DeepSmith

Nov 26 · AEO & AI Visibility

15 min read

How to Set Up AI Visibility Tracking for the First Time

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Abstract monochrome illustration of layered cards, connected nodes, and a search bubble behind the words Set Up AI Visibility Tracking.

If leadership has asked whether your brand shows up in ChatGPT and you only have a few manual searches to point to, this guide is for you. We'll walk through how to set up AI visibility tracking from zero, so that by the end you have a starting set of prompts, your brand and a few competitors entered, the engines chosen, and your first collected answers open in front of you. It's an AI visibility tracking guide for a first setup, so it stays on the basics and leaves tool buying and deeper strategy for other articles.

You'll need about an hour, a list of the products or services you sell, the main web address of your site, and the names of two or three companies you already think of as competitors. If you can find a few customer questions from sales calls or support tickets, that helps too, but you can start without them.

An AI visibility tracker measures how a chosen set of AI engines answer a chosen set of prompts over time. That's worth keeping in mind from the start, because it means the results describe the questions and engines you picked, not every conversation people have with AI tools. Your first numbers are a sample, and that's fine for a first baseline.

Step 1: Decide what the tracking should tell you

Before you open any dashboard, write down the one question you want the first round of tracking to answer. A good one sounds like this: when a possible buyer asks an AI engine about our category and use cases, does our brand show up, and do the answers link to our pages?

Keeping it that narrow makes the later steps easier, because you'll know whether a prompt belongs in the list or not. You can also write down which signals you care about. Most teams look at three of them (the best AI visibility tools roundup shows how tools package them): whether the brand is named, whether an answer links to one of your own pages, and which competitors appear instead.

You're done with this step when you can say, in a sentence or two, who the buyer is, what situations they ask about, and what you want to observe.

Common mistake: Treating a visibility score as a measure of all AI discovery. A score only reflects the prompts, engines, location, and time period you configured, so it's best to say that plainly when you share it with anyone.

Step 2: Build a small prompt set

A prompt is the full question someone types into an AI tool. It's worth starting from real buyer situations, not from short SEO keywords, because people tend to ask AI engines long, conversational questions. If you'd rather find ideas from scratch, the prompt discovery playbook covers that. Good places to look for ideas are customer emails, sales and support conversations, the questions on your own FAQ pages, and your existing search queries. Queries from Google Search Console can be a decent starting point if you rephrase them as full questions. If you want a method for the sales and support side, there's a separate guide on how to mine support tickets and sales calls for prompts.

One thing to know is that AI platforms don't give marketers a public feed of what people ask them. So your first prompt list is a deliberate sample that you chose, and not a list of questions with known search volume.

A manageable first batch is about 10 prompts. That number comes from third-party setup advice, so it's a practical starting point and not a rule or a guarantee that 10 prompts is statistically representative. DeepSmith's plan limits are higher (Pro tracks 50 prompts, Grow 100, and Scale 200), which gives you room to grow later. If you're wondering how big the list should eventually get, the article on how many prompts to track goes into it.

Aim for distinct buyer decisions instead of ten rewrites of the same question. A starter set can include a few of each of these, with the brackets replaced by your real category, audience, and use cases:

  • Category discovery: "What should a [buyer type] look for in a [category] solution?"
  • Problem or use case: "What tools help a [buyer type] solve [specific problem]?"
  • Constraints: "What is a good [category] option for a team with [meaningful constraint]?"
  • Evaluation: "How should a [buyer type] compare options for [use case]?"
  • Brand awareness: "What does [brand] offer for [use case]?"
  • Comparison: "How does [brand] compare with [competitor] for [use case]?"

These are templates, not real customer queries, so fill them in with the language your own buyers use. Include prompts that name your brand and prompts that don't. The ones without your name show whether you appear when nobody cues the engine, and the ones with your name show how the engine describes you. There's a longer explanation of this split in the piece on branded, unbranded, and competitor prompts.

