If your boss or a teammate has asked whether AI tools mention your brand, you can answer with real evidence this month, and you don't need to buy anything to do it. This guide shows you how to monitor LLM brand mentions with a spreadsheet, a short list of questions, and a routine you repeat on a schedule. By the end you'll have a log that shows where your brand appears in AI answers, where it doesn't, and which of your pages get cited.
You'll need a spreadsheet program you already use (Google Sheets works fine), a way to open each AI platform you care about, and about an hour for the first round.
Keep in mind that each answer you collect is one observation from one platform, for one question, on one day. It's useful, but it isn't proof that your brand always shows up or never does, so the method below compares like with like.
Step 1: Choose the questions buyers actually ask
Start by writing down a small, steady set of questions, which we'll call your prompt panel. It's the manual version of what AI brand monitoring tools do on a schedule. Pull them from real buyer needs: the problem your customers are trying to solve, the kinds of solutions they compare, the criteria they use to judge options, and a few questions that name your brand directly. Include plain, non-branded questions as well as branded ones, because the non-branded ones are where you find out whether AI brings you up on its own.
Keep each question specific enough to get a useful answer, but don't word it so that it assumes your brand should be in the answer. "Why is our tool the best for agencies?" tells the AI what to say, so ask the open version instead.
A good starting size is 5 to 10 prompts. That number is a workload suggestion and not a rule from any research, so pick what your team can keep up with. Give each prompt a stable ID such as P01, save its exact wording, and add a category like awareness, consideration, or decision, plus a short note about the buyer question it stands for. If you want help thinking about what kinds of questions to include, the guide on the four types of prompts worth tracking is a good companion.
You're done with this step when every prompt has an ID, exact text, and a category, and the list covers the questions your marketing team really wants to watch.
Common mistake: changing the wording between runs and then blaming the change in the answer on a change in visibility. If you have to reword a prompt, keep the old ID and old wording in your log, give the new wording a new ID, and note the date you switched.
Step 2: Set the conditions you'll compare under
Next, decide which platforms matter for your business and write down the conditions you'll use each time. Record the platform or product surface, the country or region, the language, the date, and whether you were signed in or had personalization turned on. Try to keep these the same from run to run. If you test from one country, label it that way, and don't assume the results hold in every market.
For each platform, use the same mode or search setting each time. When a fresh or temporary conversation is available, use it. When it isn't, open a new thread for every prompt, skip follow-up questions that add context, and write down any profile details or custom instructions that stay switched on. A temporary chat doesn't remove every source of personalization.
Here is what to watch for on the platforms most teams start with.
ChatGPT. Web search is available on the free plan and several others, and the help documentation says people who aren't signed in can use it too. To start a search yourself, open a conversation, choose View all tools, and then pick the search tool. ChatGPT can also decide to search on its own, so note whether a given answer used search. If memory is on, it can be used when ChatGPT rewrites your query. Temporary Chat gives you a choice between Personalized and Unpersonalized, and the Unpersonalized option doesn't use memory, custom instructions, or plugins. The ChatGPT search help page explains how the search tool works. If ChatGPT is your main concern, the guide on monitoring product-level mentions in ChatGPT shows what to watch for once you're past the basics, and the walkthrough on how to check your ChatGPT visibility manually goes deeper on that one platform.
Gemini. A Temporary Chat works for a one-off conversation and isn't used to personalize Gemini, according to Google's announcement. Google's Privacy Hub says temporary chats are kept with the account for 72 hours, so don't describe them as erased right away. On managed Google Workspace accounts, controls can depend on your administrator.
Perplexity. If you also want to cover Perplexity, Grok, Meta AI and DeepSeek, the same log works. A session keeps a running conversation, and follow-up questions build on earlier answers. So open a separate new thread for each independent prompt if you want to compare prompts and not hold a conversation. Your account settings can include things like occupation, company name, and custom instructions, so write those down or neutralize them if you can. Perplexity's page on what a thread is explains how sessions work.
