Quick answer
To check whether AI recommends your business, open a fresh session in ChatGPT, Perplexity or Gemini with memory and personalisation switched off, then ask for a recommendation in your category and city without naming your business. Record whether you appear, who does, and the reasons given.
Repeat across at least three assistants, because each pulls from different sources. Then ask about your business by name and check whether the facts are right. Fifteen minutes, no cost, and the result tells you more than most paid audits.
Something quiet happened over the last two years. A meaningful share of your customers stopped starting at Google.
They open ChatGPT instead. They ask Perplexity. They read the AI summary at the top of the results page and never scroll. And in every one of those cases, a machine picks two or three businesses to name — out of the dozens that could have been named.
So here is the uncomfortable question. When that happens in your category, in your city, does your name come up?
Most owners have never checked. Not because they do not care, but because nobody told them it was checkable. It is. And it takes about a quarter of an hour.
Why this is worth fifteen minutes
Two numbers explain the whole opportunity. One is about customers. The other is about competitors.
The gap between how people search and how businesses market
United States, 2026
Sources: GrowthPro AI (2026) for local business AI strategy adoption and consumer usage; Similarweb (2025) for zero-click share, up from 56% the prior year; Acquia/Frase (2026) for the belief-versus-implementation gap. Bar widths are the stated percentages.
Look at the first bar and the last one together. Nearly nine in ten local businesses have done nothing. Only one in five organisations of any size has started. Meanwhile a third of consumers are already using AI to find local services.
That is not a trend you need to get ahead of eventually. That is a gap sitting open right now.
And the traffic that does arrive from these systems behaves unusually well. Visitors referred from ChatGPT convert at around 15.9%, and from Perplexity around 10.5%, against roughly 1.76% for ordinary organic search. The reason is simple: they arrive already holding a recommendation. Somebody they trust — or something they trust — already vouched for you.
The uncomfortable version
In a traditional search result there are ten links, and being seventh still puts you on the page. In an AI answer there are two or three names. Being fourth is the same as being invisible. That compression is the whole reason this matters more than a ranking drop.
Setting up a clean test
Before you run anything, remove the ways this test can lie to you. There are three.
- Turn off memory and personalisation. If you have discussed your own business with an assistant before, it may name you out of context rather than merit. In ChatGPT this means disabling memory and using a temporary chat. In others, log out or use a private window.
- Use a fresh session for every prompt. Assistants carry context forward. Ask six questions in one thread and answers two through six are contaminated by answer one.
- Do not name your business first. The first five prompts must be category questions. The moment you say your name, you have told the assistant the answer.
Then pick your inputs. You need your service as a customer would say it — "plumber", not "hydronic systems specialist" — and your city or neighbourhood.
Have a notepad open. You are recording three things per prompt: whether you appeared, who did, and what reason the assistant gave.
That third column is the one people skip, and it is the most valuable. The reasoning tells you exactly what the system thinks matters in your category.
The six prompts
Run these exactly as written, swapping in your service and city. Copy each one, run it in a fresh session, record the result, move on.
Prompt 1 · The basic recommendation
"Who is the best [your service] in [your city]? Give me three options and explain why you picked each one."
What to look for: whether you appear at all, and whether the reasoning mentions reviews, years in business, specialisation or something else entirely.
Prompt 2 · The buying-intent version
"I need a [your service] in [your city] this week. Who should I call, and what should I ask them before hiring?"
What to look for: urgency changes the answer. Some businesses surface for research questions but vanish for ready-to-buy ones.
Prompt 3 · The specialisation test
"Which [your service] companies in [your city] specialise in [your most profitable service]? Be specific about who does what."
What to look for: this is where small businesses win. General category answers favour big names. Specialisation answers favour whoever documented their niche clearly.
Prompt 4 · The comparison
"Compare the top [your service] providers in [your city]. Make a table with pricing approach, specialisms, and what customers say about each."
What to look for: comparison content is the format most cited by AI. If a competitor dominates this answer, look at what they published to earn it.
Prompt 5 · The neighbourhood test
"I live in [your neighbourhood or nearest suburb]. Which local [your service] businesses actually serve that area?"
What to look for: service-area accuracy. Many businesses are visible city-wide but invisible for the specific suburbs they actually work in.
Prompt 6 · The brand check (do this one last)
"Tell me about [your business name] in [your city]. What do they do, who do they serve, and what do customers say?"
