How to Check If Your Competitors

Short answer: Yes, you can check, but not with a single prompt. Build a set of 20 to 30 buyer-style questions, run each one several times in ChatGPT, Perplexity, Copilot and Gemini, and log which brands are named and which URLs are cited. Then calculate your mention rate, citation rate and share of mentions against each competitor, and list the third-party sites that get cited for them but not for you. Those source gaps tell you what to fix first. Treat any one-off answer as an anecdote, because AI tools almost never return the same brand list twice.

You open ChatGPT, ask for providers in your category, and a competitor is named first. Your brand isn’t there. It is tempting to treat that as proof you are losing, or as a glitch to ignore. It is neither. It is one sample from a system designed to give different answers every time.

This guide gives you a repeatable method instead: what to ask, how to run it, what to log, which numbers to calculate, how to score sentiment, and how to turn the results into a fix list. It is vendor-neutral and works with free accounts and a spreadsheet. If you only want the overview of why AI citations matter, start with our earlier post on competitors being cited in AI search. This post is the measurement manual that follows it.

What you need: a spreadsheet, accounts on the four AI tools (free tiers are enough to start), a list of three to five competitors, and roughly three to four hours for the first full run. After that, a monthly refresh takes about half that.

Why can’t you just ask the AI once and compare?

Because the answers are probabilistic. In research published in January 2026, SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google’s AI features, for 2,961 runs in total. The chance that any two responses contained the same list of brands was under 1 in 100, and the chance of the same order was closer to 1 in 1,000. The lists also varied in length, from 2 or 3 recommendations to 10 or more.

That sounds like bad news for tracking, but the same research found a usable signal. Across many runs of the same prompt, some brands appear again and again. In one example, a single agency appeared in 85 of 95 Google AI responses. The authors concluded that visibility percentage across dozens to hundreds of prompt runs is a reasonable metric, while a brand’s exact rank position in an AI answer is mostly noise. Note that Gumshoe, which co-ran the study, sells AI tracking, and the authors list open questions such as how many runs are statistically enough.

So the method has two rules: measure how often you appear across many runs, not where you appear in one, and compare yourself to competitors using the same prompts under the same conditions.

What should you measure?

Five numbers cover most decisions. The first three tell you whether competitors are ahead. The last two tell you why.

Metric

How to calculate it

What it tells you

Mention rate

Responses that name your brand ÷ total responses

How often you are in the consideration set

Citation rate

Responses that cite a URL from your domain ÷ total responses

How often your own content is used as a source

Share of mentions (share of model)

Your mentions ÷ all brand mentions across tracked brands

Your slice of the conversation versus competitors

Source gap

Domains cited alongside competitors but never alongside you

Where to earn presence

Sentiment and accuracy

Score each mention as positive, neutral or negative, and correct or incorrect

Whether being mentioned helps you

Keep mention rate and share of mentions apart. Similarweb’s published example shows why: a retail brand appeared in 32.77% of tracked AI answers, yet held only 4.71% of all brand mentions across them. It was present often but outnumbered by competitors named alongside it.

Bing’s documentation draws a similar line between volume and share. Total citations measure how often your content is referenced, while citation share measures your portion of the citations for one grounding query.

How do you check which competitors AI search engines cite? A 9-step method

Step 1: Define who and what you are tracking

List your brand and three to five direct competitors. Include the variants a model might use: the full company name, the short name, the product names, and the domain. Add one or two “wildcard” competitors, such as a directory, marketplace or large publisher that tends to appear in answers for your category. AI answers often name sources you do not think of as competitors.

Step 2: Write 20 to 30 prompts in your buyers’ words

People do not prompt like they search. SparkToro’s survey of 142 volunteers found almost no two prompts alike, even when the intent was the same, yet the AI tools still returned a fairly consistent set of brands for that intent. So write prompts that cover intent, not exact wording. Use four groups:

  • Category: “What are the best providers for a [company type] in [market]?”
  • Problem: “Our [metric] dropped after [event]. Who can help and what should we do first?”
  • Comparison: “Compare [Brand A], [Brand B] and [Brand C] for [use case].”
  • Brand and reputation: “Is [Brand] a good choice for [use case]? What do customers say?”