Google AI Mode. It can give a full answer with links, but sometimes it returns only web links when it isn't confident that an AI answer would be helpful. Log that as "no AI answer / links only." It isn't the same as your brand being absent. AI Mode can also use your past searches and connected Google apps when Personal Intelligence is on, so note whether it was active.
You're done when you have a short written protocol that covers platforms, region and language, personalization, and how to record a missing answer.
Common mistake: switching account settings, countries, languages, or chat context without writing it down. Any of those can change what you see, and later you won't know why.
Step 3: Build a spreadsheet that keeps the evidence
Set up one row for each completed prompt, platform, and run. Here is a layout that works.
| Column | What to record |
|---|---|
| Run date | The date you submitted the prompt |
| Cycle | A steady week or period label, entered the same way each time |
| Platform | ChatGPT, Gemini, Perplexity, Google AI Mode, and so on |
| Prompt ID | Your stable ID, such as P01 |
| Prompt text | The exact wording you submitted |
| Prompt category | Buyer stage or question type |
| Run number | Which attempt this is, if you repeat a prompt in one cycle |
| Brand mentioned | 1 if the answer names your brand, otherwise 0 |
| Own site cited | 1 if a source link points to your own site, otherwise 0 |
| Competitor named | Names of competitors that appear in the answer |
| Competitor site cited | Competitor URLs or domains shown as sources, if any |
| Sources shown | The exact pages or domains cited, or "none shown" |
| Description accuracy | Accurate, incomplete, incorrect, or not stated |
| Sentiment | Positive, neutral, negative, mixed, or not stated, if you find it useful |
| Status | Complete, no AI answer, error, or not run |
| Evidence note | A short excerpt or the file name of the saved answer or screenshot |
Use dropdowns for any column with repeatable labels, so that "Complete," "complete," and "Done" don't turn into three separate categories. Google Sheets can create an in-cell dropdown list and reject anything that doesn't match it. Keep "not run" separate from a completed answer where your brand wasn't mentioned.
Keep the full answers and screenshots in an evidence folder your team can open, because the sheet is an index and can't replace the original answer. Put the date and the exact prompt next to any excerpt, so someone else can see exactly what was checked.
You're done when a teammate can look at any row and tell the exact prompt, the platform, the conditions, what the answer said, and where the proof is saved.
Step 4: Run the panel on a fixed routine
Pick a cadence you can actually keep. A weekly run is a sensible default for most teams, though it's an editorial suggestion and not a scientific benchmark. Use the same prompt panel and the same platforms every cycle. If the full set is too much work, cut the number of prompts or platforms for the whole panel, and don't quietly skip some of them. When a check fails or gets skipped, record it as skipped or failed.
For each prompt and platform pair, do this in order:
- Open a fresh or unpersonalized conversation if the platform offers one.
- Paste the exact saved prompt, with no follow-up questions that add new context.
- Save the full answer and any sources the platform shows.
- Record whether the brand is named, and record separately whether your own site is cited.
- Mark the row complete, or choose the right exception status.
This is also how you track ChatGPT mentions manually, and it works the same way on the other platforms.
For your most important questions, think about running the same prompt more than once in a cycle and logging each run on its own row. Repeats help you see whether an answer holds steady. A few manual runs still won't give you a statistically reliable estimate, so if you can't repeat everything, repeat the prompts that matter most and say in your notes that you did.
You're done when every scheduled prompt and platform pair has either a recorded observation or an explicit reason it wasn't completed.
Pro tip: Keep the wording and conditions fixed for your main trend panel, and put your curious, one-off questions in a separate tab. Otherwise the exploratory tests will blur the time series you're trying to build.
If you're wondering how much answers can move from one run to the next, a 2026 study gives some numbers, and we'll get to them in Step 7.
