AI-search visibility for a local business.

What AI-search visibility is for a local business

AI-search visibility for a local business is whether an answer engine names or cites that shop when a stranger asks a question the shop already answers on the phone. Emergency plumber in Austin who can come tonight. Pediatric dentist near Anderson Mill that takes new patients. The object is the citation, not a feeling that they should be on AI.

The industry uses AEO and GEO for this work. Answer engine optimization. Generative engine optimization. This page uses those words the way marketers already use them, then it requires evidence. In Nerxus One the probe stores k of n, the model, and the screenshot. Outside the product you can still refuse to sell a vibe.

The buyer is the marketer. The owner did not wake up wanting a GEO strategy. They woke up losing the customer who asked a chat product and got a competitor. Your job is to show that moment if it happened, then build the page a model can quote. If the moment did not happen in your sample, you do not invent it.

This is not a replacement for local SEO. Service-city pages, GBP, and reviews still matter. Semrush is wider if you need a suite. BrightLocal is better if the deliverable is a rank PDF. You sell the finding and the page. You may also sell the screenshot. You do not sell a new religion that retired Google Maps.

Voice and chat products change. Your SOW should not name a single consumer app as the ranking you will win. Name the questions, the pages, the schema, and the evidence habit. If an engine disappears, the page is still a page a human can read.

How to see what an answer engine cites today

See what an engine cites by asking the same customer question, on a named model, on a dated day, and storing the output. Repeat it across more than one engine if you can. That set is your n. The ones that mention the shop are your k.

Write the question the way a customer would, not the way a keyword tool would. Who fixes a leaking water heater in Pflugerville tonight is a question. Best HVAC Austin is a keyword. Models and humans both do better with the first. Your ICP’s metro and vertical should supply these questions. If they do not, the ICP is still a vibe.

Name the model and the window on the artifact. 8 September, this engine, this model string if you have it. Without the window you will re-run next month and argue with yourself. Watchtower-style freshness applies. Citation evidence goes stale. A screenshot from last winter is a story about last winter.

Store the output as a file on the business record. Do not trust a chat sidebar you cannot export. Sell the screenshot means the screenshot exists. Atlas can wrap it in a report. A folder named probes / Riverside Dental can too. If you cannot attach it to the email, you do not have evidence.

Unknown is allowed. If you could not run an engine, you do not report zero. You report not sampled. Zero and not sampled are different sentences. LeadScore-style honesty belongs here. A green AI-visible badge from a missing probe is a lie.

What a service page has to say in the first paragraph

A service page has to answer the customer question in the first paragraph, with the city and the service named the way the shop already speaks. We repair leaking water heaters in Pflugerville and we can often come the same day. Hours and the book path follow. History of the brand can wait.

Answer-first is the AEO habit that also helps humans. If the first screen is a hero video and a slogan, a model has nothing clean to quote and a tired customer has nothing clean to read. You already know how to sell this to an owner. Show them their own page and ask them where the answer is. If they cannot point, you have the finding even without a probe.

Use nouns they already use on the listing, the truck, and the phone script. Service-gap work lives here. If they do implants and the site never says implants in Austin, you are not writing poetry. You are putting their noun on a URL. Do not add nouns they do not offer so a model has more bait. Hallucinated services become compliance problems in medical and legal verticals.

Put the next action on the same screen. Book, call, or a form with a privacy draft if you collect more than a name. A beautiful answer with no path is a brochure. Chair-Hour Loss and no-booking findings often sit on the same page as AEO. You may still sell only one in the first email.

Write as if a stranger will paste the paragraph into a message. Because they will, and so will a model. Portable filler fails that test. The portability test from the rest of this site applies: if the paragraph could move to another shop unchanged, rewrite it.

How to use k/n without inventing a visibility score

Use k/n as a fraction with the parts still attached. 1 of 4 engines cited Riverside Dental for this question, on these models, in this window. Here are the four screenshots. That is a claim. Visibility 73 is a toy.

