Define the buyer questions and the audit boundary
Choose questions a potential buyer would ask before selecting a service: what the service includes, who it fits, how to compare approaches and what evidence to request. Include questions without your brand as well as questions about it. Keep branded accuracy checks separate from unbranded discovery.
Write a fixed prompt set for each market and language. Record the exact wording and why the question matters. Avoid a large collection of near-identical prompts that makes a percentage look precise without covering different decisions. Do not put private customer, prospect or account information into a public model.
Log the answer and the limits of what you observed
Load the complete ten-answer fictional register or add your own observations. Record exact prompt wording when you adapt the library; preserve engine/context and unassessable source panels. Build a report, export CSV or print.
Record exact prompt and buyer purpose.
Include engine, model if visible, date, language and context.
Separate mention, exact citation, unavailable source panel and next check.
Private, ungated workbook. Add up to 50 entries, load the complete fictional example, build a report and export CSV or print. Inputs stay in this page and disappear on reload. No automatic audit, score or verified outcome is produced.
Entry 1
Entry 1
No completed entries. Unknown is not zero.
No account, upload or data collection. Inputs disappear when this page reloads. Build and export the report before leaving. Without JavaScript, use the blank fields and the completed example in the article manually.
Make each observation reproducible
For every run, record engine and visible model/version, timestamp, language, location if known, signed-in or clean context, exact prompt and whether browsing or sources were available. Preserve a concise answer excerpt or reference where appropriate. If a setting or model version is not disclosed, write unknown.
Personalization, model updates and answer variability limit comparisons. Repeat the same questions under matched conditions across documented review dates; do not compare one personalized answer with an unrelated clean session as though only your page changed. Manual observation is a sample, not a complete census of AI search.
- Capture whether the brand is mentioned and what is said about it.
- Record each cited URL exactly and verify whether it supports the associated claim.
- Mark source-panel availability separately from citation presence.
- Log factual errors, unsupported statements and inaccessible evidence without filling gaps by inference.
Keep mentions, citations, referrals and bookings separate
A mention means the answer names the business. A citation means a source link identifies a particular URL. A referral means an observed visit from an AI source. A genuine booking is a separate event that needs scheduling evidence. None of these proves the next stage occurred.
| Measure | What to record | What it does not establish |
|---|---|---|
| Brand mention | Exact observed name and statement in the sampled answer | A link, recommendation or visit |
| URL citation | Cited URL and the statement it supports | Traffic, endorsement or a new booking |
| AI referral | Source-aware session or landing evidence with coverage limits | That a particular prompt caused the visit |
| New website booking | Original-created booking, reconciled source and landing context where available | That all earlier AI exposure is known |
Check factual accuracy and ordinary search eligibility
Review your own service explanations, organization facts, authorship, source dates and language consistency. The page should answer the buyer’s question plainly, show useful original help and distinguish approved facts from examples. Accurate visible content and matching structured data make the information more coherent; they do not force an engine to cite it.
Google’s AI-features guidance says the normal SEO foundations apply. Supporting pages need to be indexed and eligible to appear with a snippet. There is no special AI file or schema required for these features. Check access, indexability, canonical references and internal discovery before adding speculative technical files.
A page being crawlable is not proof of indexing; indexing is not proof of citation. Do not block useful search access as an untested visibility tactic or treat an llms.txt file as a ranking switch.
Review answer gaps and competing sources carefully
Inspect the pages actually cited for your buyer question when source links are available. Note which decision they help with, their evidence and what your existing page misses. A result that names a competitor does not establish why that competitor appeared. An unavailable source panel does not prove there were no citations.
Improve the relevant existing page when it can answer the question well: clarify scope, direct answers, responsibilities, sourced distinctions and useful examples. Create a new URL only for a distinct buyer job. Do not copy a cited page, invent client proof or add broad claims merely to resemble a sampled answer.
