How to Track Competitors in AI Search Results, Step by Step

45% of marketing leaders cannot accurately measure brand visibility in AI answers, and only 9% have the tools to track every relevant metric across platforms, per the Semrush 2026 AI Visibility Index, built on 126 million U.S. AI search prompts. Most teams are guessing about themselves. About their rivals, they are guessing harder.

Meanwhile your rank tracker still says position three, and the answer above that result names three other brands and never yours. Organic CTR for informational queries with AI Overviews fell 61% from June 2024 to September 2025, per Seer Interactive’s study of 3,119 search terms across 42 client organizations, and Seer’s April 2026 update on a larger dataset found CTR on AI Overview queries rebounding from 1.3% in December 2025 to 2.4% in February 2026, still below early 2025 levels. Either way, the old scoreboard tracks clicks and never shows who the answer named.

What follows is the method we use to track competitor rankings in AI search results for our own clients: the prompt suite, the run count that makes a number real, the clean-room rules, and the log schema.

How do you track competitor rankings in AI search results effectively?

You run a fixed set of buyer questions through each engine, several times, in a clean session, and count who gets named. Six steps: pick the panel, write the prompt set, set the clean room, run and repeat, log mentions plus cited URLs, then compare across a fixed window.

1. Pick the panel. Five to eight rivals, chosen once and frozen. Include the two teams you lose listings to and one smaller agent who keeps showing up. Swapping the panel mid-quarter destroys the comparison.
2. Write 40 prompts your buyer would actually type, split across archetypes (the archetype table below has the split).
3. Set the clean room before the first run. Fresh session, memory off, locale written down.
4. Five runs per prompt, per engine. Same day, same conditions.
5. Log two things every time: which brands appear, and which URLs the engine cites.
6. Compare month over month. Never day over day.

One word to drop is rank, because there is no position one in an AI answer. There is mentioned or not mentioned, cited or not cited, named first or named fifth, and that scoreboard behaves nothing like a ten-blue-links report. Track mentions and citations, not positions.

What is share of voice in AI search, and how do you calculate it?

Share of voice is your brand’s mentions divided by every panel brand’s mentions, calculated per engine, per window. If your panel got named 283 times across 200 logged answers and you were named 34 of those times, your share of voice is 12%.

12%

is your share of voice if a panel got named 283 times across 200 logged answers and you were named 34 of those times.

The formula in plain words:

Your mentions ÷ all panel-brand mentions, for one engine, over one fixed window.

Here is a worked example. The numbers below are illustrative, not measured client data, and they exist to show the shape of the sheet.

Notice the fourth column. Competitor A is not just mentioned more, it is being pulled from three times as many sources. That gap is the actual work order, and it is far more useful than the percentage next to it.

A share-of-voice number with no run count attached is decoration, so write the denominator on the chart: 40 prompts, 5 runs, 4 engines, this month. Without it nobody can check you, including you in six months. It is the headline metric when you benchmark AI visibility against competitors across GEO platforms, and the easiest one to fake by accident, which brings us to sample size.

How many prompts and how many runs make a competitor number real?

Prompt count sets your coverage and run count sets your confidence. Forty prompts, five runs each, per engine, is the working floor. Most tracking setups get the second number wrong, which is why so many AI visibility dashboards show dramatic weekly swings that mean nothing.

Variance study

34.8%

of the variance in LLM brand answers came from within-prompt resampling, against 0.7% for the brand’s own score, in a July 2026 preprint.

Photo: ulrichw, Pixabay

Start with the evidence. A July 2026 variance-components preprint on non-determinism in LLM brand answers decomposed 12,933 responses about 20 Central and Eastern European brands, asked in eight languages across three models. On the 7,173 responses where the same prompt was repeated, within-prompt resampling accounted for 34.8% of the variance, while the brand’s context-free true score accounted for 0.7%. Ask the same question twice and the answer changes, all by itself. The study also found that a repeat past the fifth cuts relative-error variance by about 0.0003, which is the practical argument for stopping at five.

Two limits on how far to lean on it. The study scored each response’s sentiment toward the brand, not whether the brand was named, so treat 34.8% as the shape of the noise and measure your own. And its decision analysis found that, block for block, adding languages cut relative-error variance about fifteen times as much as five more repeats, adding models about six times as much and rewording prompts about four times as much. Applied here, spend the next hour on a second engine or a reworded prompt rather than a sixth run.

That gives you one rule worth taping to the wall:

Do not act on a share-of-voice change smaller than your own run-to-run spread.

Finding that spread takes an afternoon. Run your existing prompt set twice on the same day, same engine, same conditions, and measure the gap between the two results. That gap is your noise floor. Anything under it is weather.