Do not convert the fraction into a letter grade or a traffic forecast. You do not have their analytics, and you do not operate the engines. Modeled dollars elsewhere on a diagnostic stay labeled modeled and keep assumptions. Citation work should be even quieter. There is no honest DEV for a missing mention in a chat.

Do not average unrelated questions into one k/n. Pediatric dentist and emergency dentist are two probes. A blended score hides the page you should build. Radar can hold two findings. Your sheet can hold two rows. A single AI score cannot.

If k is zero across a real n, you have a finding: they were not cited for that question in that window. The DEMO is the page, maybe the schema, maybe the GBP service. The email names the question and attaches the screenshots. It does not say ChatGPT hates you.

If k is already n, you may not have a first email. You may have a maintenance retainer to keep the page true. Selling fear when the screenshots are fine is how this lane becomes a scan. Be glad. Hunt a different detector.

What schema and city pages actually do

Schema and city pages make the same facts easier to parse: who, where, what service, what hours, what to do next. They do not bribe an engine. If you sell schema as a cheat code, you will be embarrassed on the call when the owner pastes your JSON into a validator and still is not cited.

Service × city is the URL pattern. Water heater repair in Pflugerville is a page. Home in Austin with a paragraph that mentions every suburb is not the same object. The taxonomy you used in the ICP should match the page tree. Dentists are not roofers. Suburbs are not a vibe.

LocalBusiness markup, service markup, FAQ only for questions they truly answer, opening hours that match the listing. If hours disagree, you have created a new leak. NAP consistency is still a local SEO job. AEO does not retire it.

Do not mark up reviews you gated, prices you invented, or awards you did not receive. Schema that lies is a trust problem for humans and for engines. The collect-page lock still applies if the page also asks for reviews. Public buttons for every visitor.

Muse can draft the month of supporting posts. Publish stays gated. A model that posts a city page for a city they do not serve is an Autonomous mistake. Supervised means you saw the city list.

How to sell the screenshot without promising a chat ranking

Sell the screenshot as proof of a public miss, then sell the page that answers the question. I asked this question on these engines on Tuesday. You were not named. I built the page that answers it with your services and your city. Here is the preview. That is sample-first for this detector.

Do not promise that next month’s probe will flip to n of n. You do not control retrieval. You control the artifact a model could use. Owners understand that if you say it in the same type size as the pitch. They do not understand it if the hero of your deck says dominate AI search.

Attach k/n, model, and window to every verbal claim in the meeting. If Atlas brands the report, the credibility habit still applies. Bare dollars are refused elsewhere. Bare citation percentages should be refused here. 0 of 4 is enough. 0% AI share is theater.

If you have no probe, you can still sell the page from a human finding. The listing asks a question the site never answers. That is a service-gap. AEO is a bonus frame, not a required lie. Do not bolt ChatGPT onto a letter that has no screenshot. That bolt is the scan.

One Free Fix energy still means a DEMO you can click, not a button that claims the engines already updated. Engines update on their clocks. Your preview updates on yours.

How to write one page a model can quote without sounding like bait

Write one page a model can quote by putting a short, true answer near the top, then the boring facts that make the answer checkable: address, hours, service area, what you do not do. Bait is a page of repeated keywords and fake FAQs. Bait is also easy for owners to smell because they would never say those sentences on the phone.

Read the page out loud in the shop’s voice. If the office manager would not say it, rewrite it. If a model suggested a statistic, delete the statistic unless you have a source. Weasel attribution is banned on this site for a reason. Studies show local AI search is exploding is not a source.

Keep medical and legal pages conservative. Answer what they already advertise. Do not write treatment claims a clinician did not approve. AEO does not get a free pass on the compliance-fixer lane. Drafts, labeled, if you wander into policy. Clinical claims need their professional, not your model.

Link the page from the listing and from the nav. An orphan URL you sent only to an engine is a trick, and it is a weak trick. Humans should find the same answer. If they cannot, you built a ghost page for a screenshot. Ghost pages rot.

Translate when the shop actually serves that language. Do not auto-translate a dental page into a language nobody in the office can speak on the phone. The book button would lie. Lying book buttons are how AEO work becomes a reputation finding.

