AI Visibility Tracker Setup in 15 Minutes

You can set up an AI visibility tracker in about 15 minutes if you bring three things: 10-20 real buyer prompts, a list of 3-5 competitors, and a decision on which engines matter. Paste the prompts, set engines and market, add brand name variants, choose a schedule, then read the first run's raw answers before trusting any chart.

On this page

• From MentionX data: what a working tracker looks like
• What do you need before you start?
• Minutes 0-3: How should you enter prompts?
• Minutes 3-6: Which engine settings matter?
• Minutes 6-9: How do you set up competitors?
• Minutes 9-12: What schedule and alerts?
• Minutes 12-15: How do you check the first run?
• Worked example: Ledgerline's first 15 minutes
• How many runs per prompt do you need?
• What should you know about each engine before you start?
• What should you do in week one?
• What are the most common setup mistakes?
• How do you know the tracker is working?
• Which numbers should you look at first?
• How do you report it?
• Setup checklist
• What should your first 20 prompts look like?
• How do you sanity-check the first results?
• What should you leave out at the start?
• What should you document on day one?
• Common questions
• Sources

Settings checked against OpenAI and Google documentation on 30 October 2026. "Our view" marks MentionX opinion. Ledgerline is fictional.

From MentionX data: what a working tracker looks like

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 54 scans, 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

Mention rate by intent, all engines:

What this means for setup. A prompt set of around 15 is a normal starting size. Some runs will fail, so a tracker needs retries and should report how many runs completed. Answers vary, so single runs mislead. And intent mix changes the headline rate more than almost any other setting.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from consumer apps.

What do you need before you start?

If you have no prompts yet, follow Building your first prompt set and come back.

Minutes 0-3: How should you enter prompts?

• Paste as full questions, the way a buyer would type them.
• Tag each with intent: learn, compare or buy.
• Leave out branded prompts for the main set; track them separately if needed.

Minutes 3-6: Which engine settings matter?

• ChatGPT: whether search is used. Answers with search can cite current pages; OpenAI's OAI-SearchBot is what surfaces sites in ChatGPT search (OpenAI).
• Google: AI Overviews and AI Mode separately. Google says they may use different models and show different links (Google).
• Location and language: match your market. A US answer may name different vendors than a German one.

Minutes 6-9: How do you set up competitors?

Add each competitor with every variant buyers and engines use: "HubSpot", "HubSpot CRM", "Hubspot". Do the same for your own brand. Missed variants are a common cause of under-counted mentions.

Minutes 9-12: What schedule and alerts?

Avoid alerting on every single-answer change; you'll mute the alerts within a week.

Minutes 12-15: How do you check the first run?

Open five raw answers and check:

1. Was the right engine mode used?
2. Are mentions of your brand and competitors detected correctly?
3. Are all cited URLs captured?
4. Is the location right (local vendors, currency)?
5. Does anything look obviously wrong, such as a mismatched brand?

Fix settings now; it's much harder to correct months of data later.

Worked example: Ledgerline's first 15 minutes

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

• Minutes 0-3: The marketing lead pastes 16 prompts from sales call notes and tags them: 4 learn, 7 compare, 3 buy, 2 branded accuracy.
• Minutes 3-6: ChatGPT with search, Google AI Mode, Perplexity and Copilot; market UK, language English.
• Minutes 6-9: Three competitors, plus variants: "Ledgerline", "Ledger Line" and "Ledgerline AP".
• Minutes 9-12: Weekly, three runs per prompt, alerts only for a mention rate outside the last four weeks' range.
• Minutes 12-15: Reading five raw answers, she spots that one competitor is being counted in answers about an unrelated accounting textbook with a similar name. She adds an exclusion term before the second run.

The first run shows Ledgerline named in 31% of unbranded answers, but 47% in compare prompts and 8% in learn prompts. That split is more useful than the average: the gap is in early-stage questions, where the brand barely appears.

How many runs per prompt do you need?

More than one, and the right number depends on what you want to detect. A mention rate is a proportion, so its uncertainty shrinks with the number of answers behind it (NIST). Roughly, for a true rate near 30%:

Twenty prompts × three runs × one engine gives 60 answers a week, enough to spot a large change but not a five-point one. To see smaller shifts, report four-week rolling rates, or combine engines for the headline and keep per-engine figures as context.

