An AI search audit should produce a prioritized, verifiable backlog rather than a list of general recommendations. Start by documenting the buyer question and the observed evidence, then distinguish confirmed failures from untested explanations. Resolve incorrect commercial facts and broken buying paths before expanding lower-impact content work. Give each action an owner, a dependency and a repeatable acceptance check. The value of the audit is the decision it enables: what to fix, why it matters and how your team will know that the specific problem has been addressed.
1. Define the sample before collecting answers
Record the market, language, platforms, interfaces, question IDs and observation dates. Include branded and unbranded questions across discovery, evaluation and buying intent. Keep the sample stable long enough to compare like with like.
A failed request, unavailable platform and a completed answer without a brand mention are different outcomes. Your records must preserve that distinction. Do not fill missing observations with zero visibility.
2. Reconcile the facts a buyer needs
- Brand name and legal operating entity are consistently identified.
- Actual company address and contact routes are correct.
- Products and services have clear definitions and availability.
- Material specifications and claims have an approval owner.
- Third-party references are linked or recorded where relevant.
If two pages disagree on an offer, fix the source of the discrepancy first. Schema should describe the same facts a person can read on the website.
3. Inspect the website path
| Check | Record | Decision |
|---|---|---|
| Access | HTTP behavior, robots rules and hosting restrictions | Fix a confirmed block or failure. |
| Preferred URL | Canonical and redirect chain | Consolidate competing versions. |
| Readable content | Returned text, headings and relevant links | Make the key explanation available. |
| Discovery | Navigation and sitemap placement | Connect an orphan page to a relevant topic. |
| Conversion | Inquiry steps, errors and confirmation | Repair the point where the buyer gets stuck. |
Local checks and production checks serve different purposes. A source file can prove a link is generated correctly; it cannot establish a firewall rule or a live search index state.
4. Test the usefulness of the answer
Can a buyer understand who the offer suits, compare the relevant options and identify the next step? Is the content supported by product evidence or a clearly attributed source? Does the page add something beyond a generic definition?
Record the exact missing decision. “Improve content” is too broad to assign. “Explain supported deployment modes on the product page, approved by the product owner” creates a concrete task.
5. Turn observations into an owned backlog
| Field | What to write |
|---|---|
| Finding | The observed gap and the affected question or URL |
| Evidence | Dated record or source supporting the finding |
| Action | Specific change and why it helps the buyer |
| Owner | Person responsible for approval and implementation |
| Acceptance | What can be checked when the task is complete |
| Review | When to repeat the relevant observation |
Prioritize blocked buying paths, incorrect facts and important unresolved questions. The $990 TANTU AI diagnostic uses a defined 90-observation plan and provides a prioritized action discussion. It does not replace an unlimited technical audit or guarantee a recommendation.
Triage findings by consequence and certainty
Use three practical queues. Immediate corrections contain confirmed failures that prevent a relevant inquiry or materially misstate the offer. Planned improvements address important unanswered buying questions with available evidence. Investigation items contain uncertain causes, incomplete records or inconsistent observations. This keeps an alarming but unverified interpretation from displacing a concrete problem your team can fix.
Hypothetical example: a quote form returns an error during a repeatable test, a product page gives conflicting deployment details and one sampled answer omits the brand. Repair the form and reconcile the approved product facts first. Preserve the answer omission for investigation, including its question and collection conditions. The omission alone does not establish a technical block or justify a sitewide rebuild.
Define closure before assigning the task
An acceptance check should test the original problem directly. For a broken form, repeat the relevant path and confirm the expected submission handling. For a conflicting fact, compare the corrected public pages with the approved source. For a missing buyer explanation, have the responsible reviewer verify the decision and its limits. “Page updated” is a delivery status, not proof that the finding was resolved.
Record the result, date and any remaining limitation. Keep website verification separate from platform observation: a corrected page can be checked immediately, while an external answer may continue to vary. If the root cause is still unknown, close only the work that was verified and leave the investigation open. A short evidence trail makes subsequent reviews faster and prevents the same generic recommendation from recurring without progress.
Add the property’s generative AI setting to the access review
For Google, review the Search Console generative AI inclusion setting, including parent-property inheritance, alongside crawl and indexing checks. Record the setting observed and the property it applies to. An audit recommendation should explain an actual mismatch with the owner’s intended policy, rather than instructing every site to change the default.
Keep training permissions, search eligibility and user-triggered retrieval separate in the audit. Different providers document different controls; apply the relevant official guidance and confirm production behavior before marking an access issue resolved.
Google: Search generative AI control · OpenAI: agent purposes
Common questions
Can we use this without hiring TANTU AI?
Yes. It is a planning checklist. Your team can use the structure for an internal review.
Should we audit every possible question?
No. Begin with a justified, commercially relevant sample. Explain its limits instead of implying exhaustive coverage.