What to look for: accuracy. Wrong hours, an old address, services you dropped years ago, or a flat "I don't have information about that business" — each points to a different problem.
Do this bit properly
Write the answers down. Do not just read them and form an impression. In ninety days you will re-run the same six prompts, and a written record is the only way to tell whether anything you changed actually worked.
How to read your results
You now have six answers across three or four assistants. Sort what you found into one of four buckets.
| What you saw | What it means | Where to start |
|---|---|---|
| Named in most prompts, facts correct | You are visible and your entity is clean | Protect it — keep reviews and content fresh |
| Named sometimes, mostly for general prompts | Broad presence, weak specialisation signals | Publish depth on your most profitable service |
| Named but details wrong or outdated | Entity inconsistency across the web | Fix your facts everywhere before anything else |
| Never named, competitors always are | You are effectively invisible to these systems | Start with crawler access and review signals |
| Assistant says it has no information | No usable entity has formed at all | Google Business Profile, citations, then content |
| Named in Perplexity but not ChatGPT | Live web presence fine, training-data presence thin | Third-party mentions — press, directories, industry sites |
That last row deserves a note, because it confuses people. Perplexity leans heavily on live retrieval, so it reflects the current web quickly. Assistants that lean more on training data reflect a slower, older picture built from how widely you are referenced across the internet.
So being visible in one and not the other is diagnostic, not contradictory. It tells you your website is fine but the wider web barely mentions you.
The 12-point technical scorecard
The prompts tell you whether you are visible. This checklist tells you why. Tick everything that is genuinely true — be honest, since nobody sees this but you.
AI Visibility Scorecard
Twelve checks. Tick what's true. Your score updates as you go.
Access — can AI read you at all?
Entity — does the web agree on who you are?
Proof — is there reason to recommend you?
Format — is your content quotable?
Your AI visibility score
Tick the boxes above to score
Every item is weighted by how much it actually affects whether an AI names your business.
Weightings reflect our own experience across Denver client work plus the published research cited throughout this article. Treat the score as a priority guide, not a precise measurement.
What AI actually looks at
Strip away the mystique and these systems are doing something fairly understandable. They need to name a business, and they need a defensible reason. So they look for evidence.
Consistency. If twelve sources agree your name, address and phone, a confident entity forms. If four sources disagree, the system hedges — and hedging usually means naming somebody else.
Corroboration. Your own website claiming you are the best is worth almost nothing. Independent sources saying anything factual about you is worth a great deal. This is why a mention in a local paper or a trade association listing outweighs another page of marketing copy.
Extractability. Content that answers a question plainly can be lifted into an answer. Content that builds atmosphere for four paragraphs before saying anything cannot. This is a genuine format difference, not a writing-quality judgement.
Recency. Roughly half of the content cited in AI answers is under thirteen weeks old. A site untouched for two years is not just stale to humans.
Specificity. Research from Princeton and the Allen Institute for AI found that adding statistics to a page increased its visibility in generative answers by about 26%, and adding quotations by about 28%, with targeted optimisation lifting visibility by up to 40% overall. Vague pages lose to precise ones.
The practical translation
Say specific, checkable things. "We've served Denver since 2012" beats "we have years of experience." "Most campaigns show movement in 3–6 months" beats "results vary." Precision is not just better writing — it measurably increases the chance of being quoted.
The llms.txt myth
Now for something you may have been sold, or may be about to be.
Over the last year a file called llms.txt has been marketed hard as an AI visibility essential. It is a Markdown file placed at your site root, proposed by Jeremy Howard of Answer.AI in September 2024, intended to summarise your site for language models.
Plenty of agencies now bill for creating and maintaining one. So it is worth knowing what Google says about it.
Google published its first official AI optimization guide on 15 May 2026, then updated it on 15 June. It lives in the Generative AI fundamentals section of Google Search Central, and the relevant line is not ambiguous:
Google Search Central, AI optimization guidance, 15 June 2026
"You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them."
And on the file specifically: "It's completely fine if you decide to create and maintain LLMS.txt files (or other similar files) for other services or systems that use these files. Doing so won't harm (nor help) your visibility or rankings in Google Search, as Google Search ignores them."
Google's position has been consistent, and was reported at the time by Search Engine Land. Gary Illyes said at Search Central Live in July 2025 that Google does not support llms.txt and had no plans to. John Mueller compared the idea to the old meta keywords tag — a signal spammers stuffed and search engines eventually ignored entirely.
