If a rival keeps showing up when people ask ChatGPT about your category, it's natural to want a clearer picture than a few screenshots. This guide is for marketing leads who want a steady way to track the competitor mentions ChatGPT returns, along with other AI engines, and to see how those mentions compare with their own. By the end you'll have a simple process that records when a rival appears, what role the answer gives them, whether their pages are cited, and how that stacks up against your brand. It's one part of wider AI brand monitoring, and it fits into a routine you can repeat every few weeks.
This guide covers the monitoring itself. It doesn't cover choosing which competitors to track, and it doesn't cover how AI answers describe your whole category. Here we assume your competitor list is already settled.
You'll need a spreadsheet or a monitoring tool, a short list of buyer questions, and access to the engines you care about.
Step 1: Fix the questions you will monitor
Competitor monitoring AI answers only works if the questions stay the same from one round to the next. If you change the wording, you can't tell whether the answer moved because of a competitor or because of your edit.
Start by picking a defined set of buyer questions that matter to your brand, and write down the exact wording of each one. Give the set a version name, such as "v1", so you can say which set a result came from. For each prompt, record these fields:
- Prompt ID and exact prompt text
- Prompt-set version
- Intent (for example, comparing options or looking for a recommendation)
- Persona or qualifier, if the question names one
- Date added, and whether it's still in the tracked set
Write questions the way your buyers would ask them, and don't write them to push the answer toward a result you want. If you'd like a head start, the guide on how to build a repeatable prompt set walks through it in more detail. DeepSmith's Discover Prompts can also suggest a starting list from your product, persona, and buyer-stage context. You choose which suggestions to track, and you can write your own too. It's a way to build a test set, and it won't cover every real search.
You're done with this step when there's a named, versioned list of prompts with the exact wording and intent recorded.
Common mistake: rewriting a prompt between reporting periods and then crediting the change in the answer to a competitor.
Step 2: Make collection conditions as comparable as possible
Next, decide how each run will be collected, and write it down. For every run, record the engine and, where you can see it, the model or mode. Also record the collection time, the region or language, and any account or personalization setting that might matter.
When you sample by hand, ask each prompt in a fresh conversation. Don't carry over a follow-up from the previous question, and don't edit an answer with follow-ups and then treat that conversation as the original answer.
Account settings can change what you see. ChatGPT search may rewrite a query using saved memories. A ChatGPT Temporary Chat has an Unpersonalized option that doesn't use memory, custom instructions, or plugins, but you have to choose it before you send the first message, because it can't be changed afterward. Google says AI Mode can use your Search history to personalize responses, and you can opt out in the personalized-recommendations settings. Where a clean test is possible, use the same setting every time. Where the platform doesn't offer a comparable control, write the limitation down rather than assuming the results are neutral.
Engines also behave differently. ChatGPT search responses may show citations and a Sources panel. Perplexity describes its answers as grounded in web search, with numbered source citations. Gemini's Double-check uses Google Search to find content that's likely similar to or different from what it generated, and those results aren't necessarily the sources Gemini used. In Google AI Mode, web links can show up instead of an AI response. Whatever the interface returns, record it, including the cases where there was no answer or the run failed.
DeepSmith AI Visibility runs your tracked prompts on a schedule you set, then captures and analyzes the answers. DeepSmith covers ten engines in total: ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews, Google AI Mode, Grok, Meta AI, Microsoft Copilot, and DeepSeek. What you get depends on your plan. Pro tracks ChatGPT, Grow adds Perplexity, Scale adds Gemini, and Enterprise covers all ten. So if you mostly care about competitor mentions ChatGPT shows and competitors ranking in ChatGPT, every plan covers that, and you can add engines as you move up.
You're done when each comparison shows the same prompt, the same engine or mode, and the conditions you documented.
Pro tip: use a clean conversation and the same personalization setting each time, but don't claim this makes the test identical to what every real user sees. It only makes your own rounds comparable to each other.
Step 3: Save the complete answer and its run metadata
A screenshot of a brand name isn't enough to work from later. Keep the full answer text, along with the details of the run. For every prompt, engine, and run, preserve:
- The exact prompt and its version
- Intent and any persona or qualifier
- Engine, and the model or mode when you can see it
- Run timestamp, ideally in UTC
- Full answer text
- Source URLs or citations the engine showed
- Whether your brand was mentioned, where it appeared, and its sentiment if you track that
- Whether your domain was cited, and which page
- Which competitors were mentioned or cited in the same answer
- Operator notes, such as an outage, a mode change, or a model release
One row per answer-run is the easiest setup to audit. Keep failed attempts separate from completed answers. Note why the run failed and leave it out of your answer counts. If you quietly count a failed run as an answer with no mention, your rates will drift down for a reason that has nothing to do with the market.