Step 5: Code mentions, citations, and answer quality separately
Choose clear coding rules and apply them the same way every time.
- Brand mention: the answer text names your brand. If only a URL shows up, record it as a citation, and don't quietly count every linked domain as a spoken mention.
- Own-site citation: the answer shows a source link to a page on your website. Write down the actual URL or domain when you can see it. This is a different thing from being named in the text.
- Competitor mention: record each competitor you monitor that is named in the answer. Don't count a competitor as named only because its URL appears, unless your monitoring definition says linked sources count and you apply that rule everywhere.
- Description quality: code separately whether the answer describes you accurately, incompletely, incorrectly, or not at all. Keep a short excerpt that supports any "incorrect" or "incomplete" rating.
- Sentiment: this one is optional and is a judgment call. If you include it, use a few simple categories and keep the exact text for review, so a subjective label doesn't get passed around as if the platform produced it.
Gemini needs one extra note. It can show sources or related content for parts of an answer, and it also has a Double-check response feature. Those are two different things. Double-check uses Google Search to find content that is likely similar to or different from the statements in the answer. A green highlight means Search found likely similar content, but the linked result isn't necessarily what Gemini used to write its response. Orange means likely different content, or nothing relevant found. So don't count Double-check links as the answer's original citations.
You're done when two people on your team could take the same answer, apply your rules, and reach the same coding decision.
Common mistake: treating "the answer named us" and "the answer cited our page" as a single measure. An answer can name you without linking to your site, link to a page without naming you in the text, do both, or do neither, so keep those as independent columns.
Step 6: Calculate rates with a steady denominator
Now you can turn the rows into rates. Use completed answers as the denominator, and report exceptions separately. For a given platform and cycle:
- Mention rate is the number of completed observations that named your brand, divided by the number of completed observations.
- Own-site citation rate is the number of completed observations that cited your site, divided by the number of completed observations.
- Competitor mention rate is the number of completed observations that named a specific competitor, divided by the number of completed observations.
In Google Sheets, COUNTIFS handles the counting. Say the cycle is in column B, the platform is in column C, the brand-mentioned flag is in column H, and the status is in column P. If the cycle you want is in A2 and the platform is in B1, a mention-rate formula looks like this:
=COUNTIFS($B:$B,$A2,$C:$C,B$1,$H:$H,1)/COUNTIFS($B:$B,$A2,$C:$C,B$1,$P:$P,"Complete")
For own-site citations, swap in the own-site-cited column where H appears. Make sure there's at least one completed observation before you divide, or the formula will return an error. Show the numerator and denominator next to the percentage, for example "7 of 20," so small samples are easy to spot and nobody mistakes a rate for broad market coverage.
A pivot table can summarize the data by platform, cycle, prompt category, or competitor. Keep the raw sheet untouched and build the rates on a separate summary tab. If the panel or completion rate changed between periods, say so before comparing.
You're done when your report shows the counts and denominators, the rate for each platform, and how many observations were missed or invalid.
Common mistake: counting only the answers that mention a brand in the denominator. That makes the rate look higher than it is. Use every completed observation for the platform and cycle you're reporting.
Step 7: Review the pattern, then look into notable changes
Compare the same prompts, on the same platforms, under the same conditions, across cycles. Look at the change by prompt as well as the overall platform rate, because a total can hide something important. For example, you might show up for branded questions and not for neutral category questions, and the overall number would blur that. Read the saved answers and sources before you suggest a cause for anything.
A review sequence that works well:
- Check the data quality first: missing runs, reworded prompts, a changed platform mode, a different region or language, or coding that wasn't consistent.
- Compare mention rate and own-site citation rate separately, with their counts and denominators.
- Find which prompt and platform observations changed.
- Read the earlier and current answers side by side and note what changed: whether your brand appears, how it's described, which page was cited, or whether there was no answer at all.
- Write down your explanation as a hypothesis. Don't say a content change caused a visibility change unless the evidence really backs that up.