Prompt count is a coverage decision

The 50 to 100 prompt figure is coverage advice. Onely’s tracking framework sets it at 50 to 100 queries per test suite, and most vendor pages repeat some version of that number. It tells you how much of your category you can see, not how much you can trust. Forty well-chosen prompts run five times beats 100 prompts run once, every time.

Prompt wording moves the number as much as prompt count

Wording is not a detail. In a Peec AI study of 37,804 AI responses across five engines, reported by Search Engine Journal, brand mentions dropped roughly 50% when prompts fell to 0.35 to 0.39 cosine similarity with the original prompt, and list or ranking prompts produced up to 20% higher average visibility. Rewording that kept the meaning left brand visibility stable above roughly 0.50 to 0.60 similarity, so the swings came from drifting meaning and from prompt format. So freeze the phrasing.

Here is the suite we build for an agent or brokerage, sized at 40 prompts. It is a template, not a measured benchmark, and you rewrite the brackets for your own market.

“This sounds like a lot of work” is the fair objection to any honest way to track competitor rankings in AI search results. The smallest version is 40 prompts, 5 runs and 4 engines, which is 800 logged answers a month from one person with a spreadsheet and a private window, in about an afternoon.

How do you set up a clean room so the answer you log is real?

Your own account contaminates the result. Chat memory, past conversations, saved preferences, account history and location all personalize what an engine says, so a logged-in run tells you about you, not about the market. The clean room removes the personalization so the answer you log is closer to what a stranger sees.

The protocol, every single time:

• Fresh session or a private window. No exceptions, no “it is probably fine”.
• Memory, personalization and custom instructions switched off.
• One prompt per session. No prior turns in the thread, because the previous answer steers the next one.
• Write the locale into the log. US English and AU English return different brands.
• Same model version across the whole panel, for the whole batch.
• Run the entire batch inside one day.

Record the model name and version on every row. Engines update silently and nobody sends you a memo, so a share-of-voice drop in March can turn out to be a model change rather than a competitor win. Without the version column you will never know which one it was.

People skip this step, then wonder why their numbers disagree with their agency’s, but you cannot track competitor rankings in AI search results from a logged-in tab.

What log schema makes your comparison repeatable in six months?

The schema is the deliverable. Method without a log is memory, and if a column is missing the comparison cannot be re-run, which makes it an anecdote with a chart on top.

These are the exact fields we log for every single answer, one row per run.

Those thirteen columns give you every view you need: share of voice by engine, mention rate by prompt archetype, citation source overlap between you and each rival, and month-over-month movement with the noise floor drawn on it. The same discipline applies to your own AI visibility audit.

Screenshots age badly. A sheet re-runs.

Why do competitors appear in AI search results when your brand does not?

Usually not authority. Usually sources. The engine is quoting places your competitor shows up and you do not, and those places barely overlap between engines, so a rival can own Perplexity while being invisible in ChatGPT. Tracking tells you which pocket you are missing.

The overlap evidence is blunt. Profound’s analysis of 680 million citations, August 2024 to June 2025, found that Reddit is 46.7% of the citations to Perplexity’s ten most-cited sources, and Wikipedia is 47.9% for ChatGPT’s top ten. Those are shares of each top-10 list, not of all citations, and they describe two different internets. In one analysis, The Digital Bloom’s 2025 AI Visibility Report, only 11% of cited domains are shared by ChatGPT and Perplexity, though the report does not say what data that figure comes from.

So the tracking output turns into a work order. For every domain citing a competitor and not you, sort it into one of three buckets:

• Fix. You have a page on that topic and it is weak, thin, or buried under a hero image.
• Build. No asset exists yet. Write the thing the engine keeps quoting from someone else.
• Influence. Third-party surface you do not own: a roundup, a subreddit, a review site, a comparison page.

We see the gap in our own numbers. In our 500-query Houston study, portals and agent referral networks took 43.6% of the citations behind the answers, while a firm’s own domain was cited in only 8.4% to 10.4% of the answers naming it. Our case studies show what closing that source gap looks like in practice, and why AI skips a brand that ranks fine on Google covers the diagnosis.

One more from Semrush’s index: only 36 global brands held top-100 visibility on all four platforms in every month of the study. Nobody wins everywhere, so pick your engine and win there first.

How do you track competitors in AI search as a real estate agent or brokerage?

Build the panel per area rather than per city, and keep each engine on its own line. In CiteVantage’s 500-query Houston study, an estimated 82.6% of the agents, teams and brokerages AI named appeared in just one of 40 areas, and 83.3% on only one of four surfaces, so your real competitors change from one neighborhood and one engine to the next.

