DeepSmith

Nov 26 · AEO & AI Visibility

18 min read

How to Set Up AI Brand Monitoring in One Afternoon

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Abstract monochrome cover with the words Monitor AI Answers Fast over layered checklist cards, a clock, and connected question and answer boxes.

If someone on your team has asked what ChatGPT or Perplexity says about your company, and nobody could give a straight answer, this guide is for you. You can set up AI brand monitoring in about three hours of focused work: pick one question to answer, check your brand settings, save a small list of real buyer prompts, and review the first completed answers. By the end you'll have a saved baseline you can compare against later. This is a practical brand monitoring AI how to, with a few common mistakes flagged along the way, and it assumes you're choosing from the AI brand monitoring tools that track answers on a schedule instead of checking by hand.

The times below are planning estimates for a focused session, not product limits. How fast the first answers come back depends on the engines and the collection schedule in your workspace, so the review step is written to work even if the first collection finishes after you've wrapped up for the day.

The plan for the afternoon

Here is how the three hours break down. Each block ends with something you can point to, which is how you know you can move on.

StepSuggested timeYou're done when
1. Decide what the first pass should answer15 minutesOne audience, market, and question are written down
2. Check your brand and monitoring scope25 minutesBrand name, domain, variants, competitors, and engines are confirmed
3. Build a starter prompt list40 minutesAbout 12 to 15 real buyer questions are drafted
4. Add the prompts to AI Visibility20 minutesThe saved list matches the draft
5. Make the list comparable15 minutesYou know which prompt and engine pairs to expect
6. Review the first answers45 minutesEach observation points to a real answer
7. Save a baseline20 minutesSomeone else could read your notes and understand them
8. Name an owner and the next review10 minutesA person and a date are on the calendar

You'll need access to your DeepSmith workspace (or a free trial), a spreadsheet or doc for notes, and ideally one person from sales or support who can tell you what buyers really ask.

Step 1: Decide what the first pass should answer

Start by writing one sentence that says what you want to learn. A good one sounds like this: "When buyers ask about our product category, which brands and sources do AI answers bring up?" That's plenty for a first pass. Along with the sentence, pick the audience and buyer stage you care about most right now, and one market and language for the baseline if your workspace lets you set those.

It's tempting to cover every market, persona, and topic at once, but a small set that your team can understand and repeat is more useful as a first baseline than a big, loosely defined one. AI tools answer the whole question as it's phrased, not just a keyword, and market and language can change what comes back. So write down the market you chose, and don't assume what you see for one region holds for another.

If you want a broader picture of what AI brand monitoring tracks and how repeated checks turn into reports, that's worth a skim before you start, but you can do this step without it.

You're done when a teammate can say which audience, market, and question your first prompt set stands for.

Common mistake: trying to watch every engine, market, and persona on day one. You can add more later, and it's much easier to read a small result.

Step 2: Check your brand and monitoring scope

Next, open the AI Visibility setup in your workspace and look at what it's using to recognize your brand. Check the brand name, your website domain, and any common variants, such as a short name people use or a product name that often appears next to yours. Then look at the competitor list, the categories, the collection schedule, and the engines your plan includes.

DeepSmith tracks buyer questions on a schedule and keeps the full answers, with mentions and citations counted separately. The engines available depend on your plan, so check what your workspace can actually use instead of assuming every engine is on. Pro covers ChatGPT, Grow adds Perplexity, Scale adds Gemini, and Enterprise covers every engine DeepSmith supports. If you're comparing options, AI visibility tools differ a lot on which engines they cover, which is a good reason to confirm this before you build a prompt list around an engine you can't collect from.

Pay attention to names that are easy to confuse. If your brand name is also a common word, a mention check can match the wrong thing, and you'll end up cleaning results by hand. It also helps to decide now how you'll choose which competitors to track, because a competitor list that grows without a reason makes the first review harder to read. A short list of the rivals buyers actually mention is plenty.

You're done when the brand and domain are correct, the name variants are covered, the competitor list is on purpose, and you know which engines you'll be collecting from.

Common mistake: changing brand matching or the competitor list after collection starts, and then treating the new numbers as if nothing changed. If you do change something, write down what and when, so you can explain a jump later.

One thing to keep straight from the start, since it comes up again in the review. A mention means the answer names your brand. A citation means the answer links to one of your pages as a source. Neither one means the AI recommended you, and they can happen separately. An answer can name you without linking to your site, or link to your page without saying anything good about the company.

Step 3: Build a buyer-led starter prompt list

This is where most of your time goes, and it's worth it, because the prompts decide what you can learn. A prompt is just a question a buyer might type into an AI tool. Aim for about 12 to 15 to start. That number is a suggestion, not a minimum or a plan limit, and your plan will allow more tracked prompts than this (Pro includes 50, Grow 100, and Scale 200).

