Field guide for local SEO agencies
How agencies can audit local-business visibility in AI answers.
A useful AI visibility audit is a reproducible field sample: a frozen client brief, balanced prompts, exact observations, cited sources, and an implementation queue. It is not a permanent ranking and it cannot promise future mentions.
Freeze what you test, record what you observe, separate inference from fact, and rerun only after public changes have had time to be discovered.
01 · DEFINE THE MEASUREMENT
Measure a dated answer sample—not an invisible universal rank.
An AI visibility audit documents whether, where, and how a business appears across a defined set of answer interfaces and buyer questions. The result belongs to that exact field window. Outputs can change with wording, model or interface updates, location, account context, and time.
The audit should answer four practical questions:
- Does the business appear when a buyer has not already named it?
- Which competitors appear for the same intent?
- Which public sources or facts appear to support the answers?
- What can the client or agency change and verify next?
02 · FREEZE THE CLIENT BRIEF
Write down the entity and market before writing prompts.
A frozen brief keeps the research from drifting toward whichever result looks most interesting. It also makes the eventual rerun comparable.
03 · BUILD THE PROMPT MATRIX
Balance buyer intent instead of repeating one recommendation question.
Use natural language that a plausible buyer could use. Keep most questions non-branded so the test measures discovery, then add a smaller branded set to inspect factual accuracy and trust.
| Intent | What it tests | Illustrative pattern |
|---|---|---|
| Discovery | Unprompted category presence | Who provides [service] in [market]? |
| Comparison | Competitive framing and evidence | Compare reputable [service] companies in [market]. |
| Trust | Credentials, reviews, and corroboration | Which [service] providers in [market] show evidence of [approved criterion]? |
| Purchase | Availability, process, and price context | How should I choose a company for [urgent or high-value service]? |
| Problem-aware | Whether the brand connects to the buyer's symptom | What should I do when [specific problem] happens in [market]? |
| Branded | Fact accuracy and entity clarity | What services does [business] offer, and where? |
Do not lead the system toward a desired answer, ask the same question with cosmetic rewrites, or silently remove unfavorable observations. A compact balanced set is more defensible than a large biased one.
04 · KEEP AN EVIDENCE LOG
Make every conclusion traceable to a recorded observation.
For each run, preserve enough context for a reviewer to understand what happened. Store no private customer data and do not treat the interface's prose as a fact until it is checked against an authoritative public source.
- Exact prompt text and sequence number
- Interface and model name when displayed
- Date, time zone, locale, and signed-in or signed-out state
- Whether the client and each comparison business appeared
- The answer's material claim—not selective praise
- Cited or visibly supporting URLs
- A screenshot or export where permitted
- A separate verifier note for each material fact
Keep three labels distinct throughout the report: observed describes the captured answer, inferred explains a plausible pattern, and recommended states the next action. An inference should never be presented as something the model or source explicitly said.
05 · AUDIT THE SUPPORT LAYER
Look beyond the answer to the public evidence a system can retrieve.
Answer visibility rarely reduces to one page or one schema block. Review the public sources that repeatedly shape buyer understanding and check whether they agree about the business.
Prioritize contradictions and missing facts that matter to a buying decision. A typo in an old low-value listing is not equal to conflicting hours, locations, service availability, licensing, or warranty language on first-party pages.
06 · INTERPRET WITHOUT OVERCLAIMING
Use counts only inside the frozen sample.
It is reasonable to report that a business appeared in 6 of 20 documented checks. It is not reasonable to turn that number into a universal market share or permanent “AI rank.” If you use a readiness score, publish its inputs, weights, limitations, and the evidence beneath it.
“The business appeared in 6 of 20 recorded checks run under the field conditions listed in the appendix.”
“The business owns 30% of AI search” or “will rank after these changes.”
Patterns matter more than a single surprising answer. Look for repeated absence by intent, recurring competitors, recurring sources, consistent fact errors, and the difference between branded and non-branded questions.
07 · BUILD THE ACTION QUEUE
Turn each gap into an owned, verifiable task.
An implementation queue should name the action, owner, dependency, effort, and proof of completion. “Improve AEO” is not a task. “Publish the client-approved emergency-service process on the service page and link it from the location page” is.
| Field | Question it answers |
|---|---|
| Evidence | What recorded observation or verified source created this task? |
| Action | What exact public change should be made? |
| Owner | Who approves, writes, implements, or verifies it? |
| Dependency | Which fact, credential, approval, or technical change must come first? |
| Effort | Is it a small correction, page revision, profile update, or outreach project? |
| Proof | What public URL, validation result, or approved record shows completion? |
08 · IMPLEMENT, WAIT, AND RERUN
Compare like with like after the public evidence changes.
Verify that the implementation is public, crawlable, internally linked, and—where relevant—discovered by search systems. Then rerun the same frozen prompt set under comparable conditions. Record new answers alongside the original sample; do not overwrite the baseline.
A rerun can show movement inside the sample. It still cannot prove causation from one change or guarantee that the pattern will persist. Report improvements, regressions, and unchanged results with the same discipline.
09 · ONE-PAGE FIELD CHECKLIST
The compact version your strategist can reuse.
- Approve the brief.Entity, market, services, comparison set, and publishable facts.
- Freeze the matrix.Balanced intents, exact wording, branded/non-branded mix, and run order.
- Record field conditions.Interface, visible model, date, locale, and account context.
- Capture complete observations.Mentions, competitors, material claims, and supporting URLs.
- Verify material facts.Use authoritative public sources; flag unresolved contradictions.
- Separate labels.Observed evidence, inference, and recommendation remain distinct.
- Prioritize implementable gaps.Owner, dependency, effort, and proof of completion for each task.
- State limitations.No permanent rank, guaranteed mention, traffic, lead, or revenue claim.
- Preserve the baseline.Rerun the same matrix only after verified public changes.
10 · SEE THE METHOD IN CONTEXT
Review a fictional worked example before using client data.
The Local Answer Lab sample uses fictional business data to show how the evidence log, limitations, source gaps, and implementation handoff fit together. It is an illustrative format—not a client result or performance claim.