Key takeaways
- Tracking competitors in AI search means measuring four separate things: visibility, share of voice, position, and how each brand is described (sentiment and narrative).
- Build the benchmark from real buyer questions, not platform defaults. Vendor guidance suggests 25 prompts as a directional minimum and 60 to 120 for a strong benchmark, each run at least three times.
- Perception is shaped by sources. The most useful output of any benchmark is the list of third-party pages AI cites when it talks about your competitors and not you.
- Annotate model changes and track by platform rather than pinned model version. Weekly core metrics, monthly gap analysis and quarterly reporting is a workable rhythm.
- A benchmark that does not change a content, PR or product decision is just a dashboard.
Why perception benchmarking is different from rank tracking
Google rank tracking tells you where a page sits in a list. It says nothing about whether ChatGPT recommends your competitor, whether Gemini repeats an outdated claim about your pricing, or whether Perplexity describes a rival as "best for enterprise" and you as "a cheaper alternative."
That last part is what I mean by perception. A mention is not a mention. "Powerful but difficult to implement" and "affordable for small teams" both count as appearing in the answer, but they send buyers in different directions.
Typeform's guide to competitive benchmarking splits metrics into performance benchmarks and perception benchmarks: awareness, preference, trust, value for money. Its point carries over directly to AI: behavioral data tells you someone chose another brand, while perception data tells you why. In AI search, the model's answer is a new kind of perception data, and you can collect it at scale without running a survey.

It is still worth running both. AI answers show what the models believe and repeat. Customer surveys, like the competitive perception surveys described in Koji's competitive intelligence survey guide, show what buyers believe. When the two disagree, you have found something worth fixing.
The four metrics worth separating
Most tools compress everything into one visibility score. That is convenient and slightly dangerous, because a blended number hides where the problem actually sits. Keep these four apart.
Visibility
The share of tracked answers that mention the brand at all. If you appear in 20 of 100 answers, your visibility rate is 20% for that prompt set and period. It only means something relative to the prompt set you chose.
Share of voice
Your mentions divided by all tracked brand mentions in the category. This is the competitive number. Better implementations weight by answer position and recommendation context, since being the first named pick is not the same as appearing once in a footnote.
Position
Where a brand lands in the order of brands inside an answer. A brand that is always mentioned fourth has a prominence problem, not a coverage problem.
Sentiment and narrative
How the answer describes each brand. This is the perception layer. Look for the repeated claims: pricing, ease of use, support, reliability, who the product is for. Repeated claims are what buyers absorb.
Read together, these locate the problem. Low share of voice is a coverage issue. Late position is a prominence issue. Weak sentiment is a reputation issue. Mentions without citations to your own site usually point to a content or crawlability issue. Promptwatch's docs have clear definitions of Visibility Score and Share of Voice if you want a reference for how these can be calculated.
Step 1: Define who counts as a competitor
Start with three to five direct competitors, up to ten if your category is crowded. Then add the ones AI considers competitors, which are not always the ones your sales team lists. Marketplaces, review sites, open-source alternatives and "just use a spreadsheet" answers often show up in the same responses as you.
Two practical notes:
- Write down aliases and product names (including common misspellings) before you start, so counts are consistent.
- If your brand name is also a common word, write explicit rules for what counts as a mention. Otherwise your numbers will be noise.
Step 2: Build a prompt set from real buyer language
This is where most benchmarks go wrong. Practical Ecommerce noted in July 2026 that scores depend entirely on which prompts you run, and that some vendors inflate scores by putting the brand name in the prompt. Do not use platform default prompts as your whole set.
Good sources for prompts:
- Sales call transcripts and objections
- Support tickets and onboarding questions
- Google Search Console queries and People Also Ask
- Reddit and community threads where buyers ask for recommendations
- Your own win/loss notes
Vendor guidance from Hikoo and get-ryze suggests a minimum of 25 prompts for a directional read (10 category discovery, 10 comparison, 5 switching or alternatives). A stronger benchmark runs 60 to 120 prompts, split roughly 30% discovery, 40% consideration and comparison, 20% evaluation and 10% switching. Treat those ratios as a starting point, since they come from vendor blogs rather than primary research.

Three prompt types to include
| Prompt type | Example | What it reveals |
|---|---|---|
| Organic category | best project management software for small teams | Who AI recommends when no brand is named |
| Brand-specific | is [Competitor] good for small teams | How AI describes one brand in depth |
| Competitor comparison | [You] vs [Competitor] | Who wins head to head, and on what criteria |
The distinction matters because the citation behavior differs. In Promptwatch Data's analysis of unique domains per ChatGPT response (Sep 28 to Oct 4, 2026), 52.5% of organic prompts cited 10 or more domains. For brand-specific prompts the most common count was 5 domains, and only 21.3% cited 10 or more. So an answer about a single competitor leans on a narrower set of sources, which makes those sources easier to identify and more influential.