Independent measurement backs this up. One test domain logged 62,100 AI bot requests over ninety days. Just 84 of them touched the llms.txt file. That is around 0.1% of AI crawler traffic.
To be fair, the picture is not entirely one-sided. Some systems outside Google do read the file, and Google confusingly added an llms.txt check to its Lighthouse audit tool. If you run documentation-heavy software where AI coding assistants parse your docs, there is a genuine if narrow use case.
But for a local service business? It does essentially nothing, and it should never appear as a headline deliverable on an invoice.
If someone is selling you this
Ask them one question: "Which specific AI systems read this file, and what evidence do you have that it changed our visibility?" A good answer names systems and shows before-and-after data. A vague answer about future-proofing means you are paying for a text file.
The same test applies to any AI-search deliverable. The work that moves the needle is unglamorous: accurate facts, real authority, clear content, strong reviews. That is harder to sell than a magic file, which is precisely why the magic file gets sold.
Fixes, in priority order
Work down this list. Do not skip ahead, because each step makes the next one more effective.
1. Unblock the crawlers (today, 10 minutes)
Open yoursite.com/robots.txt in a browser. Look for Disallow rules affecting GPTBot, ClaudeBot, PerplexityBot, Google-Extended, OAI-SearchBot or a blanket block on all agents.
If you find one, decide deliberately. Publishers who sell access to their own content have a reason to block. A plumber does not. If you want AI to recommend you, it has to be able to read you.
2. Make the web agree about you (this week)
Fix your Google Business Profile first — hours, categories, services, description, photos. Then your own site. Then the directories.
The standard is exact match. "Suite 660" on one and "#660" on another is a small inconsistency to you and a genuine ambiguity to a machine trying to decide whether those are the same business.
3. Restart your review flow (this month)
Reviews are among the heaviest inputs to these recommendations, and recency counts as much as volume. If your newest review is four months old, that is the highest-return fix available to you.
We covered the whole method in how to get more Google reviews without being annoying — including the two rules that carry real penalties. The fastest mechanism for collecting them is covered in our NFC marketing guide.
4. Write the facts down plainly (this month)
Somewhere on your site, in ordinary readable text, state: what you do, where you do it, when you started, who you serve, what it typically costs, and how to reach you.
Most business websites never say these things directly. They imply them through imagery and adjectives. A machine cannot extract an implication.
5. Publish something worth quoting (next quarter)
Comparison content earns the largest share of AI citations of any format. So a page that honestly compares options in your category — including when someone should not choose you — is unusually likely to be cited.
Add real numbers. Add named sources. Add specifics you can defend. That is what the Princeton research measured, and it is the cheapest lever most businesses never pull.
6. Then check the destination
An AI recommendation sends someone to your site with unusually high intent. If they land on something slow or confusing, you have wasted the hardest-won click in marketing. Our post on why your website gets traffic but no calls covers the symptoms, and conversion optimization covers the repair.
The order matters
Publishing brilliant content while your robots.txt blocks AI crawlers is wasted effort. Fixing crawler access while your business facts contradict each other across six directories is also wasted effort. Access, then identity, then proof, then content. In that order.
How often to re-run it
Every ninety days. Same six prompts, same assistants, same notepad.
Ninety days is long enough for changes to be crawled and reflected, and short enough that a slide shows up before it costs you real money. Any more often and you are measuring noise — these answers vary between sessions and regions, so week-to-week movement means very little.
Track three things over time: how many of the six prompts named you, how many assistants named you at least once, and whether the facts returned were accurate. Three numbers, four times a year. That is a complete AI visibility programme for a local business.
One caution on expectations. Changes take weeks, not days. Live-retrieval engines like Perplexity reflect updates fastest. Assistants leaning on training data move far more slowly, because they are waiting on the wider web to catch up with you.
What this audit cannot tell you
Worth being straight about the limits.
These answers vary. Between sessions, between users, between regions, and between model versions. So no single run is definitive, and anyone selling you a precise "AI visibility percentage" is selling precision that does not exist yet.
What the audit does reliably is separate three states: broadly visible, occasionally visible, and absent. Absence across six prompts and three assistants is not noise. That is a finding, and it is actionable today.
It also cannot tell you how much business you are losing. Nobody can measure the customer who asked ChatGPT, got three names that were not yours, and called one of them. That loss is invisible by design — which is exactly why it goes unnoticed for so long.