Orbit Media notes that prompts are more likely to trigger a live web search, and therefore citations, when they are very specific, ask for current information, ask for sources, or concern a high-stakes decision. Include some of each, such as “in 2026” or “with sources”. Draw wording from sales calls, support tickets and the questions prospects ask on discovery calls rather than from a keyword tool.

Step 3: Standardise the test conditions

The SparkToro volunteers used default settings and results still varied, so you cannot remove variance, but you can keep it from distorting your comparison:

  • Start a new chat for every run so earlier answers do not bias later ones.
  • Turn off memory and custom instructions where you can, because ChatGPT may use saved memories when it rewrites a search query.
  • Use the same country and device. ChatGPT may infer approximate location from your IP address, and a VPN can change it.
  • Record the date, tool, plan (free or paid) and whether web search was on.
  • Change the prompt set only on a planned schedule, and version it (v1, v2), so month-to-month numbers stay comparable.

Step 4: Run each prompt several times on each platform

A workable minimum for a small business is 25 prompts × 3 runs × 4 platforms = 300 responses. A more robust run is 5 runs per prompt. SparkToro suggested 60 to 100 runs of a single prompt for a precise picture of one consideration set, which is impractical across a full prompt library, so treat your own numbers as directional and widen the sample over time. That recommendation is ours, not an established standard.

Interfaces change often, so log what each tool actually shows rather than assuming:

Platform

How to get sources

Record

Watch out for

ChatGPT

Use the Search tool (View all tools, then Search) or ask for current information; open Sources or the citation links

Brands named, cited URLs, whether search ran

In August 2026 some accounts showed sources only under a three-dot menu, so check there. OpenAI notes citations can be incomplete or incorrect

Perplexity

Read the source links shown with the answer

Brands named, cited domains

Run in a fresh thread each time

Microsoft Copilot

Open the links or references shown with the answer

Brands named, cited domains

Bing’s AI Performance report covers Copilot citations for your own site only

Gemini

Check the sources or links the answer displays

Brands named, cited domains

If no sources appear, log “no citations”, which is itself a result

Google’s AI Overviews and AI Mode are separate surfaces. Measure those with Search Console (Step 9), and add them to the manual run only if your buyers use them for research.

Step 5: Log every response in a simple sheet

One row per response. Paste the full answer into a separate column or tab so you can recheck it later.

Date

Tool

Prompt ID

Run #

Search used?

Brands named (in order)

Your brand named?

Cited domains/URLs

Sentiment (-1/0/+1)

Accuracy (ok/outdated/wrong)

Notes

 

 

 

 

 

 

 

 

 

 

 

Two habits keep the data clean. First, log brands and domains exactly as shown, then normalise them (for example, “G2”, “g2.com” and “G2 Crowd” become one entry) in a second pass. Second, record the order if you like, but do not build a KPI on it, since order is the least stable part of an AI answer.

Step 6: Calculate your share of model

Once the log is full, three formulas give you the comparison. The numbers below are illustrative only, to show the arithmetic, not real results.

Suppose you ran 25 prompts × 4 runs × 4 platforms = 400 responses. Across them, tracked brands were named 500 times in total.

Brand

Responses naming it

Mention rate

Share of mentions

Responses citing its own domain

Citation rate

You

84

21%

16.8%

36

9%

Competitor X

188

47%

37.6%

112

28%

Competitor Y

120

30%

24.0%

52

13%

In this example Competitor X is named in more than twice as many responses as you and gets its own pages cited three times as often. That is a real gap. A smaller difference, such as 21% against 24%, could easily sit inside the noise. With 400 responses, a proportion near 50% carries a margin of roughly ±5 percentage points at best, and less at the extremes, and runs on the same prompt are not fully independent, so treat that as a floor. Only act on gaps that are large, that appear on several platforms, and that repeat in the next monthly run.

Break the table down by prompt group and by platform too. A common pattern is strong mention rate on brand and comparison prompts but weak presence on category and problem prompts, which is where new buyers start. Similarweb’s published analysis found the same shape: broad prompts surfaced familiar brands, while specific prompts about attributes or use cases changed the answers completely.