Pull the questions from real places: sales calls, customer interviews, support tickets, and what your own team hears all day. Write them as natural, complete questions, and cut near-duplicates that test the same thing. Cover three kinds of question so the list isn't lopsided:

  • Unbranded discovery: the buyer doesn't know you yet.
  • Category and comparison: the buyer is weighing options.
  • Branded accuracy: the buyer asks about you by name.

Here's an example list for an AI visibility platform. Swap in your own brand, product, and audience, and delete anything your customers wouldn't really ask. These are examples, not evidence that a buyer asked them or that any engine will answer a certain way.

Unbranded discovery

  1. How can a marketing team track whether AI assistants mention its brand in answers?
  2. What tools monitor brand mentions and citations across AI search engines?
  3. How do AI visibility platforms measure brand mentions, citations, and share of voice?
  4. How can a team find which website pages AI engines cite?
  5. What is the difference between AI brand monitoring and traditional brand monitoring?

Category and comparison

  1. What are the best AI visibility tracking tools for content marketing teams?
  2. How do AI visibility platforms compare when tracking prompts and cited pages?
  3. Which tools monitor competitor citations in ChatGPT, Perplexity, and Gemini?
  4. How should a marketing team set up prompt tracking across AI answer engines?
  5. What should I look for in software for LLM response monitoring?

Branded accuracy and consideration

  1. What does [brand] do?
  2. What is [brand]'s AI visibility platform for?
  3. How does [brand] track citations across AI engines?
  4. What are the main alternatives to [brand] for AI search monitoring?
  5. Which AI brand-monitoring tool is a fit for [specific buyer or use case]?

You're done when every prompt has a reason to be there, and the list has both unbranded and brand-specific questions.

Pro tip: give the list a version label and a date, something like "v1, starter set." If you change prompts later, you can mark it as v2 and compare like with like instead of quietly mixing two different tests.

Step 4: Add the prompts to AI Visibility

Now put the list into the product. Open the Prompts area under AI Visibility, where you can see every question being tracked. If you want ideas beyond your own list, Discover Prompts can suggest a starter set based on your product and persona context. Treat those as suggestions. Keep the ones that sound like your buyers and reject the rest.

Add the questions you chose on purpose, which is the core of any LLM response monitoring setup, then check the engines and the schedule your workspace is set to collect. The schedule is configurable, but this guide can't tell you an exact interval or a time for the first result, so look at what your own workspace shows instead of assuming a daily run. Other tools describe their own cadence in their help pages, and that doesn't carry over.

This is a good moment to run through your LLM response monitoring setup one more time as a checklist: brand and variants correct, competitors chosen, prompts saved, engines confirmed, schedule known. If one of those is off, fixing it now costs a minute, and fixing it after the first review costs you the clean baseline.

You're done when the saved prompt list matches the list you meant to track, the wording reads clearly, and you know which engines will be asked.

Common mistake: assuming a suggested prompt is relevant because the tool generated it, or assuming an answer on one engine tells you what the others say. Read the suggestions, and check which engines you actually have before you read any results.

Step 5: Make the prompt list comparable across engines and time

A baseline is only useful if you can repeat it. When your workspace covers more than one engine and you want to compare them, keep the wording of each question and the market the same across engines. Then keep using the same list on the same schedule, so each review looks at like for like.

Write down anything that could make two results hard to compare: an engine that wasn't available yet, a different market, a prompt you reworded, or a collection that hasn't finished. Later you'll want to say which prompt and engine pairs you expected, which ones completed, and which were missed or need another try.

Answers also move around, even when you ask the same thing twice. Research on repeated runs shows a good deal of variation in which sources and brands show up, which is a good reason to rely on several observations over time instead of treating one answer as a stable ranking. Those findings come from a specific study and shouldn't be read as a benchmark for DeepSmith or as a number to expect for your own brand.

You're done when you can list which prompt and engine combinations you expect to see, and you know what you'll compare next time.

Common mistake: treating a single answer as your ranking, or comparing a reworded prompt to its old version as if nothing changed.

Step 6: Wait for collection, then review the actual answers

Once at least one collection has completed, you can start the review. If it hasn't finished by the end of your afternoon, that's fine. Your setup is complete, and you can do this step as soon as the first answers are in. Nothing here promises the full set of results will be there in three hours, so don't plan around it.

Begin with the Overview in AI Visibility. It shows mention rate, citation rate, share of voice, trends, and a breakdown by platform. Those numbers are a good map, but they aren't the review. Next, open the Prompts view and read the answers behind the metrics. Then look at the cited pages and the competitor citations to see which exact pages and sources came up.

A DeepSmith answer view showing the full text an AI engine gave for a tracked buyer question, with the brand named in the list, the sources it cited, and the prompt's mention rate and citation rate beside it.

