Who does AI search recommend for the service you sell in the market you serve? Start there, not with a screenshot of an unrelated query. A useful audit identifies the businesses selected, the reasons given and the sources supporting those reasons.
For owners, CEOs and marketing leaders evaluating local visibility, the deliverable should be a defensible audit—not a trophy screenshot. The method below helps you compare providers, preserve evidence and decide what to fix before buying a larger reporting program.
What counts as an AI search recommendation?
For an audit, count a recommendation when an answer explicitly presents a business as an option for the stated need. Record citations separately: a linked source is evidence to inspect, not automatically an endorsement of that source’s owner. The useful result is a shortlist tied to a particular prompt and test setup, with each selection’s support checked.
Buy the audit scope that answers your decision
Before comparing software or agency proposals, name the decision the report must support. Checking whether your company appears is a different assignment from explaining why a competitor was selected or connecting visibility to qualified inquiries.
When reviewing Percepture’s GEO services or another provider’s proposal, use the same evidence requirements. Compare the promised deliverables, not the confidence of the sales presentation.
Six useful audit scopes to compare
| Scope | Audience fit | Evidence to request | Technical constraint | Effort or reliability check | Tradeoff | Verification source |
|---|---|---|---|---|---|---|
| Spot check | Owner exploring one buyer question | Complete answer and exact prompt | One recorded setup | Keep conclusions narrow | No repeatability assessment | Original answer capture |
| Controlled baseline | Marketing team choosing priorities | Repeated runs under a fixed setup | Separate sessions and conditions | Request all completed answers | More review work | Run log and captures |
| Cross-platform comparison | Leader comparing answer surfaces | Matched buyer prompts on selected platforms | Record each platform’s exposed settings | Keep platform denominators separate | No single blended ranking | Platform-specific records |
| Geographic comparison | Multi-location operator | Matched service questions by market | Separate requested city from test location | Hold other conditions steady | Additional market-level analysis | Prompt and location records |
| Recurring tracking | Team evaluating changes over time | Baseline, follow-up captures and change log | Document setup changes | Retain historical exports | Ongoing collection cost | Dated answer archive |
| Inquiry attribution | Leader evaluating commercial relevance | Referral evidence and qualified-inquiry records | Define matching and privacy rules | Separate observed referrals from self-reports | Not a complete customer journey | Analytics and permitted inquiry records |
The important comparison is between a screenshot and an auditable record. Use a screenshot to inspect wording. Use a record containing the prompt, setup, answer, sources and coding decisions to evaluate what that wording means. Do not purchase a recurring dashboard until you know which of those records you can export.
Apply the AI Citation Reproducibility Protocol
Use this protocol as a repeatable audit worksheet, not a prediction of which companies will appear. Its purpose is to make the collection process inspectable. Repeating the setup and reproducing identical answer text are different standards.
1. Freeze the buyer question before collecting answers
Choose a real service, an explicit market and a plausible buying situation. Leave your business name out of the discovery prompt unless the assignment is specifically to evaluate branded consideration. Keep branded and unbranded tests in separate groups.
For a residential heating contractor, two distinct buying scenarios could use these prompts:
Which companies should I consider for residential heat-pump installation in Newark, New Jersey? Explain why each fits this service and cite supporting sources.
Which companies should I consider for emergency furnace repair in Newark, New Jersey? Explain why each fits this service and cite supporting sources.
Do not combine those scenarios into one service score. A recommendation for installation should not stand in for evidence about emergency repair.
2. Record the setup without guessing hidden details
For every attempt, record the exact prompt, platform, displayed model label when available, exposed search settings, requested city, tester location, session context, and date and time with a time-zone offset. Distinguish a setting that is not shown from one that is known to be disabled.
Use fresh conversations where the interface permits. If a test depends on earlier messages, retain that conversation context as part of the input. Do not quietly compare a personalized follow-up conversation with a first-turn discovery prompt.
3. Repeat each pairing and preserve the full record
For a starter design, choose two platforms, two buying scenarios and three fresh-session runs for each pairing. That produces 12 planned answers per market. Keep each pairing separate in analysis. This is a workload example, not a statistical sufficiency rule.
Save the full answer, visible citations and a screenshot together. Preserve failures and answers that contain no shortlist rather than keeping only favorable results. Use permitted collection methods, and exclude private account details or customer information from shared reports.
4. Code mentions, recommendations and support separately
For each business, record whether it was merely mentioned, explicitly recommended, or linked as a source. Copy the recommendation wording. Then open the cited page and check whether it supports the specific reason given, such as the service offered or area served.
Use separate fields for an inaccessible source and an accessible source that does not support the statement. Neither should be silently counted as checked support. A persuasive explanation is not a substitute for inspecting its evidence.
Calculate recommendation frequency within each fixed setup: completed answers recommending the business divided by all completed answers in that setup. Include completed answers with no shortlist; report failed attempts separately. Keep source-support checks beside that frequency rather than hiding them inside one composite score.
Can this result support a decision?
