You get recommended by an AI assistant the same way you get recommended by a neighbor: by being the business whose public answers match the question that was actually asked. That is not a trick you buy. It is a small workflow you own, and it runs on one person spending under an hour a week. Pick that person. Have them ask ChatGPT, Claude, Gemini, and Perplexity what those assistants currently say about your business, using the words a customer would use, on a fixed day every week. Then fix the one gap that surfaced. The assistants read your listings, your reviews, and your service pages, so that is where every fix lands. The rest of this piece is how to run that loop so it does not quietly die in week three.
The position: the gap is almost never ranking, it is that nobody wrote the answer down
Here is the claim, and it is falsifiable: when an assistant fails to recommend you, the cause is usually a missing sentence in public text, not a missing budget. Test it yourself in ten minutes. Ask an assistant the question your best customer asked you last month, phrased the way they phrased it, and then search your own website and your own listings for the words in that question. Most of the time you will find that the phrase appears nowhere you control. You call it "emergency HVAC." Your customer typed "furnace died overnight, house is 48 degrees, who comes today." Those are not the same string, and nothing in your public text bridges them.
The mechanism is boring, which is why it is reliable. An assistant answering a local recommendation question is assembling a response out of text it can retrieve: search results, map and business listings, review bodies, directory entries, and your own pages. It is not consulting a secret ranking API you can petition. It reads what is there. If the vocabulary of the question appears in your public text, attached to your name and your town, you are a candidate. If it appears only in your head, or only in the quotes you email to customers, you are invisible to that question no matter how good you are at the work.
That mechanism has a pleasant consequence: this is fixable by writing, and writing is something a small business already knows how to do. It also has an unpleasant consequence. Nobody can do it for you once and be finished, because the questions change, your services change, and the assistants change what they surface. Which is exactly why the answer is a cadence and an owner, not a project.
Name the workflow, name the owner, pick the day
Give this a name and a single owner or it will not survive the first busy week. Call it the Friday Ask. One person runs it, same day, same time, thirty to forty minutes. Not the marketing agency, not "whoever gets to it," not a committee. If the owner is you, it is you every week. Shared ownership of a weekly habit reliably becomes nobody's ownership by the end of the first quarter, and this workflow fails silently: no alarm goes off when you stop checking.
The owner keeps one log, and the log lives somewhere other than their memory. A shared doc or sheet named assistant-check-log is enough. Five columns: date, prompt asked, which assistant, what it said, gap found. One row per prompt per week. The log is the entire reason this compounds instead of resetting, because it turns "I think ChatGPT mentioned us a while back" into a record you can read across eight weeks and see a pattern in. If you have already written your processes down so an AI tool can follow them, this is the same discipline pointed outward.
The five prompts to run every week
Run the same five prompts every week, changing only the details in brackets, because repetition is what makes drift visible. Variety feels productive and tells you nothing.
1. "Best [service] in [your town]." The category question. Are you in the set at all. 2. "Who should I call for [problem described as a symptom, not a service name] near [neighborhood]." The symptom question, and the one most businesses fail. 3. "Is [your business name] any good?" The reputation question. What does it cite, and from where. 4. "Does [your business name] do [a service you actually offer]?" The scope question. 5. "Compare [your business] and [a real competitor] for [service]." The head to head. This one is uncomfortable and it is the most useful.
Run each in a fresh chat with no prior context, because an assistant that has been talking to you about your own business for twenty minutes will flatter you. Run them across at least two assistants. Copy the answers into the log verbatim, including the parts that are wrong, especially the parts that are wrong.
The five gaps you will find, and what each one actually means
Every answer you get will fall into one of five buckets, and the bucket tells you which surface to fix.
**Silence.** You are not mentioned, and businesses you consider peers are. This means your name is not attached to that vocabulary anywhere retrievable. The fix is a service page that uses the customer's words in its heading and first paragraph, plus the same words in your business listing description.
**Wrong facts.** You are mentioned, but the hours are stale, the address is a suite you moved out of, the phone number is the old one, or it lists a service you stopped offering two years ago. This is a listings problem, and it is the highest damage per minute of fix. Wrong hours send a ready buyer to a locked door. Go straight to the source records: Google Business Profile, Apple Business Connect, Bing Places, your industry directories, and any old profile you forgot you created.
**Right business, wrong scope.** The assistant knows you and describes you as doing one third of what you do. You are "a residential plumber" when half your revenue is commercial backflow testing. The fix is a dedicated page per real service line, not a bullet buried in a list on your homepage. One page, one service, question-shaped heading, plain answer in the first two sentences, the town named.