Step 7: Score sentiment and accuracy

Being named is not the same as being recommended. Orbit Media makes the point that visibility is the start, and the real goal is to be recommended. Score every response that names you with a short rubric:

Dimension

Scale

How to score

Sentiment

-1 / 0 / +1

+1 if described as a strong fit or recommended, 0 if listed neutrally, -1 if cautioned against or described with a weakness

Role

Recommended / Listed / Alternative / Warned

Where the answer places you in the decision

Accuracy

Correct / Outdated / Wrong

Check services, pricing, location and claims against your own site

Descriptors

Free text

The two or three words the model uses for you, such as “affordable” or “agency-focused”

Average sentiment by competitor, then read the actual descriptors side by side. They show you how each brand is positioned in the model’s eyes, which is often more useful than the number itself. Wrong or outdated statements are the cheapest wins, because you can usually trace them to a specific source and correct it.

If you use an AI tool to help classify a long log, treat it as a first pass and manually check a sample of about one in ten rows, since the classifier is subject to the same variability.

Step 8: Run the source-gap analysis

This step answers the question in your query: why are they cited and you are not. For every response, list the domains cited. Then classify each domain:

  • Owned: the brand’s own site.
  • Earned: news, press, industry publications and roundups.
  • Community: forums and Q&A sites such as Reddit and Quora.
  • Intermediaries: review platforms, directories and marketplaces.

Then build a simple domain table: for each domain, count the number of distinct prompts where it was cited, and note which brands appeared in the same answers. Scrunch calls a version of this an influence score, defined as the share of responses citing a source multiplied by the number of unique prompts it appears for. A hand-built count of prompts per domain is enough to start.

The gaps fall into three groups, borrowing TopCited’s useful framing:

  1. You have no content on the question. The engine cannot cite a page that does not exist.
  2. You have content, but you are not cited. Usually a structure or authority problem, not a subject problem.
  3. The question is not yours to win. Some answers come from a competitor’s own documentation or a reference site. Do not chase those.

The third-party domains that appear repeatedly across prompts and platforms are your target list. Similarweb’s example showed a handful of listicles and publisher pages shaping answers across dozens of prompts, which meant getting included there could matter more than publishing another page on the brand’s own site.

Step 9: Cross-check with the data AI platforms give you

Manual testing shows you competitors. Native reports show you yourself. Use both.

Google Search Console, Generative AI performance report. Google launched it on 3 June 2026 for a subset of sites, and its help page says it rolled out to all websites worldwide on 31 August 2026. It reports impressions for your pages in AI Overviews and AI Mode, split by page, country, device and date. It does not report clicks or competitors, so it tells you whether your pages appear, not who beat you.

Bing Webmaster Tools, AI Performance. Launched in public preview on 10 February 2026, it shows how often your site is cited across Microsoft Copilot, AI summaries in Bing and select partners. You get total citations, average cited pages, grounding queries and page-level counts. Bing says the data is a sample, a citation is not a click, and the newer Citation Share view does not reveal which competitors hold the rest.

Because neither report names competitors, the manual log is still the only way to see them. Use the native reports to confirm direction: if your Bing citations are flat while a competitor appears more in your Copilot tests, that is a consistent signal.

Which prompts should you use to check competitor citations and sentiment?

Copy these into your tools and replace the brackets. Run each several times in fresh chats.

Category discovery

I’m a [role] at a [company type] in [country]. Which 

 providers would you shortlist for [goal], and why? Please cite sources for each recommendation.

Head-to-head comparison

Compare [Brand A], [Brand B] and [Brand C] for [use case and budget]. Give strengths, weaknesses and who each is best for, with sources.

Reputation and sentiment

What do customers, reviewers and industry sources say about [Brand]? Separate positive points from complaints, and say where each claim comes from.

Problem-first (where new buyers start)

Our [symptom, e.g. demo requests fell while rankings held]. What are the likely causes, and which agencies or tools could help? Include sources.

Alternatives

I’m considering [Competitor]. What are the best alternatives for [use case], and what are the trade-offs?