Keep a short worksheet with a row for each prompt and engine. These columns are enough:

  • Prompt and engine
  • Collection status and time
  • Was your brand named
  • Was one of your pages cited
  • Was a competitor named or cited
  • How the answer framed or recommended anything
  • Notable source
  • Follow-up question

Go through each answer in this order:

  1. Confirm the prompt and engine really ran.
  2. Read the whole answer before you code anything.
  3. Count a brand mention only if the answer names your brand, using your own rules for the name and its variants.
  4. Count a citation only if a visible link points to a page on your site.
  5. Record whether the answer recommends or favors you, separately from a bare mention.
  6. Note the competitors and sources, and open the important sources to check they support what the answer says.
  7. Mark anything inaccessible or missing as "unable to assess" or "missing," not as "no mention."

That last point matters more than it sounds. A run that failed or never finished says nothing about whether your brand was left out, so keep those apart from answers that were complete and simply didn't name you.

Opening the linked pages is the part people skip. A good citation audit means reading the answer and the source it links, then checking what the source really says. A link doesn't prove that page produced the whole answer, and a mention alone doesn't prove the AI endorses you.

If you'd like a wider view while you're in there, it can also help to see how AI explains your whole category, not only your own brand, and to note what the answers say about the space.

You're done when you can point to the answer behind every observation, tell completed answers from missed runs, and name a few concrete patterns without stretching them. Something like "three of the discovery prompts named a competitor and none named us" is the right size of finding.

Common mistake: looking only at a score or a count of domains. The score tells you where to look, and the answer tells you what happened.

Step 7: Save a baseline and read the first signals carefully

When the review is done, save what you'd need to repeat it. Write down the prompt set version, the date, the brand variants, the competitors, the market, the engines, and your review notes. Keep the underlying answers too, because a number without the answer behind it is hard to explain to anyone else.

For later comparisons, use the same prompt set and configuration across completed collection windows. DeepSmith's rates and trends are the platform's own metrics, so use them for the trend. If you also do a manual spot-check, say what you counted. For example, you might report how many completed prompt and engine answers you reviewed and how many were missing. That's a handy review aid, but it's your own count and not DeepSmith's formula for mention rate or citation rate. Other vendors also define their visibility metrics in their own ways, so don't expect numbers from different tools to line up.

Be careful with share of voice. It compares how visible the tracked brands are against each other, and it isn't market share or a sales measure. These presence numbers are useful for watching direction over time, and they don't prove business results. When a number moves, a useful question to ask is what counts as a meaningful change before you decide to act on it.

You'll probably want your data in other places too, such as a team report or a spreadsheet. DeepSmith has a read-only API and an MCP server for pulling workspace data out, and there's a good overview of exports and API options if you want to plan for that later.

You're done when someone else can tell what you measured, which answers you actually saw, and what changed since the last review.

Common mistake: reading a small shift as a trend after one collection. Wait for a few completed windows on the same prompt set.

Step 8: Name an owner and the next review

Last, give the work a home. Pick the person who owns the prompt set and the next review, and put a reminder on the calendar that matches the collection schedule and what your team can handle. At the next review, use the same prompts and market where you can, and note what changed instead of editing the baseline in place.

You're done when an owner knows when to look at the next completed collection, where the baseline lives, and how a change to prompts or brand matching gets recorded.

Common mistake: expecting a dashboard to make the decision for you. The monitoring shows you what the engines said. A person still has to read the answer and decide whether the sources back it up.

A quick note on alerts. DeepSmith tracks on a schedule, and sending a notice to Slack or email is a separate step you'd build around that. If you want it later, here's how to set up alerts for changes in how AI mentions your brand, and it's best to wait until you have a couple of reviews behind you.

What to do next

Keep your starter prompt list steady for the next few collections and review each one against the baseline you saved today. If you want a lighter option for a team that isn't ready for a platform, you can monitor LLM brand mentions without buying software using a spreadsheet and a routine. If you're still comparing platforms, the guide to tools a non-technical team can run is a good place to look.

When you're ready to do this with real data, you can start a DeepSmith free trial and work through the steps above in your own workspace.

Frequently asked questions

What should I track first?

Start with a small set of real buyer questions across discovery, comparison, and brand-specific accuracy. About 12 to 15 prompts is a good starting size. It's a suggestion and not a product rule, and a short list of distinct questions will teach you more than a long list of near-duplicates.

Does a brand mention mean the AI recommended us?

No. A mention means the answer names your brand. A recommendation is a separate thing, and so is a citation, which means the answer links to one of your pages. An answer can cite you without recommending you, or name you without linking to your site, so record each one on its own.

How soon will the first results appear?

DeepSmith collects your tracked prompts on a schedule and keeps the full answers, but there's no published promise for exactly when the first result lands. Check whether a collection has completed. If it hasn't by the end of the afternoon, your setup is still done, and you can review the first completed answers afterward.

Should I change my prompts if the results look quiet?

Check the basics first. Make sure the prompt ran, the engine you wanted was available, and your brand variants are set up correctly. Then keep the starter list steady long enough to build a comparable baseline. If you do change a prompt or brand matching, record the change and the date so you aren't comparing two different setups as if they were one.