Keep two prompts separate if a change in audience, constraint, or use case could change the answer or the sources. Then sort the list into a few families, such as category, use case, and evaluation, so you can read your first report by theme.

DeepSmith's Discover Prompts can help here. It suggests candidate prompts from your product, persona, and buyer-stage context, and you pick which ones to track. You can also type your own, and other trackers such as Ahrefs describe the same idea in their custom prompts documentation. Read the suggestions before you accept them, because some will overlap, and some won't match how your buyers talk.

You're done when you have a list of distinct questions that fit your audience and that you can group into a few families.

Pro tip: Keep a stable "Core" group for your baseline and put new ideas in a separate "Explore" group. If you ever swap out a core prompt, write down the change, because a different prompt set isn't a perfect like-for-like comparison with earlier results.

Step 3: Enter your brand and a few competitors

Next, tell the tool who you are, which is the point where you really start tracking AI mentions of your own name. Use the brand name your customers and AI answers actually use, along with your main public domain. The name is what lets the tool spot mentions in text, and the domain is what lets it spot links to your own pages as citations. Check the spelling, spacing, and product names carefully, because a small typo here can quietly skew everything after it.

In DeepSmith, the settings cover what counts as your brand and who your competitors are. The public materials don't spell out every detail of the screen, so go by what your workspace shows you. In particular, don't assume there's an alias field unless you can see one.

Then add a short list of competitors you already recognize as relevant to the same buyer decisions, using their usual names and main domains. Two or three is plenty for a first pass. Deciding which competitors deserve a spot is its own topic, so for now just add the ones you already know are in your buyers' shortlists.

If your brand name is also an everyday word, an abbreviation, or the name of something else, read the first answers closely for false matches.

You're done when the workspace shows the right brand and domain, and the competitor list holds only the peers you need for a first comparison.

Common mistake: Entering a parent company name when customers and AI answers use a product name, or typing a domain that isn't your site's main public address. Double-check both before you read any results.

Step 4: Choose the engines and the market

Now pick the AI engines to track. You can only choose from the ones your plan includes, so look at what your workspace offers. On DeepSmith, Pro covers ChatGPT, Grow adds Perplexity, and Scale adds Gemini, while Enterprise and Custom plans go beyond those. Write down which engines are in your baseline so you can remember later what you measured.

Pay attention to location and language as well. AI answers can differ depending on where the question is asked, so results from two markets shouldn't be treated as the same measurement. If your account shows a location or language setting, note the choice and keep it the same from run to run. Not every plan has the same region controls, so you may not see one. If you do work across several markets, the guide on tracking visibility across countries and languages covers it.

You're done when the engines you picked are available to your workspace and you've written down the engines and market in your baseline notes.

Common mistake: Comparing results across engines or locations as if they were one measurement. If ChatGPT and Perplexity look different, that's a platform difference worth reporting as such, not a sign that something broke.

Step 5: Set the schedule and start the first run

DeepSmith runs your tracked prompts on a schedule and keeps each full answer so you can read it later. Choose one of the schedule options your workspace offers and then leave it alone for the baseline, since a steady rhythm makes later comparisons cleaner.

Before you start, take one more look at the brand name, the domain, the prompts, the competitor list, and the engines. It's much easier to fix a typo now than after a week of collected answers. DeepSmith's public page says tracking starts within minutes, though that doesn't mean every prompt and engine pair will have a finished answer right away. Give the first run time to finish before you draw any conclusions, since the point of this step is to start tracking AI mentions on a steady footing.

You're done when your prompts are scheduled and the workspace shows collected answers for the engines you chose.

Common mistake: Reading an empty result as proof that your brand wasn't mentioned. First check that an answer was collected at all. If one wasn't, you have nothing to judge yet, and it's a matter of waiting or checking the run, not of being invisible.

