An AI search audit is a structured investigation of eligibility, business facts, evidence, buyer questions and observed discovery. It is not a way to buy a ChatGPT recommendation. The useful output is a prioritized backlog with a repeatable baseline and clear limits.
Google’s current guidance says its AI Overviews and AI Mode use the same foundational SEO practices as Search, require a page to be indexed and snippet-eligible, and do not require special AI files or special schema. That makes the technical starting point familiar: crawl access, index controls, helpful textual content, internal discovery and accurate structured data.
1. Record the audit scope and baseline
Start with the business decision, not a tool score. Define the services, products, locations, languages and buyer stages being reviewed. Record the audit date, the engines and interfaces observed, and whether each query was signed in, localized or personalized.
- Choose 10–20 real buyer questions across discovery, comparison, risk and action.
- Include branded fact checks and non-branded category questions.
- Keep the exact wording; small query changes can produce different answers.
- Save the answer, cited URLs, date, interface and visible limitations.
- Separate “brand named,” “site cited,” “correct fact,” and “qualified referral.”
2. Audit crawl, index and snippet eligibility
Review important URLs from the outside in. A page should return the intended status, resolve to the intended canonical URL, remain reachable through internal links and expose its main content in usable HTML. Then inspect the controls that can prevent crawling, indexing or excerpts.
| Control | Audit question | Common mistake |
|---|---|---|
| robots.txt | Can the relevant crawler access the page and required resources? | Blocking a crawler while expecting it to read a page-level directive |
| HTTP status and redirects | Does the preferred URL resolve directly and consistently? | Chains, loops, soft errors or conflicting host versions |
| Canonical | Does the declared preferred URL match the page and internal links? | Pointing a useful page at an unrelated or stale canonical |
| Indexing | Is the page indexable and actually indexed where tools provide evidence? | Treating a sitemap entry as proof of indexing |
| Snippet controls | Are nosnippet, data-nosnippet or max-snippet choices intentional? | Restricting the evidence or answer text the business wants surfaced |
| Rendered content | Are key facts present in accessible text after rendering? | Hiding the main answer inside an image, canvas or failed script |
Google documents eligibility and controls for AI features in Search and the effects of robots meta and snippet controls. Eligibility is necessary, but it does not guarantee crawling, indexing, a ranking, a citation or a referral.
3. Separate OpenAI search access from training preference
Do not treat every AI crawler as one setting. OpenAI’s current documentation assigns different purposes to its agents:
- OAI-SearchBot is used for surfacing websites in ChatGPT search features.
- GPTBot is used for content that may be used to improve generative AI foundation models.
- ChatGPT-User can fetch pages for user-initiated actions and is not the control for Search inclusion.
The audit should record the current policy choice for each agent and verify the actual robots response. Allowing search access does not require allowing training. See OpenAI’s official crawler documentation for the current distinctions and published IP references.
4. Build an entity-fact ledger
List the facts that must remain consistent across the website and authoritative profiles. For a local business this may include legal and trading name, category, address or service area, phone, hours, locations, services, credentials and policies. For a store it may include brand, product identifiers, availability, price, shipping markets and return terms.
| Fact | Primary source | Website location | External corroboration | Owner / review date |
|---|---|---|---|---|
| Business identity | Owner-approved business record | About, contact, footer | Relevant business profiles | Named role and date |
| Service coverage | Current operating policy | Service and location pages | Current listings where applicable | Operations owner |
| Product availability | Commerce/inventory system | Product and collection pages | Current merchant feeds where used | Catalog owner |
| Claim or result | Dated evidence artifact | Case, service or article | Source link where public and permitted | Claim owner and expiry |
Use local SEO to repair location discovery and business-profile consistency where geography drives the decision. Use the e-commerce service for catalog, feed and store-measurement issues. Structured data can describe visible facts, but it should not invent or replace them.
5. Audit source-backed proof, not just topical coverage
For each important claim, ask what a buyer or system can verify. Strong pages distinguish first-party operating facts, dated results, demonstrations, cited external facts and opinions. Remove unsupported superlatives, anonymous numbers and copied claims whose source cannot be inspected.
