AI Search / Practical guide
What should an AI visibility audit actually deliver?
Before commissioning an audit, I want to know what it will help the business decide. Here is what I would expect in the scope, evidence and handover.
Buy a decision, not just a visibility score
My first question when reviewing an audit proposal is simple: what will the business be able to decide when this is finished? A colourful dashboard can show that an answer mentioned a competitor. It does not, by itself, tell your team whether to fix a service page, correct a business listing, investigate access or leave the website alone.
The report should distinguish what the reviewer saw from what they think caused it. A recorded answer is evidence; the explanation for it may still need checking.
This guide is a buyer’s checklist for commissioning or reviewing the work. For the scope of my own service, see the AI visibility audit page. Agree the actual deliverables before booking.
1. Agree what the audit will actually examine
Start with the business, not the software’s default prompt list. Which services or products matter? Who is buying, in which market, and what are they trying to establish before making contact? A local service business and an ecommerce retailer need different questions, comparisons and source checks.
Name the platforms and modes being tested. An AI Overview, an answer in AI Mode and a conversation in another product are different checks, not one AI ranking.
- Commercial focus: the services, product categories or locations included, with exclusions.
- Question set: the exact buyer questions, how they were selected and which buying tasks they represent.
- Research conditions: platform, mode, language, market, dates and any relevant account or session context.
- Coverage: planned checks, repeat checks, incomplete answers and unavailable features.
- Business access: whether analytics, Search Console and enquiry records are available, and what remains unverified without them.
I would rather see a smaller set of relevant questions than hundreds of slight rewrites. If a provider quotes a fixed prompt count, ask which services and buying decisions it covers.
2. Look for these six deliverables
Look for these six things in the proposal. The format matters less than whether your team can use the work.
| Deliverable | What makes it useful | What to question |
|---|---|---|
| Scope and question register | Exact questions mapped to services, markets and buying tasks, with dates and conditions. | A generic question list with no explanation of relevance. |
| Traceable observation log | Answer excerpts or private captures, source URLs, context and missing checks. | Percentages without a denominator or any inspectable observations. |
| Representation and source review | Specific inaccurate facts, unsupported claims or source conflicts, checked against verified business information. | Treating every absent mention as misinformation. |
| Relevant competitor comparison | The same research conditions and a reason each competitor belongs in the comparison. | Comparing unrelated businesses or assuming citation proves superiority. |
| Prioritised fixes with evidence | Each recommendation has evidence, an owner, dependencies and a completion check. | A long list of generic SEO tasks without a connection to findings. |
| Handover and recheck plan | Editable working files, a review discussion and a defined way to assess changes. | A PDF with no explanation of implementation or repeat measurement. |
What I would ask to see
Scope and question register
Review question: Can you trace each question to a service, market or buying task?
For example, if enquiries come from two countries, ask which questions and platforms will be checked in each. A single total can hide a gap in one market.
Traceable observation log
Review question: Can you inspect the answer behind a reported percentage?
Ask the reviewer to open one recorded answer and show how it contributes to the reported percentage. Check where failed or unavailable answers are counted.
Representation and source review
Review question: Which specific business fact is inaccurate, and who verified the correction?
If an answer gives an old price, ask what confirms the current price and who can update the page or profile involved. An absent brand mention is a different issue.
Relevant competitor comparison
Review question: Were comparable businesses checked under the same recorded conditions?
Ask why those competitors were chosen. A national chain and a local specialist may answer different buyer needs even if both appear for the same broad question.
Prioritised fixes with evidence
Review question: Can your implementer act on one finding without guessing the requirement?
Ask to see one proposed fix: the affected page, what is wrong, who would change it and how the team would check the result. Section 4 shows an example.
Handover and recheck plan
Review question: Who receives the working files, answers questions and owns the next review?
Ask what you receive besides the PDF, whether your developer can ask questions and who will check completed changes. Get any follow-up fee in writing.
Keep mentions, citations and business results separate
A brand mention, a source link and an enquiry are different events. Record them separately so the report does not count every appearance as a business result.
Ask for the completed sample size beside any percentage, together with the checks that failed or could not be completed. My guide to measuring AI search visibility explains the distinction in more detail. If an enquiry cannot be connected to an AI visit, leave that connection unconfirmed.
Show the evidence without sharing private information
You should be able to check a finding without exposing account or customer details. Agree who can see the records and whether a redacted screenshot is enough.