Our view: three runs per prompt per week is the practical minimum. Fewer, and you'll mistake normal variation for a trend. In our repeated pairs, 11 of 28 flipped between mentioned and not mentioned with no change to the prompt.

What should you know about each engine before you start?

Engines differ in speed and in how often they cite, which affects how you read the first run. From the same MentionX data:

Slow engines time out more often, so set timeouts generously and retry once. Engines that rarely cite will give you mention data but little source data. Don't read a low citation count from them as a problem with your site.

What should you do in week one?

• Share the first results with sales and ask: "Do these answers match what prospects tell you?"
• Add 5 prompts from their feedback (bump the prompt-set version).
• Note the top cited domains; they're your first off-page list.

What are the most common setup mistakes?

• Mixing branded prompts into headline rates, which makes you look stronger than you are.
• Wrong market, often defaulting to the US.
• No name variants, so mentions are missed.
• Changing prompts every week, which breaks trends.
• Never reading raw answers, so a scoring error goes unnoticed for months.

How do you know the tracker is working?

After two or three weekly runs you should see stable ranges for each rate, a consistent list of top cited domains, and at least one specific action, such as a page to rewrite or a source to pitch. If the numbers jump wildly with no change in the market, look at settings first.

Which numbers should you look at first?

Resist the urge to build a dashboard on day one. After the first two or three runs, look at four numbers:

1. Completed runs against planned runs, so you know the base is sound.
2. Mention rate on unbranded prompts, split by intent.
3. Share of answers citing any page about you, split by engine.
4. The ten most-cited domains in answers where competitors appear and you don't.

Each one points to a different kind of work: fixing the setup, writing content for a weak intent, fixing crawler access or pages, and outreach to the domains engines rely on. Anything beyond these four can wait until the tracker has run for a month.

How do you report it?

Use the monthly report template: headline rates, by engine, by topic, sources, accuracy and traffic, plus a method box.

Setup checklist

• 10-20 prompts, tagged by intent
• Engines and modes selected
• Market and language set
• Own brand and competitors with name variants
• Weekly schedule, 2-3 runs per prompt
• Alerts limited to meaningful changes
• First run's raw answers reviewed
• Prompt set saved as v1

What should your first 20 prompts look like?

Split them across intent so the first run tells you something useful. Illustrative example for a project management tool:

Write them the way a buyer would type or speak, not as keywords. Tag each with intent, product line and market before you import them, because adding tags later means re-mapping history.

How do you sanity-check the first results?

Pick five prompts and run them yourself in the consumer interface with the same engine settings. Compare three things: whether the same brands appear, whether yours is mentioned, and whether the cited domains overlap. Full agreement isn't expected, because answers vary between runs, but if your manual checks and the tracker disagree on most prompts, check location, language and mode settings before trusting any trend.

What should you leave out at the start?

• Long-tail prompts you can't act on yet.
• Engines your buyers rarely use.
• Daily schedules; weekly or monthly is enough until you know your variance.
• Custom scores. Start with plain mention, recommendation and citation rates.

You can add all of these later; removing them after reports exist is harder.

What should you document on day one?

Write a short method note: prompt set version, engines and modes, location and language, run frequency, competitors tracked and who owns the tracker. Store it next to the data. When numbers shift in month three, this note is how you tell a real change from a settings change.

Common questions

Do I need to track every engine? No. Pick the two or three your buyers use most. Add others once the process works.

How do I handle failed runs? Retry them once, and report completed runs as the base for every rate. A 2.8% failure rate doesn't matter much; a sudden jump does, because it usually means an engine setting or quota has changed.

When should I change the prompt set? Rarely, and on purpose. Add or remove prompts in batches, give the set a new version number, and note the date in your report. Changing prompts every week makes trends impossible to read.

What if a competitor is never mentioned? Check the name variants first, then read a few answers. Sometimes they're named under a parent brand or product name. If they really don't appear, keep them anyway: their absence is part of the picture.

Is weekly too often? For most brands, weekly runs with monthly reporting work well. Daily runs mostly add noise unless you're watching a launch or a known problem.

Should I use the API or the consumer app? APIs are repeatable and scalable; consumer apps are what buyers see. Use the API for tracking and check a sample in the app each month.

Sources

• OpenAI, Overview of OpenAI crawlers
• Google Search Central, AI features and your website
• OpenAI, Web search tool; Google, Grounding with Google Search; Perplexity, API overview
• NIST/SEMATECH, Confidence intervals for a proportion