A customer asks ChatGPT, Perplexity, Gemini, or an AI-powered Google result which supplier to choose. Your business is mentioned—or it is not. These systems are often called large language models, or LLMs. In this guide, “AI answer” means the customer-facing response they produce.
That sounds like a ranking problem. Vendors will happily sell you a clean number: “You rank number four in AI.” The number is comforting, but it can be fiction.
AI answers are not a fixed list of ten blue links. The response can change when the prompt changes by one word. It can change by location, language, model, search mode, logged-in context, and the sources available at that moment. Ask twice and you may get different recommendations. A platform may mention your brand without linking it, cite your article without naming the company prominently, or send a highly qualified visitor from a prompt your tracker never tested.
You can still measure visibility. You just need to treat it like a customer-research sample, not pretend there is one universal rank. You can start with a spreadsheet; a paid tracking platform is optional.
This guide gives a practical system for a small business:
- a stable panel of real customer prompts;
- repeatable test conditions;
- saved answers and citations;
- referral and conversion tracking;
- clear confidence labels;
- a monthly improvement routine.
Measure a sample, not a mythical rank
Connect buyer questions to business evidence.
- 1
Real customer questions
Sales, support, reviews, site search, and Search Console
- 2
Fixed prompt panel
Document the wording, market, and intent
- 3
Repeated observations
Run across selected answer surfaces on a cadence
- 4
Mentions + citations + accuracy
Record presence, source, position, and correctness
- 5
Business outcome
Referral visits, branded demand, qualified conversations
A spreadsheet is enough to start. Consistency matters more than dashboard polish.
What “LLM visibility” actually means
LLM visibility is a collection of observable signals. Keep them separate.
1. Mention visibility
Does the answer name your company, product, founder, or proprietary research?
Example: “For a small Berlin bakery, options include LocalSupplier, YourBrand, and…”
A mention can build awareness even without a link. It can also be inaccurate or negative, so count the context.
2. Recommendation visibility
Does the system present you as a reasonable choice for the user’s actual need?
“YourBrand exists” is weaker than “YourBrand is a good fit for a five-person company that needs X.” Recommendation fit matters more than raw mention frequency.
3. Citation visibility
Does the answer cite or link to a page on your website?
ChatGPT Search provides linked web sources in search-backed answers. Perplexity and Google’s AI search features also surface supporting links. A citation is observable, but it is not an endorsement; your page may simply support one factual sentence.
4. Answer accuracy
When the brand appears, are the price, location, product, availability, and positioning correct?
An incorrect recommendation can be worse than invisibility. Track harmful errors separately and fix their likely source.
5. Referral traffic
Did someone click from an AI experience to your website? OpenAI’s publisher guidance says ChatGPT search referral links include utm_source=chatgpt.com, which analytics tools can track. Referrals provide stronger evidence than a test prompt, but only when a clickable visit occurs and tracking survives browser and privacy controls.
6. Business outcome
Did the visit become a call, quote request, store visit, trial, or sale? This is the scoreboard that matters. Ten AI referrals that produce two qualified enquiries can be more useful than 10,000 unqualified impressions elsewhere.
Why a single “AI rank” is misleading
Imagine a family-owned commercial cleaning company in Manchester. It tests the prompt:
What is the best commercial cleaning company in Manchester?
The company appears third in one answer. Is its rank three? Not really.
Another buyer asks:
Who can clean a 20-person medical office in South Manchester every evening and use fragrance-free products?
That answer is a different market. It may use different searches, sources, and selection criteria. It is also closer to a sale.
The first prompt is broad and easy to track. The second is specific and valuable. A dashboard that averages both into “rank 3.7” hides the important difference.
Report the observations instead:
- mentioned in 9 of 30 tested answers: 30% mention rate;
- recommended as a fit in 4: 13% recommendation rate;
- website cited in 6: 20% citation rate;
- facts correct in 8 of 9 mentions: 89% observed accuracy;
- 17 measurable referral sessions and 3 enquiries this month;
- prompt coverage: 30 fixed English prompts, UK location, two platforms, one run each.
That is less magical and more useful.