If no source is shown, say so. Even when a citation is visible, do not assume it was the answer’s only source.
3. Check the website before recommending changes
The audit should check relevant website conditions, but I would not accept a missing citation as proof of a crawling fault. First establish which URL and answer are being discussed, then examine access, indexability, visible content and the match with the buyer’s question.
For Google’s AI features, the Search Central guidance ties eligibility to indexing and snippet eligibility. It does not require a special AI schema or AI text file. That is guidance for Google’s Search features, not a claim that every other answer product follows identical rules.
If the page is already eligible, a useful finding may concern incomplete service information, conflicting business facts or poor internal discovery. If those checks pass too, I would ask for more evidence before changing the page. The indexed but missing from AI answers guide provides a step-by-step diagnostic route.
A broken link gives your team something specific to fix and test. Whether an AI system recommends the business afterwards is a separate result.
4. Ask to see one finding written as a real task
Here is a fictional example of how I would write the task.
Example: an answer describes the wrong service area
Observed: a saved answer describes Example Advisory as serving only one city. The displayed source is an old public directory profile. The business owner confirms that the service now covers a wider, specified area.
What remains unknown: we cannot tell whether that directory was the only source used or whether correcting it will change a later answer.
Recommended work: the business owner verifies the current service area; the profile owner requests a factual correction; the website editor checks that the relevant service page states the same accurate scope. Do not add locations the business does not actually serve.
How I would check the fix: the corrected directory record and website wording can be inspected. Record the URLs, review date and any third-party request still pending.
Recheck: repeat the original question under comparable recorded conditions. Keep the result in the log even if the answer has not changed or is unavailable.
This is more useful than a ticket saying “improve entity authority”. Your team knows the fact to verify, who can act and what constitutes a completed correction. If the answer itself remains inaccurate, the business-fact correction guide explains how to preserve the evidence and separate source corrections from platform feedback.
5. Decide who owns the work after the audit
A finding is not implemented just because it appears in a report. Before commissioning the audit, agree whether the fee includes diagnosis only, a handover session, implementation support or a later review. Those are separate responsibilities, even when the same specialist can provide them.
I would want the top priorities written so your existing writer, developer or agency can use them. For each task, name the person doing it, what they need first and how the fix will be checked. The SEO audit prioritisation guide shows how to turn a backlog into work that fits your team’s capacity.
- At handover: review the strongest findings, limitations, editable files and unanswered questions.
- Before implementation: assign owners, approvals, access and completion checks; protect sensitive evidence.
- After changes: confirm the repairs separately from any change in observed AI answers.
- At review: compare a consistent question panel and discuss suitable enquiries as well as visibility.
One-off diagnosis and ongoing AI search monitoring serve different needs. Before paying for regular monitoring, ask what you will do differently with the next report.
Before you approve the proposal
Use the interactive checklist below to check what you have agreed with the provider.
Be wary of guaranteed recommendations or a standard prescription for more pages and backlinks. Ask how the proposed work follows from what the audit found.
Is the audit scope ready to approve?
Use this checklist to spot gaps in the proposal before you approve it.
Common questions about AI visibility audit deliverables
Is an AI visibility audit the same as an SEO audit?
No. An AI visibility audit starts with answers, business representation and visible sources. An SEO audit looks more broadly at the website and search performance. Some findings will overlap.
How many prompts should the audit include?
Agree the services and markets first, then the questions and repeat checks. There is no universal minimum. Ask how the provider chose the questions and how they report incomplete checks.
Should an audit include implementation?
Only when implementation is explicitly included in the scope. Diagnosis, content changes, technical repairs and ongoing monitoring can be separate engagements. A useful audit should still provide a handover your existing team can act on.
Can an audit guarantee more citations or leads?
No. It can document observations, identify supported improvements and define how to assess subsequent work. Answer selection belongs to the platform, and enquiries depend on more than a citation. Neither outcome should be sold as guaranteed.
What if the audit finds no clear technical problem?
That is a valid finding. Review the relevance of the question set, the accuracy of business information and the usefulness of the page. If there is no supported gap, avoid reactive changes and agree what evidence would justify further investigation.
Sources
Google’s AI features guidance was checked on 16 October 2026. The checklist above is how I would assess a proposal; agree the actual deliverables with your provider before booking.