How to Benchmark Your Brand's AI Citations Against Competitors

A single check can't do this job. When SparkToro had 600 volunteers run 12 prompts 2,961 times, it found less than a 1 in 100 chance that ChatGPT or Google's AI would return the same list of brands twice. Links and names also move apart. In Semrush's June 2026 study, 62% of AI citations were "ghost citations": the page was linked, but the brand behind it was never named.

This guide gives you a fixed 20-prompt matrix, the scoring math, a worked example with three made-up brands, and a table for reading the result. It's a one-off benchmark you can run in an afternoon. For the weekly version, see our guide on how to track competitor rankings in AI search. For baselines over time, see how to benchmark AI search performance.

Why a Competitor Benchmark Needs Its Own Prompt Set

SEO tools taught us to compare rivals by keyword overlap: which terms you both rank for, and who ranks higher. AI answers don't work that way. There is no list of ten results. The engine writes a short answer, names a few brands, and links a handful of pages. So the question changes from "who ranks above me?" to "when a buyer asks, who does the answer recommend first?"

That question depends on how the buyer asks. People rarely type your category name and stop. They describe a problem ("my team keeps missing client deadlines"). They look for a way out of the tool they have ("what can I use instead of Loopcraft?"). Or they compare features ("which tools have client approvals built in?"). A benchmark built only from "best X for Y" prompts tests one of those moments and misses the other two.

Mentions and AI citations are separate wins

An answer can help a brand in two ways. It can name the brand in the text, which is a mention. It can link to the brand's page as a source, which counts toward its AI citations. The two often split. The same Semrush study found that ChatGPT linked to a domain 87% of the time it appeared but named the brand in the text only 20.7% of the time. Gemini did the reverse, naming brands 83.7% of the time and linking 21.4%.

So a rival can lead your benchmark on mentions and trail on AI citations, or the reverse. Log both, and score both. Our glossary entry on AI citations covers the difference in more depth.

Engines disagree, so benchmark each one

The 2026 study "Don't Measure Once" tested ChatGPT, Gemini, Google AI Mode and Perplexity. It advises marketers to "set engine-specific visibility baselines rather than applying a single threshold across all AI search products." Google AI Mode concentrated its AI citations most heavily on a few sources, and Perplexity spread them most evenly. A rival that dominates one engine can be a minor name in another, so every table in this guide splits by engine as well as in total.

The 20-Prompt Rival Matrix

The matrix is a fixed set of 20 prompts, split by the three ways buyers compare products. It's small enough to run by hand and broad enough to show where you win and lose. Here it is for Plannora, a made-up project management tool whose made-up rivals are Loopcraft and Taskwell.

Problem queries (8 prompts)

Write these the way buyers talk in sales calls and support tickets. Name the pain, not the product type. Use four plain versions and four with a constraint, such as team size, budget, industry or a tool the buyer already uses.

• "Our design team keeps losing track of client feedback. What should we use?"
• "What's the easiest way to plan work for a 12-person agency on a small budget?"

Problem queries tell you whether the engine links your brand to the job it does. They are often the hardest prompts to win, because nothing in them points to you.

Alternative searches (6 prompts)

Alternative searches catch buyers at the moment they switch. Split them across your rivals and add one about yourself:

• 3 prompts that ask for alternatives to your biggest rival
• 2 prompts that ask for alternatives to your second rival
• 1 prompt that asks for alternatives to your own brand

The last one is defensive. It shows which rival an engine hands your own customers when they ask how to leave.

Feature comparisons (6 prompts)

Use four feature-led prompts that name no brand, built around the features you think you win on. Add two head-to-head prompts that name you and one rival each, such as "Plannora vs Loopcraft for client work: which is better?" For those two, record the answer's verdict: which brand it recommends, or no clear pick.

Five rules that keep the matrix fair

1. Keep your brand out of 17 prompts. Only the two head-to-head prompts and the one defensive prompt name you.
2. Never score a brand that the prompt names as its own alternative. In "alternatives to Loopcraft," Loopcraft isn't eligible, so that prompt only scores you against Taskwell.
3. Freeze the wording. Run the same text every time. A new phrasing is a new prompt.
4. Pick rivals from the answers, not your sales deck. Run the matrix once, then track the two or three brands that appear most. The brand that AI recommends may not be the one you lose deals to.
5. Tag each prompt with the page you expect to win it. This turns the result into a to-do list.

Our free AI Visibility Prompt Kit writes 30 buyer prompts from your brand, category and a competitor. Use it as a pool, then sort 20 of them into the matrix.