In DeepSmith, tracked prompts keep their full answer history and show prompt-level mention and citation rates. The visibility data is also available through read-only exports and API access, which covers competitors, cited pages, and the answers themselves. It's read-only, so it reads your workspace data and doesn't create or publish anything.
You're done when another person could recover the original answer and know when and under what conditions it was collected.
Step 4: Label whether each brand appears, then describe the context
Start with a plain yes or no. Is the brand named anywhere in the answer? Count a brand once per response, even if the name appears several times, and note where it appears so a reviewer can check your label later.
Then compare your brand with one competitor at a time. Every answer gets one of four labels for that pair.
| Your brand named? | Competitor named? | Pairwise label | What it means |
|---|---|---|---|
| Yes | No | Your brand only | You appeared and the competitor didn't |
| No | Yes | Competitor only | The competitor appeared and you didn't |
| Yes | Yes | Both | Both brands appeared in the same answer |
| No | No | Neither | Neither brand appeared |
These labels are pairwise. One answer can be "both" for one rival and "competitor only" for another, so don't try to name a single winner when an answer lists several brands.
After the presence label, add a short context tag based on the wording. Useful ones are recommended or shortlisted, compared, named as an alternative, mentioned in passing, and described negatively. These are practical annotation labels and not an industry standard. Keep an excerpt that supports each tag, and use "unclear" when the wording doesn't justify a confident call. If you also capture sentiment, keep it in its own column so it doesn't blur with presence and context.
You're done when every valid answer has an auditable yes or no for your brand and for each competitor you compare against, plus a context note where it helps.
Common mistake: treating a passing mention, an explicit recommendation, and a competitor's cited URL as the same kind of win. They are three different observations.
Step 5: Record citations separately from mentions
A mention answers one question: did the response name the brand? A citation answers another: did the answer point to a page as a source? Record them separately for your domain and for each competitor's domain, and save the exact page URL along with the prompt that produced it. A rival can be named without being cited, cited without being named in the text, or both.
When you can, open the source and check that the page is the one shown and that it supports the claim next to it. OpenAI's own help guidance says search results and citations from ChatGPT may be incomplete, outdated, or incorrect, so a citation shouldn't be read as proof or as an endorsement. With Gemini, keep the sources shown with an answer apart from links produced by the separate Double-check feature. Don't call Double-check links original citations unless the answer itself identifies them that way.
DeepSmith's AI Visibility tracks which of your pages earn citations, the prompts behind them, and the sources engines use most. It also shows who wins citations for your prompts, on which exact pages, and how each competitor does by engine. Separately, competitor sitemaps are checked for new pages every day. That's a different thing from detecting every answer as it happens, so don't mix the two when you describe it to your team.
You're done when mention results and citation results can be reported separately, with the cited pages kept for review.
Step 6: Calculate rates with the counts visible
For one prompt, one engine, and one period, let N be the number of valid, completed answers. Show the counts next to the percentages every time.
- Brand mention rate: answers naming your brand, divided by N, times 100
- Competitor mention rate: answers naming that competitor, divided by N, times 100
- Brand-only, competitor-only, both, and neither rates: the count in each pairwise label, divided by N, times 100
- Citation rate for a domain: answers that cite that domain, divided by N, times 100, as long as you use that definition every time
- Percentage-point gap: the competitor's mention rate minus yours, for the same prompt, engine, and period
For each pair, the four label counts should add up to N. Your mention count is your "brand only" count plus your "both" count, and the competitor's is "competitor only" plus "both". Those two checks catch most tally mistakes.
Here's an example with made-up numbers, just to show the arithmetic, and not data from any DeepSmith customer. Say you collected 40 valid answers for one prompt on one engine. Your brand appeared in 12, a 30% mention rate. Competitor A appeared in 22, a 55% rate, so the gap is 25 percentage points. In the pairwise labels, 6 answers named only your brand, 16 named only Competitor A, 6 named both, and 12 named neither. That's 40 in total. If your pages were cited in 8 answers and Competitor A's in 10, the citation rates would be 20% and 25%.