- Traceable input: The exact prompt and relevant settings are preserved.
- Complete collection: The record includes unfavorable answers and collection failures.
- Checked support: Each cited reason has an inspectable source assessment.
- Bounded conclusion: The finding names its service, market, platform and collection period.
Write the conclusion at the same level as the evidence. Report how often a business was recommended within the recorded setup. Do not rename that finding “the best company,” “market share” or a permanent ranking.
Investigate differences before assigning a cause
When two answers disagree, first check the inputs. Was the service identical? Was the same city requested? Did the conversation include earlier instructions? Were the platform settings and collection period comparable? Treat these as candidate explanations to investigate, not automatic causes.
For a follow-up audit, retain the baseline prompts and keep a separate change log for website edits, business-information corrections and campaign activity. If several things changed together, avoid crediting one action with the entire difference.
Compare business overlap separately from source overlap. Two answers could name the same business while pointing to different supporting pages. Review both the selection and the quality of its stated support before deciding which content or information needs attention.
Turn the finding into a specific work order
Do not start with “we need more AI content.” Start with the gap you can document. If a service or coverage area is unclear on your own site, define the correction and its owner. Use content marketing planning to address that reader question rather than multiplying near-duplicate pages.
If the audit identifies an inaccurate third-party description, investigate that source directly. Consider digital PR as a separate editorial workstream; do not equate a placement with a guaranteed recommendation.
For website publishing decisions, consult Google’s people-first content guide. When reviewing markup, use Google’s introduction to structured data as an implementation reference. Evaluate the resulting pages and markup separately from observed AI recommendations.
For commercial measurement, define a qualified inquiry before reporting one. Use attribution and analytics planning to distinguish recorded referrals, customer self-reports and later sales outcomes. Keep those measures separate from answer appearances.
When several channels need coordinated work, frame the assignment through omnichannel marketing: one buyer need, clear owners and a shared measurement plan. Do not add unrelated channel activity simply to make an AI visibility proposal look comprehensive.
Ask a provider to show the record behind the score
Before signing, ask for an example export and the definitions used to classify answers. You should be able to inspect the underlying material without relying on a presentation slide.
- Which services, markets and buyer questions are included?
- How are mentions, recommendations and citations distinguished?
- Which setup details are recorded, and which are unavailable?
- How are empty answers, failures and unsupported citations reported?
- Can you export prompts, captures, source URLs and historical records?
- What collection, review and reporting work does the fee cover?
- Which proposed changes will be tested, and how will inquiries be evaluated separately?
Ask the provider to demonstrate the coding rules on an ambiguous answer. That discussion is more useful than a confident score whose underlying decisions you cannot inspect.
Questions about evaluating AI recommendations
How are AI citations different from recommendations?
Use separate definitions in your audit: a citation is a linked source; a recommendation explicitly presents a business as an option for the buyer’s need. Record both. Check whether the cited page supports the recommendation instead of counting every linked domain as a recommended provider.
Should I treat an AI answer as a stable ranking?
No. Treat an individual answer as one observation under a recorded setup. To assess consistency, repeat the same prompt under matched conditions and report how often each business is recommended. Preserve answers without a shortlist, and avoid turning answer order into a permanent ranking claim.
How should prompt, geography and model be compared?
Treat them as separate test conditions. Record the exact prompt, requested market, tester location and displayed model label when available. Change one condition at a time where practical. If several conditions differ, describe the comparison without assigning the difference to a single cause.
What sample size makes an AI visibility comparison useful?
Set the collection plan around the decision and the level of certainty you need. Use a small pilot to inspect the workload and coding rules before expanding. Do not present the 12-answer planning example as a statistical threshold or evidence of market-wide recommendation frequency.
What should SEO, PR and content teams do with the findings?
Assign work against a documented gap: unclear service information, inconsistent business details, an unsupported comparison or a source worth investigating. Keep each proposed action distinct. Compare follow-up evidence with the baseline, and avoid promising that any individual change will secure an AI recommendation.
How should AI visibility connect to inquiries?
Track qualified inquiries separately from answer appearances. Define qualification, preserve referral evidence where available and record customer-reported discovery as a separate field. Do not label every new inquiry an AI-generated lead or treat a recommendation count as proof of revenue.
Choose the evidence before the reporting package
Bring one service, one target market and the buyer question you want answered. Compare collection scope, source review and reporting access before committing to ongoing monitoring.
About Bob Generale
Bob Generale is President of Percepture. His work spans organic search, AI search visibility, digital PR, audience targeting and sales intelligence. His editorial approach starts with the buyer’s question, the operating requirements and the evidence an organization can provide. He connects Percepture’s research, strategy and publishing teams around earning qualified conversations rather than celebrating impressions in isolation. In this guide, that perspective centers on a practical distinction: appearing in an answer, being recommended and generating a qualified inquiry are separate things to evaluate. The audit worksheet gives business owners and marketing leaders a way to inspect a provider’s methodology, ask for the underlying records and connect proposed work to a defined business decision.