Want us to run the full version?
We'll run the prompts across every major assistant, audit the technical side properly, benchmark you against your actual map pack competitors, and give you a prioritised fix list. Free, and we'll tell you if the honest answer is that you're already fine.
Frequently asked questions
Open a fresh session in ChatGPT, Perplexity or Gemini with personalisation and memory turned off, then ask for a recommendation in your category and city without naming your business. Record whether you appear, which competitors appear, and what reasons the assistant gives.
Repeat across at least three assistants, because each draws on different sources. Finally, ask about your business by name and check the facts. The whole process takes about fifteen minutes and costs nothing.
Usually one of five reasons. Your site blocks AI crawlers in robots.txt. Your business facts are inconsistent across the web, so no clear entity forms. You have too few or too stale reviews to read as credible. Your content is marketing copy rather than extractable answers. Or you are simply absent from the third-party sources these systems lean on — directories, local press, industry listings.
The audit in this article isolates which of the five applies to you.
Not for Google. On 15 June 2026 Google updated its official AI optimization guidance to state that you do not need machine readable files, AI text files, markup or Markdown to appear in Google Search including its generative AI features, because Google Search does not use them. Google added that maintaining such files is fine but will neither help nor harm visibility.
Independent crawler logging found only 84 of 62,100 AI bot requests over 90 days touched an llms.txt file — roughly 0.1%. Some non-Google systems do read it, so it is not useless, but it should never be sold as a core AI visibility deliverable.
Every ninety days suits most local businesses. AI answers shift as models update, as source content changes and as competitors publish.
Ninety days is long enough for your changes to be crawled and reflected, and short enough that you spot a slide before it costs you meaningful business. Keep the same six prompts each time so results stay comparable.
It overlaps heavily but is not identical. Traditional SEO optimises for a ranked list of links. AI visibility optimises for being named inside a generated answer, where there is no list and usually only two or three businesses get mentioned.
The underlying work is largely the same — accurate entity data, genuine authority, structured content, strong reviews. What differs is content format and the emphasis on being quotable and factually extractable rather than merely relevant.
Yes. Review signals are among the strongest inputs AI systems use when deciding which local businesses to name. Rating, volume, recency and the actual language customers use in review text all feed that judgement.
A business with weak or stale review signals can be excluded from an AI recommendation before a search results page is ever generated.
For most local service businesses, no. Blocking removes any chance of being cited in AI answers, which is the opposite of what a business trying to attract customers wants.
Blocking makes sense mainly for publishers whose business model depends on selling access to their own content. If you want AI assistants to recommend you, they must be able to read you — so check robots.txt for rules affecting GPTBot, ClaudeBot, PerplexityBot, Google-Extended and similar agents.
Expect weeks rather than days, and treat ninety days as a realistic first checkpoint. AI systems have to crawl the change, and many answers draw on third-party sources such as directories and press that update on their own schedule.
Live-retrieval engines such as Perplexity tend to reflect changes faster than assistants relying more heavily on training data.
Content that answers a question directly, near the top, in language that can be lifted without editing. Comparison content performs particularly well, accounting for the largest share of AI citations of any format.
Research from Princeton and the Allen Institute for AI found adding statistics increased a source's visibility in generative answers by about 26% and adding quotations by about 28%, with targeted optimisation lifting visibility up to 40% overall. Freshness matters too — roughly half of AI-cited content is under thirteen weeks old.
Not in the organic answer itself. Some AI surfaces carry advertising, but recommendations inside a generated answer are earned rather than bought.
That is genuinely good news for smaller businesses. A well-documented local operator with strong reviews and clear facts can be named alongside far larger competitors who cannot simply outspend their way into the answer.
Wrong details almost always trace back to inconsistent information across the web rather than to the AI itself. Correct your Google Business Profile first, then your own website, then the major directories and industry listings.
Make sure name, address, phone, hours and services match exactly everywhere, then wait for a re-crawl. Assistants converge on whatever the web agrees on, so the fix is to make the web agree.
It is directional rather than precise, and that is enough. AI answers vary between sessions, users and regions, so no single run is definitive.
Running six prompts across three or more assistants and recording the pattern gives a reliable read on whether you are broadly visible, occasionally visible, or absent. Absence across every prompt and every engine is unambiguous and worth acting on immediately.