Share of voice needs a little extra care, because it isn't one formula across vendors. Ahrefs, for example, defines AI share of voice as the responses that mention your brand divided by the responses that mention any tracked brand, times 100. If an answer can name several tracked brands, the individual numerators overlap, so the percentages won't add up to 100. Write down your denominator before you report it. DeepSmith reports share of voice too, and if you want to work through the math, the guide to calculating AI share of voice covers it.
You're done when every rate comes with a count, a denominator, a prompt, an engine, and a period.
Step 7: Compare like with like and read the underlying answers
Review results by prompt and engine first, and only then roll them up across a defined set. Compare the same prompt version, the same engine or mode, the same personalization conditions, and the same collection window. If you combine engines or prompt types into one number, say how you combined them. Always show the valid-answer counts and any missing or failed runs next to the percentages.
Use the rates to spot a pattern, then read the actual answers. Is the competitor named again and again for the same question? Is your brand absent, or are both brands there? Is the rival recommended, just listed, or cited as a source? Does one of their pages keep coming back? And is it happening on one engine, one prompt, or one period? Reading prompt-level win and loss data against competitors is a good way to see this question by question. If a rival you didn't expect keeps appearing, it may be time to revisit choosing which competitors to track.
Try not to treat the order of brands in an answer as a stable ranking. In one limited SparkToro study, 142 human-written headphone prompts produced 994 AI responses across ChatGPT, Claude, and Google Search AI Overviews, with AI Mode used when an Overview didn't appear. The researchers reported that ChatGPT and Google AI returned the same brand list in fewer than 1% of repeated runs, and they were clear about the study's limits and the need for much larger samples. That's evidence that outputs can vary, and it isn't a benchmark for every category. If you're curious why answers change from run to run, there's a guide on that. For tracking, it means that position in a single answer is context, and what you want is repeated inclusion across runs and periods. A DeepSmith prompt-set guide suggests running each prompt at least three times per engine per collection window and using the majority outcome to smooth out one-off randomness. That's a sensible baseline to start from, and it doesn't guarantee statistical confidence, so collect more runs if you need a steadier signal.
You're done when you can trace any reported difference back to comparable answer records and not just one screenshot or a blended dashboard number.
Common mistake: reading a small percentage gap as a real competitive shift without checking the counts, or without checking whether the prompts and settings changed in between.
Step 8: Turn a repeated gap into a testable next action
Use what you've found to decide what to look at next. Not every competitor mention calls for a new article. If the same rival keeps being named for the same prompt, look at the answer context, the specific pages cited, and the rival's published pages on that topic. Compare those with what your own site offers. Then note whether the gap is about being named, being cited, or both.
Pick one content or page change that speaks to an evidence-backed gap, make it, and run the same prompt set again under the same recorded conditions in a later window. Changing several things at once makes it hard to know what helped.
DeepSmith helps with the collection and reading side of this. AI Visibility tracks your prompts, answers, mention and citation rates, the competitor leaderboard, the competitor pages that get cited, and the pages on your site that earn citations. The schedule is whatever you've configured for your workspace. If you're still comparing software, the roundup of the best AI visibility tools lays out the options. If rivals seem to shape the whole conversation about your category, the guide on category-level narratives covers that side. If you want to be told when something shifts, the guide on setting up alerts for changes in how AI mentions your brand covers a routine that doesn't depend on any one tool.
You're done when you've recorded the suspected gap, the action you took, and the date or window for the retest.
What to do next
Look at your next collection window and pick one prompt where a rival shows up more than you do. Read those answers, write down what you think explains it, make one change, and plan the retest. If you'd like the scheduled runs, answer history, and competitor citation views in one place, you can start a DeepSmith free trial and see your own prompts with real data. It runs for 7 days.
Related guides
- How to Set Up AI Brand Monitoring in One Afternoon: Get a working monitoring setup in a few hours with a starter prompt list and a first review.
- What a Good AI Brand Monitoring Dashboard Shows and Who Should Read It: Decide what a monitoring dashboard should show, such as mention rate, sentiment, sources and competitor share, and who each view is for, from executives to analysts.
- AI Brand Monitoring vs Social Listening: What Each One Covers: See what AI brand monitoring and social listening each cover, such as AI answers versus public posts, where their data and goals overlap, and when a team needs.
- AI Reputation Management: What It Is and Who Should Own It: Define AI reputation management and decide which team should own it.
- How AI Sentiment Analysis Works and Where It Breaks: Understand how sentiment is scored in AI answers and where it fails, such as hedged language, mixed claims and sarcasm, so you can read sentiment scores with appropriate.