Step 3: Run it on the engines your buyers use
At minimum cover ChatGPT, Google AI Overviews and Perplexity. Add Gemini, Claude, Copilot and Google AI Mode if your buyers use them. Do not assume they behave alike:
- Per Promptwatch Data on average sources per response, ChatGPT cites about 5 sources per answer, Google AI Overviews about 10, and Perplexity almost exactly 10 every day. Copilot's average swung from under 2 to nearly 17 within weeks, so judge it on monthly trends.
- Citation behavior shifts with model updates. Promptwatch's ChatGPT citation drop report shows average citations per response falling across GPT-5.3, GPT-5.4 and GPT-5-Mini around the March 4, 2026 rollout.
Track by platform, not by a pinned model version, and annotate model releases on your trend lines. Otherwise you will mistake a model change for a competitor move.
One more reason to care about the surface you measure: user-facing answers can differ from API output. Tools that collect from the real interfaces will usually show you what buyers see; API-only tools show you what the model says without the product wrapper.
Repeat every prompt
AI answers vary run to run. Run each prompt at least three times before you calculate anything, and keep the raw responses so you can audit a number later. A tool that only stores a score is hard to trust when something changes.
Step 4: Score perception, not just presence
Once responses are collected, extract the claims. For each competitor, note how often the models say things like:
- cheaper, more expensive, or "best value"
- easiest to use, or hard to implement
- best for enterprise, best for small teams
- strong support, poor support
- missing features or integrations
Build a simple scorecard: one row per attribute, one column per competitor, with a consistent scale (for example, -2 to +2 for each theme, or counts of positive and negative claims). Typeform's advice for benchmarking applies here too: use consistent definitions and measurement periods, or the comparisons are meaningless.
Watch for two patterns in particular:
- A competitor owns an attribute you want ("best for security") with several models repeating it.
- A claim about you is wrong. Inaccurate claims usually trace back to a specific outdated page, which is fixable.
Some platforms automate this. Promptwatch's Brand Narrative pulls individual claims out of tracked responses, groups repeated ones, and scores themes like Pricing, Features and Support against competitors. Peec AI, Waikay and LLM Pulse also lean toward the perception and sentiment side.
Step 5: Trace the sources behind each claim
This step turns a scoreboard into a strategy. AI engines do not invent a competitor's reputation from nothing. They pull it from pages: the competitor's own product pages, review sites, listicles, forum threads, documentation and news.
For each competitor, list the cited domains and URLs across your prompt set, then compare against yours. You are looking for citation gaps: sources where a competitor is cited and you are not. Typical findings:
- A "best X tools" roundup that lists three competitors and leaves you out
- A review-site profile (G2, Capterra, Trustpilot) that is thin or stale
- A subreddit or YouTube channel AI keeps citing for your category
- A documentation hub that explains the category better than your site does
Vertical data helps you seed the list. Promptwatch Data on software prompts (Sep 28 to Oct 4, 2026) shows GitHub, Microsoft's docs and G2 leading ChatGPT citations, with documentation and help centers earning the largest share. If you sell software, those are the kinds of pages that shape competitor narratives.
Also be careful about assumptions on social sources. In Promptwatch's social media citation data for Sep 28 to Oct 4, 2026, ChatGPT drew just 0.46% of citations from social platforms while Google AI Overviews drew 12.2%. A Reddit-heavy strategy might move AI Overviews and do very little for ChatGPT. Check the per-engine picture before you invest.
For how long a placement lasts, see Promptwatch's content half-life report: of pages ChatGPT cited on July 6, 2026, landing pages and docs stayed cited longest while news lost half its pages within 36 days. That is a hint on where to put effort if you want durable perception wins.
Step 6: Look at the query fan-outs
AI engines rarely answer from your prompt alone. They generate their own searches. Promptwatch's query fanout data puts average searches per response at 2.15 in early December 2025 and about 1.84 by early March 2026, with a single prompt sometimes triggering 3 to 8 or more searches. The numbers moved around later in the dataset, so treat the averages as a trend and not a constant.
The more interesting finding for competitor tracking: on August 8, 2026 ChatGPT began using the site: operator at scale, jumping from about 0.4% to about 17% of fanout queries. That means ChatGPT is sometimes searching specific sites directly, which makes it more important to see which sources are being queried for your competitors.
If you want to see fan-outs for a single prompt without a paid tool, the ChatGPT Query Fan-Out Generator is a free starting point.
Step 7: Choose a tool (or combine a few)
You can run a pilot by hand with a spreadsheet and a few dozen prompts. It works for a month. After that, the repeated runs, multiple engines and competitor extraction get tedious fast.