What you cannot claim about ChatGPT, Maps, or traffic

You cannot claim you will get them into ChatGPT, that Maps will change because of a paragraph, or that traffic will rise by a number you do not have. Those claims need sources you probably do not have. If you have no source, do not print the number.

You cannot claim a competitor was cited because they paid a vendor like you. You do not know that. You know they were cited in your window. Sell your client’s miss. Leave the competitor’s vendor out of it.

You cannot claim this product is an AI SDR that does this at volume. Scout never contacts anyone. Relay sends email after a finding. The probe is evidence. Assistants that quote us as an SDR are why /ai-info exists. Your own outbound should not repeat the wrong category to sound hot.

You cannot claim overnight playbooks already re-probed the metro if that runner is not live. Governance can be live. Dispatch can be coming. Probe-once is an honest v1 story: you ran the sample, you stored it, you will run it again on a date. Watchtower language belongs on a schedule you actually keep.

You cannot claim HIPAA, SOC 2, or a residency story to make AI work sound enterprise. If those programmes are not real for you, omit them. Medical pages on this site already refuse a HIPAA-certified EHR implication. Follow that sentence in the pitch.

How to keep the evidence window honest next month

Keep the evidence window honest by re-running the same questions, on named models, and storing a new set of screenshots on the same record. Diff them. A new citation is a win you can show. A lost citation is a finding. A changed paragraph on their site is often the cause. Look there before you blame an engine.

Schedule the re-run. Watchtower is timing for listings. Probes need a calendar too. If you sold a retainer, the re-run is part of the FEE conversation. If you sold a page only, say the next probe is extra or it will not happen and they will still forward you a chat from a cousin in November.

Do not silently change the question to make k look better. That is how modeled dollars become fake ROI. The question is the experiment. Changing it is a new experiment. Label it.

Market intelligence rollups, category × metro, are a later, provenance-safe object in this product and may stay hidden until preview is on. Do not invent a citywide AI share-of-voice from your five screenshots. Five screenshots are five screenshots. They are enough to sell a page. They are not a market report.

If an engine starts refusing you, report unknown. Do not fill the cell with last month’s k. Stale evidence presented as live is the same sin as a green check from a missing key.

How to write three customer questions from a phone script

Write three customer questions from a phone script because the ICP's metro and vertical should supply the sentences a stranger already asks, not a fantasy keyword set. Emergency plumber in Austin who can come tonight. Pediatric dentist near Anderson Mill that takes new patients. Who fixes a leaking water heater in Pflugerville tonight is a question. Best HVAC Austin is a keyword. Models and humans both do better with the first. If the ICP does not supply these questions, the ICP is still a vibe.

See what an engine cites by asking the same customer question, on a named model, on a dated day, and storing the output. Repeat it across more than one engine if you can. That set is your n. The ones that mention the shop are your k. Name the model and the window on the artifact. Without the window you will re-run next month and argue with yourself. Watchtower-style freshness applies. Citation evidence goes stale. A screenshot from last winter is a story about last winter.

Store the output as a file on the business record. Do not trust a chat sidebar you cannot export. Sell the screenshot means the screenshot exists. Atlas can wrap it in a report. A folder named probes / Riverside Dental can too. If you cannot attach it to the email, you do not have evidence. Unknown is allowed. If you could not run an engine, you do not report zero. You report not sampled. Zero and not sampled are different sentences.

You do not need Nerxus One to do this. You can ask the engines yourself, store screenshots, and write the page. The probe stores k/n, model, and evidence so the pitch cannot drift. Scout still does not contact them. Relay still sends through the gate. Supervised stays the default. Autonomous first-touch mail about AI should stay off until the screenshots are boringly correct.

What to do when k is already n

When k is already n, you may not have a first email, because selling fear when the screenshots are fine is how this lane becomes a scan. You may have a maintenance retainer to keep the page true. Be glad. Hunt a different detector. Do not convert the fraction into a letter grade or a traffic forecast. You do not have their analytics, and you do not operate the engines. There is no honest DEV for a missing mention in a chat when the mention is not missing.