Use a fictional observation to choose the next check
The example below is invented. It demonstrates coverage limits, not a measured AI result or a claim about any named engine.
| Question | Observed sample | Unknown | Next action |
|---|---|---|---|
| What should a managed campaign include? | Invented answer mentions the business; source panel unavailable | Exact citations and any resulting visits | Record citation state as unassessable; check visible service accuracy |
| How should we compare audit scopes? | Invented source panel cites a guide with clear responsibilities | Whether the citation produced a booking | Review the useful decision help; keep outcome measurement separate |
Connect AI visibility to the acquisition record
Use source-aware analytics and original scheduling records to distinguish new website-acquired calls from outreach-led calls that later visit the site. Retain known landing pages and booking identifiers only with the necessary access and purpose. A Direct visit with no source cannot safely be assigned to AI, and one ChatGPT booking does not establish a conversion benchmark.
Exclude identified internal or audit visits and retain attribution gaps explicitly. Compare matched periods and mature booking cohorts, including cancellations, calls held and qualification. Report observation coverage alongside any share of sampled answers; never label the sample as market-wide AI visibility.
Finish with a bounded action plan
Rank proposed actions by verified factual risk, relevant buyer value, evidence strength, effort and dependencies. Name the page, exact change, responsible owner and next observation date. Correct a demonstrated false claim promptly; test an uncertain visibility hypothesis with a baseline rather than presenting it as a proven cause.
The worksheet below is a manual log. It does not query models, crawl your site or calculate a proprietary AI score. If you need implementation support, distinguish the audit from ongoing SEO and AI-search work and agree the scope separately.
A reusable buyer-question protocol
Choose a small fixed set that represents actual buying decisions, then run the same wording under recorded conditions. The following prompts are a reusable template, not a claim about searches performed or answers currently produced. Replace bracketed terms with your approved public scope. Keep branded accuracy separate from unbranded discovery.
| ID | Question template | What to inspect |
|---|---|---|
| P1 · branded facts | What does [provider] offer to [buyer type] in [market]? | Service, audience and facts supported by sources |
| P2 · discovery | How should a [buyer role] find help with [specific problem]? | Whether a suitable provider is mentioned and why |
| P3 · scope | Does [provider] include [boundary item] in [service]? | Accuracy of inclusions and exclusions |
| P4 · evaluation | What should we ask a provider before commissioning [service]? | Decision criteria, responsibilities and evidence |
| P5 · comparison | When does [approach A] make more sense than [approach B] for [buyer context]? | Trade-offs and whether assumptions are disclosed |
| P6 · responsibilities | What must our team provide before [service] can begin? | Access, approvals and client responsibilities |
| P7 · constraints | What can go wrong if [dependency] is missing? | Whether the response explains genuine limits |
| P8 · commercial facts | What are [provider]’s publicly stated terms for [service]? | Current source and exact currency/scope, not inferred prices |
| P9 · proof | What evidence should we verify when evaluating [provider]? | Real permitted proof versus invented outcomes |
| P10 · language | Can [buyer] evaluate [service] information in [language]? | Actual accessible page and correct language relationship |
A complete fictional answer and source-panel review
This entire response is invented for training: “Example Provider helps industrial firms research suitable accounts and manage email prospecting. It also includes unlimited native-language cold calling. The buyer handles proposals and closing. [1]” Fictional source panel [1]: https://example.com/services/. The fictional source page states that email prospecting is included and calling requires a separate quote.
The provider name is a mention. The source-panel URL is a citation if that panel is actually visible. The calling statement is an accuracy error even though the answer includes a citation: the cited page does not support it. The source is an example-domain placeholder, not a real provider or page inspected in this audit.
Record the precise unsupported statement and the conflicting primary fact. Check the actual public source before proposing a correction. If your own page is wrong, correct that page through normal approval and validation. If the page is already clear, preserve the observation, inspect other cited sources where available and repeat the matched test; do not manufacture another service claim to match the answer.