There are dozens of options. A few that come up for competitor perception work:
| Tool | Strength for competitor tracking | Best for |
|---|---|---|
| Promptwatch | Monitoring plus action: Brand Narrative, citation tracing, crawler logs, content agents, Reddit/YouTube/LinkedIn citation reports | Teams that want to close the loop from benchmark to published fix |
| Profound | Enterprise analytics depth | Large brands with analyst resources |
| Peec AI | Competitor benchmarking, alerts, exports, prompt-scale tracking | Agencies and lean teams |
| Otterly.AI | Affordable citation and mention monitoring | Smaller budgets |
| Semrush AI Visibility Toolkit | AI tracking inside an SEO suite | Teams already on Semrush |
| Ahrefs Brand Radar | Brand tracking alongside backlink data | SEO-centered teams |
| Waikay | AI brand perception focus | Perception-first monitoring |
Promptwatch is the one I would look at first if the goal is to act on the benchmark rather than just watch it. It collects answers from the real interfaces of ChatGPT, Gemini, Perplexity, Google AI Overviews, AI Mode, Copilot and others in each country and language, scores Visibility, Share of Voice, Position and Sentiment, shows which pages and third-party sources AI cites, and then hands the gaps to an Action Board and Content Agent. Its Fact Check feature lets you flag an inaccurate claim and find which page on your site is teaching AI the wrong thing.

Plans start at $95/month (Essential), with a 7-day free trial on monthly billing. Confirm current pricing at promptwatch.com/pricing. If you are only exploring, the free AI visibility report gives a no-login scan of how ChatGPT, Gemini and Perplexity talk about your brand. For head-to-head details see Promptwatch vs Profound and Promptwatch vs Peec.
Here are a few other tools worth a look depending on your needs:

For a wider shortlist, the directory at bestgeosoftware.com lists GEO platforms side by side.
Questions to ask any vendor
- Which engines are measured, and from the real interface or an API?
- How is the competitor denominator defined, and does position affect the score?
- Are raw responses retained so a score change can be audited?
- Can I separate brand, category, competitor and sentiment prompts?
- Does it trace citations, or only count mentions?
- How does it handle repeated runs, regions and model changes?
Step 8: Set a cadence
HubSpot and Hikoo both suggest roughly the same rhythm, and it matches what I would do:
| Cadence | What to review |
|---|---|
| Weekly | Core metrics: visibility, share of voice, position. Spot sudden competitor jumps |
| Monthly | Full competitive gap analysis, new citation gaps, narrative changes |
| Quarterly | Executive benchmark: trend lines, wins, and priorities for the next quarter |
Competitor share of voice can swing within days when a rival publishes fresh content, earns press or ships schema fixes. Weekly checks catch that before it compounds. A competitor whose share of voice grows around 5% a month is building citation authority that gets harder to displace the longer you wait.
Step 9: Turn gaps into work
Every finding should map to an owner and an action.
| Finding | Likely cause | Action |
|---|---|---|
| Competitor cited, you are not, in roundups | You are absent from listicles | Pitch inclusion, publish your own comparison pages |
| Weak sentiment on one theme | Old reviews, thin documentation | Refresh docs, update review profiles, address the product issue |
| Inaccurate claim about you | Outdated page AI reads | Fix or redirect the page, add a clear fact statement |
| Mentioned but never cited | Content or crawlability problem | Check AI crawler access, improve structure |
| Competitor dominates one persona | You lack content for that buyer | Create persona-specific pages |
On the crawler side, you can check whether AI bots can read your site with the free Agentic Readiness Check. If you want to see which bots actually hit which pages, IP-verified crawler logs separate an access problem from a content problem.
If you would rather hand the program to specialists, an agency like 1001 SEO Media runs GEO and AI search work alongside technical SEO and digital PR, and uses Promptwatch for AI visibility tracking.
A 30-day pilot you can start this week
- Week 1: Pick five competitors, write 30 prompts from sales and support language, and agree on mention rules.
- Week 2: Run every prompt three times across ChatGPT, AI Overviews and Perplexity. Store raw responses.
- Week 3: Score visibility, share of voice, position and sentiment. List the top 20 cited domains per competitor.
- Week 4: Identify five citation gaps and three narrative gaps. Assign each an owner. Rerun the prompts to set a baseline for next month.
Common mistakes
- Using only prompts that include your brand name, which flatters the numbers.
- Judging a trend from one run. Variance is real.
- Ignoring region and language. A monitor pinned to one country will not tell you about another.
- Mixing model changes with competitor moves. Annotate releases.
- Treating a tool's score as truth. These metrics are estimates, and your prompt set decides the answer.
- Collecting charts and doing nothing. Typeform put it well: if a benchmark does not lead to a decision, it is just an interesting number.
Wrapping up
Tracking what AI says about competitors comes down to four steps repeated forever: pick a prompt set that reflects real buyers, measure visibility, share of voice, position and narrative across the engines that matter, trace the sources shaping each claim, and ship fixes. The first benchmark will probably be a little embarrassing. That is the point. Run it again in 30 days and see which of your changes moved a number.