**Named but unsupported.** You appear, but the assistant hedges: it can describe what you claim about yourself and nothing else. This is a review body problem, covered below.
**Competitor substitution.** A competitor is recommended for a job that is squarely yours. Read their public text before you feel anything about it. Usually they have simply written the sentence you have not.
Fix exactly one gap per week
One gap per week. Not five. This is the rule that people break, and breaking it is how the workflow ends.
The reasoning is mechanical, not motivational. A week is long enough for a listing edit to propagate and for a new page to be crawled, so a one gap cadence gives you a clean read on whether the fix worked: you changed one thing, and next Friday you ask the same five prompts and see whether that one thing moved. Change five things at once and you have learned nothing about any of them, and you have also just spent a Saturday. Suppose the fix takes twenty minutes and the check takes thirty: call it under an hour a week, every week, which is a load a real business can actually carry through a busy season. Calibrate that estimate for your own deployment after four weeks of real logs.
Write the fix into the log next to the gap, with the date. Next Friday, the first thing the owner does is reread last week's row.
Reviews are retrievable text, so stop asking for stars
Reviews matter here because they are text an assistant can read and quote, which means a five star review with no words in it is nearly worthless to this workflow. "Great service, highly recommend" supports no specific claim about you. "They replaced our sump pump in a Clintonville basement the same day I called, on a Sunday" supports several: the service, the neighborhood, the response time, the weekend availability. When an assistant is asked who handles a flooded basement on a weekend, the second review is evidence and the first is noise.
So change the ask. When you request a review, ask the customer to mention what the job was and where they are, in their own words. That sentence is not a script you write for them, and you should never write it for them, because a wall of identically phrased reviews reads as manufactured to humans and to the assistants both. You are simply asking them to be specific instead of generic. Ask at the moment the work is visibly done, which is when specificity is easiest for them and the goodwill is highest.
Watch for the reverse signal too. If your recent review bodies all describe one service line, assistants will describe you as a one service business, because that is what the evidence says. That is a gap worth logging.
The 30/60/90 shape
Thirty days: the habit exists and the log has four weeks of rows. You have fixed four gaps, almost certainly including at least one stale fact that was costing you calls. You are not judging results yet. You are judging whether the Friday Ask happened four times, because a workflow that does not run cannot work.
Sixty days: the log starts paying. Eight weeks of the same five prompts shows you which questions you reliably win, which you reliably lose, and which flip back and forth. The flippers are the interesting ones. They mean you are a marginal candidate for that question, and one good service page usually settles it.
Ninety days: you change the prompts. Not the cadence, the prompts. Pull the actual questions customers asked you this quarter, by phone, by email, by text, and replace two of your five with the two most common phrasings you did not already have covered. This is the step that keeps the workflow from calcifying into a ritual that measures nothing. Your customers' vocabulary moves. Your five prompts should follow it.
What to measure, and what not to
Measure three things, all of which come free out of the log. First, coverage: of your five prompts, how many name your business this week. Second, accuracy: of the facts the assistants stated about you, how many were correct. Third, weeks since a wrong fact appeared, which is the number you want getting larger. That is the whole dashboard.
Do not measure a rank position, because there is no stable rank to measure. The same prompt phrased two ways, asked twice, can produce different sets of businesses, and building a number out of that gives you a metric that moves for reasons unrelated to anything you did. Do not buy a tool that promises a single AI visibility score until you have run the log manually for a quarter, because you will not be able to tell whether the score is meaningful and you will have outsourced the one part of this that has to stay in house: reading what the assistants actually say about you, in full sentences, with your own eyes. The same discipline applies here as when you measure time saved on any single AI workflow: count the thing you can verify, not the thing that is easiest to put on a slide.
What kills this workflow
Three things kill it, and they are all predictable. It becomes everybody's job, so it becomes nobody's. The log lives in the owner's head, so when that person takes a week off the thread breaks and never gets picked up. Or the owner finds six gaps in week one, tries to fix all six, spends a weekend on it, and never opens the doc again. Name the owner, keep the log in writing where a second person can find it, fix one thing.
The underlying idea is the same one behind running a single AI workflow well rather than adopting ten at once: a small owned loop that actually runs beats an ambitious program that does not. Getting recommended by an assistant is not a channel you purchase and monitor. It is the visible output of whether your public text answers the questions your customers are really asking, checked weekly by a person whose name you can say out loud.
Start this Friday. Five prompts, one log, one gap.
Related
- smb-ai-strategy-one-workflow - write-down-your-process-so-ai-can-follow-it - measuring-time-saved-on-one-ai-workflow