Build a prompt panel from customer demand
A prompt panel is a fixed set of questions you test repeatedly. Start with 20–40, not 500. Every prompt should represent a plausible customer situation. If you cannot connect a prompt to a real sales call, website search, support question, review, or Search Console query, leave it out of the core panel.
Use five buckets.
Problem prompts
The customer describes pain without naming your category.
- “How can I stop missing calls when my plumbing team is on jobs?”
- “How do I reduce no-shows at my dental clinic?”
- “What is the easiest way to get monthly management accounts for a small shop?”
Category prompts
The customer knows the solution type.
- “Best virtual receptionist for a small plumbing company”
- “Affordable ecommerce accountant in Bristol”
- “AI marketing employee for a two-person SaaS company”
Comparison prompts
The buyer is choosing between approaches or suppliers.
- “Virtual receptionist vs answering service for a local trades business”
- “Should I hire a marketing agency or use an AI marketing agent?”
- “Compare Company A and Company B for weekend support”
Evidence prompts
The buyer needs confidence.
- “Which providers support UK data residency?”
- “What should commercial cleaning cost for a 2,000-square-foot office?”
- “Show examples of small companies using AI employees for marketing”
Local and constraint prompts
The customer adds the details that decide the sale.
- location;
- team size;
- budget;
- industry;
- language;
- deadline;
- integration or certification;
- accessibility, dietary, or service constraint.
Build prompts from sales calls, site search, support tickets, reviews, Search Console queries, and the questions customers ask before buying. Do not let an AI invent the whole panel from generic keyword lists.
Build prompts from the buying journey
Track the questions customers actually ask.
- 1
Problem
How do I solve this costly or frustrating situation?
- 2
Category
What type of product or service should I consider?
- 3
Comparison
Which options fit my size, budget, or workflow?
- 4
Brand
Is this provider credible and suitable?
- 5
Local or specific use
Who can solve it in my market or circumstance?
Keep a stable core panel; label the date when wording or test conditions change.
Write down the test conditions
For each run, save:
- platform and product mode;
- model if shown;
- date and time;
- country or test location;
- language;
- logged-in, logged-out, or API context;
- web search on, off, or automatically invoked;
- exact prompt;
- whether it was a fresh conversation;
- full answer;
- cited domains and URLs;
- screenshots or machine-readable export where permitted.
Use fresh conversations for the core benchmark. Otherwise the previous turns can steer later answers. If personalization cannot be disabled, label the run accordingly.
Do not mix tests from a consumer chat product, a developer API without web search, and a third-party tracker in one trend line. They measure different environments. If you change platform, search mode, location, or prompt wording, mark a break in the series instead of calling the movement growth or decline.
A simple scoring system that does not lie
Score every prompt-platform observation with separate fields.
| Field | Value |
|---|---|
| Brand mentioned | Yes / No |
| Brand recommended for the stated need | Yes / No / Ambiguous |
| Brand order among named options | First / Second / Third+ / Not listed |
| Own website cited | Yes / No |
| Third-party page about brand cited | Yes / No |
| Factual accuracy | Correct / Minor error / Material error / Not applicable |
| Sentiment | Positive / Neutral / Negative / Mixed |
| Competitors named | List |
| Source URLs | List |
| Reviewer note | Short explanation |
Calculate rates only when the denominator is visible:
Mention rate = answers mentioning brand ÷ answers tested
Citation rate = answers citing owned website ÷ answers tested
Recommendation rate = answers recommending brand as a fit ÷ answers tested
Observed accuracy = correct brand mentions ÷ brand mentions reviewed
Never combine “first recommendation,” “mentioned in sources,” and “not present” into a mysterious 0–100 score unless the weights are published. If you create a summary score for internal triage, keep the raw fields beside it.
Add confidence labels
Every result needs a confidence label based on coverage and repeatability.
Low confidence
- one run;
- a handful of prompts;
- unclear location or search mode;
- answers not saved;
- major prompt-panel changes since the previous period.
Moderate confidence
- stable panel across multiple buying stages;
- conditions recorded;
- answers and citations saved;
- two or more runs for important prompts;
- referral data checked separately.