- State the subject, period, geography, currency and metric definition for results.
- Link to the primary source when a current rule or platform feature supports the statement.
- Show methodology and limitations beside a case result, not in a hidden disclaimer.
- Keep authorship, publisher and update dates accurate.
- Review time-sensitive claims on an assigned schedule.
6. Map buyer questions to pages with a real job
Group the question set by intent: definitions, fit, comparison, cost, evidence, risk, location and next step. Assign each question to the page best able to answer it. Do not manufacture dozens of near-duplicate pages for slight query variations.
A useful answer usually gives the direct conclusion, the conditions under which it changes, the evidence source and the next action. It should still serve a human reader if no AI system ever cites it. Google explicitly says there is no special schema required for its AI features; use structured data only when it accurately describes visible content and matches an applicable vocabulary or search feature.
7. Measure citations and referrals as different events
A citation is an observed source reference in an answer. A referral is a visit arriving at the site. A brand mention may include neither. Track them separately, then connect qualified on-site actions only where the available identifiers support the link.
| Layer | Record | Limitation |
|---|---|---|
| Answer observation | Exact query, engine/interface, date, answer and cited URL | Results can vary by time, location, account and wording |
| Bing AI reporting | Citations, cited pages and available query/topic views | Aggregated citation data is not ranking, authority or answer placement |
| ChatGPT referral | Landing page, session and utm_source=chatgpt.com where present | No visit is recorded when a user does not click |
| On-site action | Accepted lead, quote, booking, cart or order under its own definition | Does not prove the AI answer caused the outcome |
| Business outcome | Qualified, sold, paid, fulfilled or retained | Requires CRM or commerce readback and careful attribution |
Bing’s AI Performance documentation states that its citation metrics do not indicate ranking, authority or a page’s role in an individual answer. OpenAI’s publisher FAQ documents the ChatGPT referral parameter and the discovery requirements in its current browser experience.
8. Turn findings into a prioritized audit table
| Priority | Finding | Evidence | Recommended response |
|---|---|---|---|
| P0 — harmful fact | A material business, safety, price or availability fact is wrong | Primary record plus affected page/profile/answer | Correct the authoritative source and affected surfaces; document recrawl limits |
| P1 — eligibility blocker | Important page is blocked, noindexed, broken, noncanonical or snippet-ineligible by mistake | Live response, directive and webmaster-tool evidence | Fix the specific control and verify the final response |
| P1 — unsupported claim | Page makes a material claim without inspectable evidence | Claim inventory and missing/contradictory source | Qualify, source or remove the claim |
| P2 — answer gap | Buyer question has no clear, useful page or important conditions are missing | Question map, page review and observed answer set | Improve the best existing page or create one distinct resource |
| P2 — measurement gap | Citations, referrals and outcomes are combined or unlabelled | Analytics, observation sheet and event definitions | Separate the layers and establish a dated baseline |
| P3 — optional enhancement | Valid markup or presentation can be made clearer | Visible content and validator output | Implement after factual and eligibility issues |
What the audit cannot guarantee
No ethical audit can guarantee that Google, ChatGPT, Copilot or another system will crawl, index, rank, mention or cite a page. It cannot make a third-party model repeat facts consistently, prove causation from a citation to a sale, or create authority through markup alone. Search and answer systems change, and their outputs can vary between otherwise similar checks.
The defensible goal is narrower: remove accidental eligibility problems, make important facts and evidence easier to verify, answer real buyer questions well, and measure observed visibility and qualified business outcomes over time. The Pandorium AEO / AI visibility service applies that scope without a placement guarantee.
Official references reviewed
- Google Search Central: AI features and your website
- Google Search Central: robots meta, data-nosnippet and X-Robots-Tag
- OpenAI: overview of OpenAI crawlers
- OpenAI: publishers and developers FAQ
- Bing Webmaster Blog: AI Performance in Bing Webmaster Tools
References were reviewed on 7 October 2026. Product interfaces, crawler policies and reporting fields can change. Recheck the current first-party documentation when implementing or repeating the audit.