In practice that means one panel per area you sell in. When most named agents hold a single area, the rivals AI names in one suburb are mostly different people from the ones it names in the next, and a single metro-wide panel blends several local markets into one share-of-voice figure that describes none of them. Three more adjustments, each tied to a measured finding:

• Split prompts by intent inside each area. In the Houston study, 94.0% of the firms named in new-build answers were absent from the generic answer for the same area, and investment queries displaced 76.1% of names.
• Expect one rival to hold a large share. Local Falcon, in a vendor measurement across the 100 largest US cities, reported that in the average city a single agent takes 26.6% of everything AI says about agents in that market, and four agents take 56.1%.
• Log URLs for portals, not just domains. The same Local Falcon index reported that 86.9% of Zillow’s AI citations point at a city list while only 13.1% point at a named agent, so “zillow.com” in your log tells you little until you see which page it was.

Both Local Falcon figures sit with their sources on our real estate SEO statistics page. Our Houston page shows recorded AI answers naming local agents, and our GEO for real estate plans re-run a fixed question set on each engine every one to two weeks.

Why do two AI tracking tools report different numbers?

Different prompt sets, run counts, regions and model versions mean two honest vendors can still hand you numbers that disagree by twenty points. Before you pick a side, check whether the gap is a measurement gap or just resampling.

The ten-minute spot check, before you trust any tool that claims to track competitor rankings in AI search results:

1. Pull five prompts from the tool’s own prompt list.
2. Run each one by hand, clean session, five times.
3. Log mentions in your sheet, using the log schema columns above.
4. Compare against what the tool recorded for the same day and region.
5. If the tool cannot show you its prompt list, its run count and its region, stop. That is the finding.

Some of that gap is not error, because a chunk of any two-tool disagreement is the model answering itself differently. That is why the noise floor comes first and the tool comparison second. For the questions to put to a vendor, our guide on how to verify what a tracking tool measures goes deeper.

How often should you track competitors in AI search?

Monthly for the scoreboard. Weekly only during a live test, and only on the prompts that test touches. A single-day swing is almost never a trend, because the mechanism that moves AI answers is slow and the noise that moves them is fast.

Content, schema and third-party mentions take weeks to work their way into what engines quote, so a weekly share-of-voice number mostly reads resampling. Monthly gives the slow signal time to show up, and quarterly is too slow to catch a competitor move you could still answer.

Weekly earns its place in three cases:

• During an active sprint, when you changed one thing and want to watch it.
• On a subset of prompts, never the full 40.
• When a rival has just launched something and you need the before picture fast.

A model version change, a competitor launch or a large third-party roundup going live in your category overrides the calendar, and any of them means re-running the batch.

Reporting weekly to a client teaches them to react to noise, and you spend the next call explaining a four-point drop that was never real. Report monthly, with the noise floor on the chart, and hold other people’s reporting to the same bar: our read of successful GEO campaigns case studies shows which published results state a baseline, name the engines and give a date range.

What tools track competitors in AI search, and can you do it for free?

Tools save time, not judgment. Every tool built to track competitor visibility in LLMs does what you can do by hand, which is run prompts in a session and count mentions. So the real question is whether it will show you its prompt list, its run count and its region. Free is genuinely possible, only slower.

Best practices for benchmarking AI answer visibility against competitors

The zero-spend stack is three pieces: the Performance report in Google Search Console, a spreadsheet built on that log schema, and one clean-room afternoon a month. That combination will tell you more about how to track SEO effectiveness in AI search engines than most dashboards, because you control every variable that produced the number. If you want the brand-side tooling instead, our roundup of tools that track your own brand mentions in ChatGPT covers that half, and the Google AI Mode tracking tools review covers the AI Mode surface.

Paid tools are worth it when the time cost of the manual log exceeds the subscription. Not before. If you get to that point, the four categories above price very differently, and our breakdown of the most popular AI visibility products for SEO shows what each type measures and compares them on price per tracked prompt and engine.

Want the competitor panel run for you first?

Building the sheet is the right long-term move if you plan to track competitor rankings in AI search results month after month, and a baseline this week is faster. Our Free AI Visibility Audit checks your brand across ChatGPT, Perplexity, Gemini and Copilot, and sends back screenshots of what those engines say on your buyer questions, plus which competitor gets named instead of you. There is no call and no card. Start with the Free AI Visibility Audit, or see how the same work runs for real estate agents and brokerages end to end.

What the gap on that sheet costs is hard to price, but Semrush’s AI SEO statistics roundup cites a 2025 Semrush figure that the average AI search visitor is worth 4.4x more than a traditional organic search visitor in conversion terms, with no method shown.