When a result matters, run the exact prompt again under the recorded conditions and keep both observations. A repeat can show whether the result comes back, but it won't tell you why it happened.
A 2026 study called Don't Measure Once: Measuring Visibility in AI Search looked at ChatGPT, Gemini, Google AI Mode, and Perplexity. It ran each of eight prompts per campaign ten times, across four campaigns and four engines, which made 128 engine and prompt groups. The overlap between the sources shown in pairs of runs averaged 32 to 43 percent, and the overlap in brands ranged from 33 percent in Sporting Goods to 48 percent in Consumer Electronics. Those numbers belong to that study's sample and method, and they aren't a target for your program.
The study has limits you should know about too. Its data came from servers in Switzerland, so region and language limit how far the results travel, and its brand detection used substring matching against a fixed word list, which can miss synonyms and match by accident. It also reports that ChatGPT web search was only active for some queries, and 57.8 percent of its runs showed zero citations. That's a good reason not to mix up "no citation shown" with "the brand wasn't mentioned."
The takeaway is a modest one. Asking the exact same question can produce different answers and different source lists, so a single manual check is a weak basis for claiming a trend. The weekly routine here is an operating suggestion and not a validated sample size.
You're done when every movement you report can be traced to the rows and saved answers behind it, with your data-quality notes attached.
Step 8: Turn what you found into a next action
Use the log to decide what to look at next. You might find a prompt where competitors are named and your brand is missing, a mention where the description of your product is wrong, or a case where a competitor's page was cited when you expected yours to be useful. Treat each of these as a lead to investigate, and open the answer and the cited page before you choose an action. Once you're ready to act on a pattern, setting up alerts for changes in how AI mentions your brand is a natural follow-on, and deciding which changes are meaningful will keep you from reacting to noise.
Then make the routine stick. Set a recurring calendar reminder, give the check to one named owner, and keep a prompt-change log. Add a notes field for changes to your website or your protocol, since those help you read later results. The routine is working when it gives you a steady, reviewable history. A pile of charts doesn't tell you that.
If the spreadsheet starts to feel heavy, and you want to see how it fits among the broader AI visibility tools, DeepSmith's AI Visibility feature is the automated counterpart to this manual routine. It queries AI engines on a schedule, captures the full answers, and tracks mentions and citations separately, with platform and page-level views of what's being cited. You don't need it for any step above, and you'll still want to read the answers yourself, but it can save the clicking once the panel grows. For a look at the options in that category, the guide to AI brand monitoring tools compares them, and free AI brand monitoring tools lists the ones that cost nothing.
What to do next
Run your next scheduled panel, read the saved answers, and keep the prompts and the conditions the same so the comparison stays fair. If you'd like to see an automated version of this routine, you can start a DeepSmith free trial and compare it against your spreadsheet.
Related guides
- What Your CMS Needs to Support for AEO: Evaluate whether your CMS supports AEO by checking structured data and schema control, clean semantic markup, fast updates, redirect and metadata handling, and what to ask vendors before.
- Enterprise AI Visibility Platform Requirements: Security, Governance and Scale: Build a requirements checklist for an AI visibility platform covering SSO, role permissions, audit logs, API access, data residency and scale limits, so you can score vendors on.
- How ChatGPT Visibility Tracking Works: Search Mode, Citations and Memory: Understand how tracking tools measure ChatGPT: why search mode and plain answers differ, how citations are captured, what memory and personalisation do to results, and why samples vary.
- How to Improve Your Visibility in Perplexity: What You Can Influence: Identify what you can influence in Perplexity: source quality and freshness, clear answer passages, authoritative third party mentions and crawl access, and which factors sit outside your control.
- How to Track Your Brand Visibility in Gemini: Set up Gemini tracking: choose prompts, run repeat samples, record mentions, sentiment and cited sources, and read results by topic so you can see where Gemini describes your.