How to Run the Matrix and Log AI Citations

Engines, runs and clean sessions

Pick the three or four engines your buyers use. For most B2B software, that's ChatGPT, Perplexity and Google AI Mode, with Gemini or Claude as a fourth. Run each prompt three times per engine, on the same days, from the same place. That gives 180 answers across three engines, which is the minimum this method is built for.

Keep the sessions neutral. OpenAI says that if memory is on, ChatGPT "may use relevant saved memories when rewriting a search query", and that it estimates your location from your IP address, which a VPN can change. Log out, or use a temporary chat with memory off, and use one location for every run. Record the model or mode each engine shows, so a model change doesn't pass as a real gain.

What to write down for every answer

Log one row per answer. Keep the brands in the order the answer first names them, and include brands you don't track. Peec AI's documentation makes the same choice: its position metric "accounts for every brand detected in a chat, not just tracked competitors." If an untracked brand comes first, your rival is second, not first.

date,engine,prompt_id,type,run,brands_in_order,verdict,cited_urls,cited_pages_naming_rivals_not_you
2026-09-22,perplexity,prob-3,problem,2,"Loopcraft; Wrenboard; Plannora",,"loopcraft.ai/guides/deadlines; stackreview.co/best-pm-tools","stackreview.co/best-pm-tools"
2026-09-22,chatgpt,h2h-PL,feature,1,,Loopcraft,"stackreview.co/loopcraft-vs-plannora",""

Three columns do the heavy lifting:

• brands_in_order feeds every mention metric: Share of Model, AI Share of Voice, First-Mention Rate, Average Answer Position and the head-to-head score.
• cited_urls feeds the metrics for AI citations. Count each domain once per answer for Citation Rate, and count each distinct URL for Citation Share.
• cited_pages_naming_rivals_not_you is your gap list. Open the cited third-party pages, such as reviews, roundups and forum threads, and note which brands each one names. A page that is cited often and names your rivals but not you is where to act first.

What counts as a recommendation

For Share of Model and the other mention metrics, count any brand the answer names, as the framework does. The head-to-head score is stricter. A brand only wins a contest when the answer offers it as an option: in a list, a table, a sentence ("Loopcraft is a good fit if…") or a closing tip. A brand named only to rule it out ("Taskwell isn't built for agencies") wins nothing, so log it as a negative mention in a notes column. Never count a name the answer just repeats from the prompt. For tables, order is row order. For prose, order is where each brand first appears.

The Scoring Formulas

Every number comes from the same log. The first six metrics use the names and formulas of our GEO Metrics Framework, which defines each one with notation and worked examples. The last two are this matrix's own. Measure the mention metrics on the 12 unbranded prompts only (the problem queries and feature-led prompts), because a prompt that names a brand is no fair test of whether the engine thinks of it. The framework asks for each engine to be computed on its own, or given equal weight when blended; a matrix with the same runs per engine does that for you.

How to count wins and losses

For each answer, compare your brand with one rival at a time:

1. Only you are recommended: a win.
2. Only the rival is recommended: a loss.
3. Both are recommended: the one named first wins.
4. Neither is recommended: no contest. The answer doesn't count toward the head-to-head score, but it still counts toward Share of Model.
5. Head-to-head prompts: the verdict decides. "No clear pick" is no contest.
6. Either brand is named in the prompt as the thing to replace: no contest for that pair.

Then check the margin before you trust a score. With the head-to-head score as p and the number of contests as n, the 95% margin is:

margin = 1.96 × √( p × (1 − p) ÷ n )

Treat this as a floor. Runs of the same prompt aren't fully independent, so the true margin is a little wider.

Worked Example: Plannora vs Loopcraft and Taskwell

Plannora, Loopcraft and Taskwell are made-up brands, and the answers are an illustrative dataset. The arithmetic is real, so you can check every line. Plannora ran the 20-Prompt Rival Matrix on ChatGPT, Perplexity and Google AI Mode, three runs each: 180 answers.

Step 1: Share of Model and order

These are the 12 unbranded prompts: 108 answers, 99 of which name at least one brand. Across them, brands appear 207 times, 70 of them for brands Plannora doesn't track. Every prompt has the same nine runs, so the pooled rate equals the framework's per-prompt average.

Loopcraft is named in far more answers. Yet when Plannora is named, it sits almost as high as Loopcraft does. Plannora's problem is being left out, not being placed low.

Step 2: The head-to-head score by prompt type

The totals check out: against Loopcraft, 9 + 6 + 22 + 1 = 38 wins and 43 + 12 + 5 + 7 = 67 losses. The alternative rows only count the prompts that don't name the rival: Plannora meets Loopcraft in the two "alternatives to Taskwell" prompts, and Taskwell in the three "alternatives to Loopcraft" prompts.