A completed two-pass observation example and its denominators
All observations here are fictional. Run prompts P1–P5 twice, in fresh signed-out threads, in one language and with the same documented engine/context. If a model name or location is not shown, record unknown instead of guessing. The ten-row example can be loaded into the workbook; no live engine query is performed by this page.
| Run / prompt | Source panel | Mention | Citation | Accuracy review |
|---|---|---|---|---|
| R1 P1 | Available | Yes | example.com/services/ | Checked scope accurate |
| R1 P2 | Available | No | No | Not assessed |
| R1 P3 | Unavailable | Yes | Unassessable | Calling-included claim conflicts with fictional source |
| R1 P4 | Unavailable | No | Unassessable | Not assessed |
| R1 P5 | Available | No | No | Not assessed |
| R2 P1 | Available | Yes | example.com/services/ | Checked scope accurate |
| R2 P2 | Available | No | No | Not assessed |
| R2 P3 | Unavailable | No | Unassessable | Not assessed |
| R2 P4 | Unavailable | No | Unassessable | Not assessed |
| R2 P5 | Available | No | No | Not assessed |
Describe observations without inventing an AI rank
This example has three mentions in ten observed answers. Six answers expose a source panel; two of those six cite the example provider. Four panels are unavailable, so their citation state is unassessable, not zero. State the counts and denominators together. The six assessed panels are not necessarily representative of all ten answers or the wider market.
A second pass can vary without a page edit. A changed answer after an edit still does not isolate the cause; model, retrieval, source availability and context may also change. Do not report a small prompt sample as universal share of voice or a stable recommendation rank. A proprietary checker’s score describes its own method, not an engine requirement.
Export observations into a report with exact prompt, visible engine/model, date, language, known context, source-panel state, mention, exact citation, factual review, evidence reference and next action. Keep actual AI-referral sessions and source-reconciled bookings in separate reports. A citation is not a visit; a visit is not a qualified meeting.
Connect source corrections to a controlled page review
For each actionable finding, record source URL, statement checked, current primary fact, proposed change, owner, approval and retest. For Google AI features, normal SEO foundations still apply: supporting pages must be indexed and snippet-eligible; no special AI file or schema is required. Google includes AI-feature activity in overall Web performance rather than a clean separate AI-impression bucket.
A manual buyer audit can be useful without pretending to be an automated grader. This workbook neither crawls submitted URLs nor queries models. It helps you retain the evidence, uncertainty and action trail that a bare score cannot provide.
Choose the next action the evidence can support
Preserve a dated prompt set and observation log. Improve useful, accurate buyer answers and technical access; judge business outcomes separately.
- Correct demonstrated factual errors.
- Keep unknown source-panel states unassessable.
- Compare matched observation conditions.
- Validate changes and review mature visits and booking cohorts.
Frequently asked questions
What should an AEO audit measure?
Buyer-question coverage, sampled brand mentions, exact cited URLs, factual accuracy and source-panel availability, alongside ordinary search eligibility, observed AI referrals and genuine new bookings. Keep each measure and its coverage separate.
What is the difference between an AEO audit and an AI visibility audit?
The names overlap in practice. Define the actual scope: engines, prompts, languages, source inspection, technical review and outcome reconciliation. The label alone does not tell you what evidence is collected.
Does Google require special AI schema or llms.txt?
Google’s cited AI-features guidance states there is no special schema or AI file requirement. Normal SEO foundations and indexed, snippet-eligible supporting pages remain relevant. That does not guarantee inclusion or citations.
How can we compare two audits fairly?
Use the same buyer questions, engines, visible model versions, languages and known context, with recorded dates and evidence coverage. Report unavailable settings and source panels explicitly; answer variation still limits causal conclusions.
Sources and methodology
Google supports the technical eligibility guidance; the publisher articles inform the method comparison. The manual worksheet is Beespoke planning help. All sample answers are invented. No live model benchmark, market-wide visibility score or causal booking result is claimed. Sources reviewed October 27–28, 2026.