Higher confidence
- repeated across dates and relevant platforms;
- enough observations to show the result is not one random answer;
- stable extraction and human spot checks;
- synthetic findings agree with referral, conversion, or customer evidence.
“Higher” still does not mean universal. It means dependable enough for a business decision within the tested scope.
The monthly LLM visibility routine
Week 1: run the panel
Run the fixed panel under the same conditions. For your five or ten most valuable prompts, run each more than once on separate days. Save complete answers rather than only the extracted score. Give the owner a short exception report: new material errors, lost high-value visibility, and customer questions with no credible source.
Week 2: inspect source patterns
Ask:
- Which sources are repeatedly cited?
- Are they official sites, directories, news, review platforms, forums, or competitors?
- Which specific claim does each source support?
- Does your own site offer a clearer, current, first-hand answer?
- Are important pages crawlable and indexable?
OpenAI says inclusion in ChatGPT Search depends in part on allowing OAI-SearchBot. Perplexity recommends allowing PerplexityBot for search-result visibility. Google says pages eligible for AI Overviews and AI Mode must be indexed and eligible to appear with a snippet; its AI features guidance says there are no separate technical requirements beyond established Search fundamentals.
Crawler access creates eligibility, not a promise of citation.
Week 3: improve one source of truth
Choose one customer question where:
- the business has real expertise;
- the current answer is absent, outdated, or weak;
- buyers care about the answer;
- you can publish verifiable, specific information.
Useful improvements include:
- a clear service-area and availability page;
- an honest pricing methodology or calculator;
- a comparison based on named criteria;
- a case study with dates, scope, and actual results;
- a maintained specification or integration page;
- a concise answer to a high-friction buying question;
- structured organization and product information that matches the visible page;
- consistent business facts across your website and reputable profiles.
Do not create 100 near-identical “best X in every town” pages. They weaken trust and create maintenance debt. Publish the answer you would be comfortable sending directly to a customer.
Week 4: connect exposure to revenue
Review:
- referral sessions from AI platforms;
- landing pages receiving them;
- calls, forms, signups, or purchases;
- self-reported “How did you hear about us?” answers;
- assisted conversions where your analytics supports them;
- sales-call mentions such as “ChatGPT recommended you.”
OpenAI referrals may be identifiable in analytics, but not every AI interaction sends a referral. A buyer can remember the name and later search it on Google. Add a short, optional discovery question to high-value forms and let salespeople record the answer without forcing customers into a rigid dropdown.
What Google’s AI reporting changes
Google says appearances in AI Overviews and AI Mode have historically been included in the overall Web search type in Search Console. In June 2026, Google also announced dedicated Search Generative AI performance reports, designed to show impressions in generative AI features while retaining them in the overall report.
Use those reports where available, but do not merge Google AI impressions with ChatGPT citation tests and call the total “LLM reach.” Platforms count different events. Keep platform-specific measurements, then summarize them in a business dashboard.
A one-page SMB dashboard
Your monthly dashboard can be simple:
Exposure
- 30 core prompts tested;
- 12 ChatGPT Search mentions;
- 8 Perplexity mentions;
- 9 owned-site citations across both;
- 2 material factual errors.
Demand
- 38 measurable AI referral sessions;
- 11 visits to commercial pages;
- 5 CTA actions;
- 2 qualified enquiries;
- source reported as AI assistant on 3 additional forms.
Action
- update the service-area page with verified coverage;
- correct an obsolete price on the public pricing page;
- request correction of a major third-party directory listing;
- leave 24 prompts unchanged;
- rerun the high-value panel next week, full panel next month.
Label these as example fields, not industry benchmarks. Your baseline is more valuable than somebody else’s invented average.
An honest visibility report
Five sampled signals—none of them a universal ranking.
Mention rate
Share of sampled answers that name the brand
Citation rate
Share that link to an owned or trusted source
Accuracy
Whether price, capability, and positioning are correct
Source diversity
Which domains and evidence types shape the answer
Customer evidence
Referrals, branded searches, demos, and conversations
Label the model, mode, location, prompt panel, sample count, and observation window.
Common measurement traps
Testing only your brand name
If the prompt includes your brand, a mention proves little. Most panel prompts should describe the customer’s need without naming you.