Plannora's Rival Matrix Score is 88 wins out of 196 contests, or 44.9%. With 196 contests, the margin is ±7.0 points, so the true figure sits somewhere between about 38% and 52%.

Step 3: Is each gap real?

• Against Loopcraft: 36.2% ± 9.2, a range of 27% to 45%. The whole range is below 50%, so Loopcraft really is recommended over Plannora more often than not.
• Against Taskwell: 54.9% ± 10.2, a range of 45% to 65%. That range includes 50%, so Plannora can't yet claim to beat Taskwell. It is roughly even.
• Problem queries against Loopcraft: 17.3% ± 10.3 across 52 contests. Even the top of that range is far below even. This is the biggest gap in the matrix.

The head-to-head verdict row is striking (Loopcraft picked 7 times out of 9), but it rests on only eight contests. On its own, it's a hint. Step 4 explains it.

Step 4: Where the AI citations come from

Across the 180 answers, the engines made 647 AI citations, each one a cited URL. 162 answers cited at least one source.

Overall, plannora.io earns almost as many AI citations as loopcraft.ai: 11.6% of cited URLs against 13.0%. The split by engine tells a different story. In Perplexity, plannora.io earned 47 of 326 cited URLs (14.4%) to Loopcraft's 36 (11.0%). In Google AI Mode, it earned 21 of 205 (10.2%) to Loopcraft's 36 (17.6%). In ChatGPT, it earned 7 of 116 (6.0%) to Loopcraft's 12 (10.3%).

The gap list explains the verdict row. One made-up review page, a stackreview.co roundup of project tools, was cited in 50 of the 180 answers (27.8%). It names Loopcraft and Taskwell but not Plannora. A second page on the same site, "Loopcraft vs Plannora," was cited in 19 answers, and its verdict favors Loopcraft. In this example, the engines echo the pages they read.

Step 5: What Plannora does next

1. Write problem-led pages. Plannora wins when buyers name features (81.5% and 84.0%) and loses when they describe the pain (17.3% against Loopcraft). It needs pages that answer the problem prompts in buyers' words, such as a guide to managing client feedback in agencies.
2. Fix the pages behind the AI citations. Plannora publishes its own fair "Plannora vs Loopcraft" page with current facts and asks stackreview.co to review Plannora for its roundup.
3. Watch the defensive prompt. In "alternatives to Plannora," Loopcraft was named in 7 of 9 answers and came first in 5. Plannora's existing customers are being pointed at Loopcraft.
4. Rerun in four to six weeks with the same 20 prompts, and compare each head-to-head score against its margin.

How to Read Your Benchmark

Two checks stop most bad readings. First, compare like with like. If a rival blocks an engine's crawler, a low score there reflects its choice, not its content. Our AI Bot Crawler Census shows how often sites block each bot. Second, don't read a single month as a trend. Our guide to benchmarking AI search performance has the table of how big a gap must be at each panel size.

If your competitor benchmark keeps moving because engines change, the control-prompt test in our GEO measurement guide separates your work from that drift.

Tools That Report Competitor Share of Voice

You don't need software for a one-off matrix. A spreadsheet works. But the old idea that SEO tools only show keyword overlap with rivals is out of date. As of September 2026, several trackers report AI share of voice against named competitors, and each defines it a little differently. Prices below are as listed on 28 September 2026, billed monthly.

The definitions matter more than the prices. Semrush's average position follows the order of citations, while Peec's follows the order of brands named in the text, so the two can disagree about the same answer. Ahrefs weights share of voice by search volume, while Peec counts mentions equally. Before you compare your matrix with a dashboard, check which formula the dashboard uses.

Two first-party reports show your own AI citations or AI impressions, though neither names rivals. Bing Webmaster Tools reports Citation Share across Copilot, Bing and select partner AI experiences, but Microsoft's June 2026 post says it "does not expose competitor domains." Google's generative AI performance report in Search Console shows how your own site performs in AI Overviews and AI Mode. For a fuller list of trackers and the questions to ask them, see our ChatGPT rank tracker guide.

Where Rankbox Helps After the Benchmark

Rankbox doesn't track AI citations, mentions or share of voice today, so it won't run the matrix for you. Use a spreadsheet, the free Prompt Kit or one of the trackers above for that part.

Rankbox helps with what the matrix tells you to fix. Answer-Space Research maps the questions buyers ask AI in your category, which is a good source for the problem queries you're losing. The Citation-Ready Writer then drafts source-backed articles for those questions, and they reach your site through Rankbox's API. The Business plan is $49.50 a month with a 7-day trial. See pricing.