Step 6: Read the answers before the score

Once answers come in, open the actual text first. For each engine and prompt, look at whether your brand is named, what the answer says about it, whether a page of yours is linked, which other brands show up, and what sources the engine used. DeepSmith keeps the full answer behind each metric, so you can always click through from a number to the text that produced it. The Pages view and the competitor citation view help here too, since they show which of your pages were cited, which prompts led to those citations, and which rival pages appear for the same prompts.

A DeepSmith answer view showing a full ChatGPT answer to a buyer question with its source links, next to a card that records the mention rate, citation rate, and position tracked for that prompt, using demo data.

It helps to know the terms the dashboard uses:

  • Mention: the answer names your brand.
  • Citation: the answer links to one of your pages as a source.
  • Mention rate and citation rate: how often each of those happens across your tracked answers. They're reported separately because an answer can name you without linking to you, or link to you without naming you. The article on what counts as an AI citation goes deeper on that line.
  • Share of voice: how visible you are compared with your competitors, within the scope you configured. The exact formula isn't published in the materials we have, so avoid comparing scores that were built from different prompts, engines, competitors, or time periods.
  • Visibility trend: how things change over time, which you should read against the dates and the prompt and engine set you recorded.

A link to someone else's page that happens to mention you isn't the same as a link to your own page, so check which one a number or an answer is showing.

You're done when you can open one real answer and say whether your brand was mentioned, whether one of your pages was cited, and which prompt and engine produced it.

Pro tip: When a number moves, open the underlying answers and the source pages before you decide what changed. A rate can't tell you whether the engine described you accurately or why a competitor was linked instead of you.

How to read your first baseline

AI answers vary. If you ask the same prompt twice, one answer might name you, the next might leave you out, and a third might feature a competitor. So the first result is a snapshot, and it isn't proof of a lasting win or loss. A pattern across your whole prompt set and across several scheduled runs tells you much more than any single answer. The article on whether AI visibility scores are reliable explains the day-to-day noise in more detail. For background on how Google's AI features relate to your site, see Google's AI features and your website documentation.

To keep comparisons fair, hold the core prompts, the engines, the market, and the reporting period steady. If you change any of them, write the change down and treat the comparison with some care. When you share a report, say what it covers: the prompt families, the audience or buyer stage, the engines, the market or language if you set one, and the dates.

There's no single number that counts as a good score, and the sources we reviewed don't set a universal benchmark or a minimum sample size for a trustworthy first reading. If you see round percentages quoted as a standard, it's safer to leave them out of your own report. The same goes for example figures shown in any product demo, which are there to show how a screen works and aren't results you should expect.

What to do next

Over the next couple of weeks, let the schedule run and check in on the same prompts each time. Once you've seen a few rounds, you'll have a feel for where you show up and where you don't. At that point you can write up a short note listing the prompts where a competitor is cited and you aren't, and use it to decide what to publish. If you want a plain way to keep the set steady, the guide on setting up a repeatable prompt set is the natural next read.

When you're ready to run this with your own brand, you can try DeepSmith free for 7 days and run these same steps with your own prompts.

Frequently asked questions

How many prompts should I start with?

About 10 is a manageable first batch, though it's a starting suggestion and not a universal rule. Choose distinct buyer decisions, and stay within the prompt allowance of your workspace plan.

Should I track prompts that don't mention my brand?

Yes. Include category, problem, and use-case questions alongside some questions that name your brand. That way you can see whether your brand appears when the question doesn't name it. It's a balanced way to measure, and it doesn't claim to cover every real AI search.

What is the difference between a brand mention and a citation?

A mention is your brand name appearing in an AI answer. A citation is a link to one of your own pages used as a source. They're separate signals and can move independently of each other.

What if the first answer doesn't mention us?

Open the full answer and confirm that the run finished. One omission is a snapshot, not proof that you're always invisible. Look across your whole prompt set and the next scheduled runs before you decide anything.