Changing the prompts every month
New prompts improve coverage but break comparability. Keep a fixed core panel and a separate exploratory panel. Promote a new prompt into the core only through a recorded change.
Treating citations as recommendations
Read the answer. A cited article may support a definition while the system recommends competitors.
Counting every mention as positive
Track material errors, poor fit, and negative context. Visibility that misrepresents your offer creates bad leads.
Optimizing for the tracker
If your team knows the exact panel, it may publish awkward pages engineered to trigger those prompts. Check whether improvements also help real customers and whether referral conversions rise.
Believing synthetic tests equal customer behavior
The panel samples possibilities. Actual customers use prompts you do not know, across settings you cannot reproduce. Referral, conversion, sales, and customer-reported evidence keep the synthetic score honest.
Reacting to one missing mention
Answer variation is expected. Require repeated decline across multiple relevant prompts or a matching business signal before changing strategy.
Stop conditions for an AI visibility loop
Pause or escalate when:
- extraction cannot distinguish a real recommendation from a citation;
- the platform or search mode changed and the series is no longer comparable;
- the prompt panel no longer reflects customer demand;
- a proposed “optimization” would make the website less helpful or less accurate;
- the only available supporting claim is invented, private, or unprovable;
- the business lacks enough referral or conversion volume to justify weekly changes;
- three content changes increase mentions but not qualified demand;
- crawler access, indexing, analytics, or canonicalization is broken;
- a material factual error could harm customers and needs human handling now.
The loop should be allowed to conclude: “visibility is stable; no action this month.”
What an AI employee can do
An AI marketing employee can prepare the repetitive parts when the required sources and platform access are connected:
- maintain the prompt registry;
- schedule runs where platform terms and tooling allow;
- save answers, citations, and conditions;
- extract fields and flag uncertainty;
- compare periods;
- join referral and conversion reports;
- identify recurring source gaps;
- draft one improvement brief;
- verify that published claims cite real evidence;
- schedule the next check.
A human should approve the prompt strategy, review ambiguous recommendations, verify sensitive facts, and decide whether visibility is producing the right customers. In Teamday, Nova, the AI marketing employee, can coordinate recurring marketing work as Missions; external tests and analytics still depend on the tools you connect. The broader AI marketing loop framework explains metrics, gates, memory, and rollback.
The practical takeaway
You cannot measure every AI answer on the internet. You do not need to.
Measure a stable sample of the buying questions that matter. Save enough context to reproduce the observation. Separate mention, recommendation, citation, accuracy, referral, and revenue. Improve one truthful source at a time. Stop when the evidence is weak.
The goal is not to win an imaginary ChatGPT leaderboard. It is to be accurately discoverable when a real customer describes a problem your business can solve—and to know whether that discovery creates a useful conversation.
Your next action
Open your last 20 sales enquiries, customer questions, reviews, and on-site searches. Write ten buying questions in the customer's language without naming your company. Add five comparison or local-constraint questions, then run that 15-prompt starter panel in fresh conversations on one AI-search platform.
Save the complete answers and citations. Record mentions, recommendations, factual errors, and owned-site citations as separate yes/no fields. Rerun the same panel next month. Do not change your website because of a single missing mention; act only on a repeated source gap, material error, or customer question you can answer truthfully.
Frequently asked questions
What is LLM visibility?
It is how often and in what context AI answers mention, recommend, or cite your business, plus the traffic and customer outcomes that follow. It is a group of signals, not one universal rank.
Can I track my rank in ChatGPT?
You can track position within a fixed set of saved answers, but that is not a universal ChatGPT rank. Report the exact prompts, platform, dates, conditions, sample size, mention rate, citation rate, and confidence.
How many prompts should a small business track?
Start with 20–40 across problems, categories, comparisons, evidence questions, and local or buying constraints. Keep a stable core and add a smaller exploratory set.
How often should I measure it?
Run the full panel monthly. Test a small set of high-value prompts weekly when the market changes quickly or after an important correction. Avoid reacting to one answer.
Does allowing an AI crawler guarantee visibility?
No. It can make your public pages eligible to be discovered, depending on the platform. It does not guarantee indexing, inclusion, citation, recommendation, traffic, or sales.