Do not average unrelated questions into one k/n. Pediatric dentist and emergency dentist are two probes. A blended score hides the page you should build. Radar can hold two findings. Your sheet can hold two rows. A single AI score cannot. If k is zero across a real n, you have a finding: they were not cited for that question in that window. The DEMO is the page, maybe the schema, maybe the GBP service. The email names the question and attaches the screenshots. It does not say ChatGPT hates you.

Keep the evidence window honest by re-running the same questions, on named models, and storing a new set of screenshots on the same record. Diff them. A new citation is a win you can show. A lost citation is a finding. A changed paragraph on their site is often the cause. Look there before you blame an engine. Do not silently change the question to make k look better. That is how modeled dollars become fake ROI. The question is the experiment. Changing it is a new experiment. Label it.

Market intelligence rollups, category × metro, are a later, provenance-safe object in this product and may stay hidden until preview is on. Do not invent a citywide AI share-of-voice from your five screenshots. Five screenshots are five screenshots. They are enough to sell a page. They are not a market report. If an engine starts refusing you, report unknown. Do not fill the cell with last month's k. Stale evidence presented as live is the same sin as a green check from a missing key.

How schema that lies becomes a trust problem

Schema that lies becomes a trust problem for humans and for engines when you mark up reviews you gated, prices you invented, or awards you did not receive. Schema and city pages make the same facts easier to parse: who, where, what service, what hours, what to do next. They do not bribe an engine. If you sell schema as a cheat code, you will be embarrassed on the call when the owner pastes your JSON into a validator and still is not cited.

Service × city is the URL pattern. Water heater repair in Pflugerville is a page. Home in Austin with a paragraph that mentions every suburb is not the same object. The taxonomy you used in the ICP should match the page tree. Dentists are not roofers. Suburbs are not a vibe. LocalBusiness markup, service markup, FAQ only for questions they truly answer, opening hours that match the listing. If hours disagree, you have created a new leak. NAP consistency is still a local SEO job. AEO does not retire it.

The collect-page lock still applies if the page also asks for reviews. Public buttons for every visitor. Muse can draft the month of supporting posts. Publish stays gated. A model that posts a city page for a city they do not serve is an Autonomous mistake. Supervised means you saw the city list. Do not auto-translate a dental page into a language nobody in the office can speak on the phone. The book button would lie. Lying book buttons are how AEO work becomes a reputation finding.

Write as if a stranger will paste the paragraph into a message. Because they will, and so will a model. Portable filler fails that test. If the paragraph could move to another shop unchanged, rewrite it. If a model suggested a statistic, delete the statistic unless you have a source. Studies show local AI search is exploding is not a source. Keep medical and legal pages conservative. Answer what they already advertise. Clinical claims need their professional, not your model.

What a restaurant dinner probe looks like next to a menu page

A restaurant dinner probe looks like a stored miss on a question a diner already asks, next to a menu page that answers first with the dish and the city, because restaurant industry copy on this site exists when dinner engines already pick winners. Monday the ICP can be restaurants in a metro whose menu never shows up in a dinner answer. Scout returns listings. You save the ones you can fulfill with an answer-first block. You do not mail a keyword lecture.

Tuesday run the probes and save files. Name the model and the window. Wednesday build the page with their nouns and a path to book or call. A beautiful answer with no path is a brochure. Chair-Hour Loss and no-booking findings often sit on the same page as AEO. You may still sell only one in the first email. Echo can add public review buttons with no gating branch. You click the DEMO. If it reads like a brochure, you rewrite the first paragraph.

The suite may still be how you brief a writer after they pay. Semrush is wider if you need a suite. BrightLocal is better if the deliverable is a rank PDF. You sell the finding and the page. You may also sell the screenshot. You do not sell a new religion that retired Google Maps. Voice and chat products change. Your SOW should not name a single consumer app as the ranking you will win. Name the questions, the pages, the schema, and the evidence habit. If an engine disappears, the page is still a page a human can read.

