How to track AI citations of your business
The short answer
Build a list of 50 to 100 prompts a potential customer might ask, run them across the major answer engines on a fixed schedule, and record whether your business is named or linked. Tracking the trend in that citation share over months is the metric that matters, not any single result.
You cannot rank in an answer. You can be named in one.
Search Console tells you where you rank. It does not tell you whether ChatGPT recommends you when someone asks for a web developer in Nairobi, or whether Perplexity cites your pricing page. Answer engines do not publish positions, so the only way to know is to ask them the questions your customers ask and write down what comes back.
That sounds crude, and it is, but it is measurable. Run a fixed set of prompts on a fixed schedule, record whether your business is named, linked, or absent, and you have a number that moves. The number itself matters far less than its direction over months, which is why the method has to stay identical every time you run it.
Two warnings before you start. Answers vary between runs even for the same prompt, so a single result proves nothing. And answers are personalised and localised, so run them from a clean session rather than from an account that has been talking to you about your own business all week.
Build the prompt set
Between fifty and a hundred prompts, written once and then left alone. Changing the set is how you lose the trend.
Start from the questions you already get
Your inbox, your WhatsApp and your sales calls are the best source. Write them the way people actually type them, not the way a keyword tool phrases them.
Cover the three intents separately
Choosing a provider ("best web design agency in Nairobi"), understanding a problem ("why is my website slow in Kenya"), and comparing options ("Shopify or WooCommerce for a Kenyan store").
Include your own name
"Who are Haryes Web Developers" style prompts tell you what the engines believe about you, which is often where the real errors show up.
Include your competitors
The same prompts without your name, so you can see who is being recommended instead and what those pages have that yours do not.
Localise properly
Kenya, Nairobi, Mombasa and the towns you serve, as separate prompts. National and local answers differ more than people expect.
Freeze it
Put the list in a sheet with an ID per prompt. Add new prompts in a second batch rather than editing the originals.
Record the same four things every time
Keep the scoring boring enough that two different people would record the same result.
Named
Your business is mentioned by name in the answer, whether or not there is a link.
The strongest outcome.
Linked
A specific page of yours appears as a source or citation. Record which page, because that tells you what is doing the work.
Note the exact URL.
Competitor named
Who was recommended instead. Over time this becomes the most useful column in the sheet.
Look at what they publish.
Wrong about you
Services you do not offer, a town you are not in, an old phone number. Inaccuracy is a separate problem from absence.
Fix the source, not the answer.
Run it on a schedule you can keep
Pick the engines that matter to you
Google AI Overviews, ChatGPT, Perplexity and Gemini cover most of it. Adding more engines than you will actually check every month is how the habit dies.
A short, honest list.
Use a clean session
Logged out or in a fresh profile, with location set to the market you are measuring. Personalisation will otherwise tell you what you want to hear.
Results that resemble a stranger's.
Run monthly, on the same week
Weekly is noise for most businesses. Monthly is frequent enough to catch a change and infrequent enough that you keep doing it.
A trend rather than a snapshot.
Score, do not interpret, while running
Fill the sheet first. Analysis after. Mixing the two is how the scoring quietly drifts in your favour.
Comparable data.
Review the quarter, not the month
Three runs is the smallest sample worth drawing a conclusion from, given how much answers vary.
A decision you can defend.
What to report
Three figures, calculated the same way every run.
- Citation share
- Named ÷ prompts
The share of your prompt set where your business is named at all.
- Link share
- Linked ÷ prompts
The narrower share where a page of yours is cited as a source.
- Accuracy
- Errors found
How often an engine states something about you that is wrong.
Your own baseline is the comparison. There is no industry benchmark worth quoting.
Doing it by hand, or paying for it
What paid tools do not solve
- They still sample a varying system, so the noise does not disappear.
- Their prompt sets are generic unless you supply your own.
- Coverage of Kenyan local prompts is usually thin.
- Monthly cost is hard to justify for a business with fifty prompts.
What a spreadsheet gets you
- Prompts written in your customers' actual words.
- Full control of the scoring and the schedule.
- The raw answers kept, so you can re-read them later.
- An hour a month, which most teams can genuinely sustain.
Paid visibility trackers are improving and are worth looking at once your prompt set is large or you are reporting to someone else. Start manual, though. Doing three runs by hand teaches you more about how these systems describe your market than any dashboard will, and it tells you whether the metric is worth paying for in your case.
What to do with what you find
Absent everywhere
The engines have nothing to cite. The work is publishing substantive, answer-first pages on the topics in your prompt set, and making sure your organisation is a clearly defined entity in your schema.
Named but never linked
You exist as a fact but not as a source. Usually a structure problem: the answer to the question is buried in the middle of a page rather than stated near the top in a liftable form.
Competitors consistently preferred
Read what they publish on those exact questions. In most Kenyan sectors the difference is depth and specificity, not budget.
Described incorrectly
Correct the sources: your own pages, your Google Business Profile, directories and any profile carrying stale details. Engines summarise what is out there, so the fix is upstream.
Fair questions about this method
The answers change every time I run the same prompt. Is the data useless?
No, but a single run is. Variation is why the method depends on a fixed prompt set and a trend over several months. Treat any one answer as a sample, never as a ranking.
How many prompts do I really need?
Fifty is enough for a small business to see movement. Below about thirty, normal variation swamps the signal. Above a hundred, you stop running it, which is worse than a smaller set.
Does being cited actually bring customers?
Sometimes directly, through the link. More often it works like a recommendation: someone hears your name in an answer and searches for it afterwards. Watch branded search volume alongside the citation share.
Can I just ask the engine to recommend me?
No, and anyone selling that is selling something else. What you can do is make your business easy to describe accurately and easy to lift from, which is ordinary publishing work done with retrieval in mind.
