation, not proof that the content edit caused it.
The next fix belongs where the missing or wrong claim should become true, clear, and checkable.
My editorial verdict is that the best tool is the one that preserves your fixed prompts, raw answers, exact citations, claim checks, and ownership fields without forcing them into a vendor score. Tool selection follows the audit method. It does not define the method.
Evaluate any manual worksheet or software product against the same requirements:
A polished dashboard is a weak choice if it hides the underlying answer. A plain ledger is a strong choice if another reviewer can trace every finding to the prompt, capture, citation, and approved reference. Pick the smallest system your team will maintain without losing those fields.
Lead with decisions, then attach the evidence ledger. Summarize mention coverage, own-domain citations, claim accuracy, framing, and decision-blocking omissions as separate findings. State the scope beside them.
A useful finding reads like this:
In the saved comparison prompts, the firm appeared, but its domain was not cited. The captured answers relied on publisher pages for the service claim. The content owner will clarify that claim on the approved services page and preserve the current rows for the next review.
That wording names the observation, the evidence gap, the action, and the owner. It does not claim causation or promise future inclusion. If you want the ledger and operating decisions reviewed together, get the operations audit.