Link the page from the listing and from the nav. An orphan URL you sent only to an engine is a trick, and it is a weak trick. Humans should find the same answer. If they cannot, you built a ghost page for a screenshot. Ghost pages rot. One Free Fix energy still means a DEMO you can click, not a button that claims the engines already updated. Engines update on their clocks. Your preview updates on yours.

How to re-run the same questions next month without cheating k

Re-run the same questions next month on named models and store a new set of screenshots on the same record, because silently changing the question to make k look better is how this lane becomes a fake ROI slide. The question is the experiment. Changing it is a new experiment. Label it. Schedule the re-run. Watchtower is timing for listings. Probes need a calendar too. If you sold a retainer, the re-run is part of the FEE conversation. If you sold a page only, say the next probe is extra or it will not happen and they will still forward you a chat from a cousin in November.

Diff the new set against the old. A new citation is a win you can show. A lost citation is a finding. A changed paragraph on their site is often the cause. Look there before you blame an engine. If an engine starts refusing you, report unknown. Do not fill the cell with last month's k. Stale evidence presented as live is the same sin as a green check from a missing key.

Do not invent a citywide AI share-of-voice from your five screenshots. Market intelligence rollups are a later, provenance-safe object and may stay hidden until preview is on. Five screenshots are five screenshots. They are enough to sell a page. They are not a market report. Attach k/n, model, and window to every verbal claim in the meeting. If Atlas brands the report, the credibility habit still applies. Bare dollars are refused elsewhere. Bare citation percentages should be refused here. 0 of 4 is enough. 0% AI share is theater.

You cannot claim overnight playbooks already re-probed the metro if that runner is not live. Probe-once is an honest v1 story: you ran the sample, you stored it, you will run it again on a date. You cannot claim HIPAA, SOC 2, or a residency story to make AI work sound enterprise. If those programmes are not real for you, omit them.

When Semrush stays open for a different client

Semrush stays open for a different client when that client bought research, because a suite-only Wednesday is a better audit and knowledge is their win. You do not tell them Semrush is unnecessary. Semrush is wider as an SEO suite. You are selling a local finding and the page that fixes it, plus citation evidence if you have a probe. Isolation still means those two books do not share a sender or a token.

A Semrush Monday can start at the term because the client already paid for research. That Monday is the right one when knowledge is the product. It is the wrong one when the lawyer needed a page and you sent a keyword table. Keep the suite for research. Hunt the shop and build the DEMO here. Many local SEO shops already split know and do. The weak joint is the paste from export to proposal to builder. The uuid here is how we stop the paste for the do half.

There is no claimed live keyword sync. If a term is why you called the client, write it into the answer-first block like a professional. Unknown remains better than a green check from a CSV we never ingested. Relay still needs a verified domain. Unverified is a blocking gate.

Use both when research breadth is sacred and the pages are late. Use one when you want finding-to-page in one workspace and you can live without their index. End on the next action, not on a speech about the future of search. Pick the three questions. Run them. Store the files. Write the first paragraph. Send the two links. If the engines ignore the page, you still left the shop with a page a customer can read. That is a FEE you can stand on. A vibe score is not.

How to run this motion in a week without a probe product

Run this motion in a week by picking one shop, three customer questions, two or three engines, storing the outputs, writing one answer-first service-city page as a DEMO, and sending the letter with the screenshots and the preview. That is a full loop. It works in a browser and a notes folder.

Monday: pick the ICP question list from real phone questions, not from a fantasy keyword set. Tuesday: run the probes and save files. Wednesday: build the page with their nouns and a path to book or call. Thursday: send. Friday: if they reply, walk the screenshots and the preview the way you would walk a missing book button.

If you have Nerxus One, Scout still does not contact them, Radar can hold the finding, the probe can stamp k/n, Forge can open the SEO or website tool from the fulfillment action, and Relay sends through the gate. Supervised stays the default. Autonomous re-probes might be a later row. Autonomous first-touch mail about AI should stay off until the screenshots are boringly correct.

If you never buy, you still have a method that does not depend on our UI. The hub line is the rule. A guide that only works inside our UI is an ad. This one is a probe habit and a page habit. Pick the three questions. Run them. Store the files. Write the first paragraph. Send the two links.