One prompt to be careful with is the common suggestion to ask a model what its training data says about you, or which sites it relies on. The answer is a plausible-sounding guess, not a log of what the system did. Orbit Media itself cautions that a model’s tips for improving your odds are partly echoes of generic SEO articles. Use the sources actually cited in real answers as your evidence.

How do you read the results?

Match what you see to the likeliest cause before choosing a fix.

What the data shows

Likely cause

First move

Competitors named, you absent, on category and problem prompts

Weak topical coverage or weak third-party presence

Source-gap analysis, then content and earned mentions

You are named but never cited, and competitors’ own pages are cited

Your pages are less direct, less specific or less authoritative than theirs

Restructure the key pages and add original evidence

Mentioned on one platform but not another

Platform-specific retrieval or crawler access

Check crawler access and indexing for that platform

Named but described wrongly or negatively

Outdated or negative third-party sources

Correct at the source, strengthen first-party facts

Absent from everything

Entity not recognised, or site not reachable

Entity and access audit before content work

How do you fix it when competitors are cited more than you?

Work in this order. Access problems are quick to fix and block everything else, while earned mentions take longest, so start those early. No platform publishes a formula for citations and none guarantees one, so treat each step as improving your odds.

1. Make sure the right crawlers can reach you

OpenAI says that to be eligible for ChatGPT search results you should allow OAI-SearchBot and confirm your host or CDN allows traffic from OpenAI’s published searchbot IP addresses. For Google’s AI features, Google says a page must be indexed and eligible to show with a snippet, and the site must be included in Search generative AI features in Search Console. Check robots.txt, firewall and CDN rules, and any nosnippet settings. Our post on how AI decides which brands to cite has the crawler-by-crawler table.

2. Close the content gaps, by type

  • No page exists: write one, but only for questions your buyers actually ask.
  • A page exists and is not cited: put a direct answer under a question-style heading, make claims stand alone, add a comparison table where it helps, and add original evidence such as your own numbers, screenshots or first-hand notes. Google says non-commodity content, meaning a unique point of view or experience, is likely to influence your presence in generative AI search more than any other tip in its guide.
  • Not yours to win: skip it.

Google also warns that creating a separate page for every query variation to manipulate AI responses can violate its scaled content abuse policy. Cover sub-questions in depth inside one strong page instead of cloning pages.

3. Close the source gaps with real mentions

Take the domains from Step 8 and ask, for each, what a legitimate way onto that page looks like: a contributed expert comment, original data worth quoting, a correction to outdated information, a genuine review profile, a directory listing with accurate details. Google states that seeking inauthentic mentions is not as helpful as it might seem, so earn them by being useful rather than buying placements.

4. Fix your brand entity

A brand entity is how consistently the web states who you are: your name, what you do, who you serve, where you operate, and which profiles belong to you. Models resolve conflicts by hedging or leaving a brand out, so contradictions cost you. Audit and align:

  • Business name, description and service list across your site, Google Business Profile, Bing Places, LinkedIn, directories and review platforms.
  • Leadership, address, phone and founding details.
  • A clear About page and one-sentence description that every profile repeats.
  • Links between your official profiles, so they point to one organisation.

5. Use structured data for clarity, not as a citation lever

Structured data belongs in this plan, with honest expectations. Google’s generative AI guide says structured data is not required for its AI features and that no special schema.org markup exists for them, though it still recommends it for rich-result eligibility. So use Organization markup (with name, URL, logo and sameAs links to your official profiles), Article markup on posts, and Product or Service markup where it fits, validate it, and make sure it matches the visible page. The value is that it states your entity facts unambiguously. Do not expect markup alone to change who gets cited. Our AIO guide covers this in more depth.

6. Correct wrong or negative narratives at the source

You cannot edit a model’s answer. If a response is outdated or wrong, find the cited page behind it, request an update, publish a clearer first-party statement, and keep key pages current. Our post on what AI models say about your business goes deeper on the reputation side.

7. Re-measure on a fixed schedule

Run the same versioned prompt set every month and compare against the previous run. Similarweb suggests monthly reviews are usually enough to spot trends, and Scrunch argues that two to three weeks of history reveals what models consistently rely on, since a single snapshot can change the next day. Report trends, not single readings.

Do you need a paid AI visibility tool?

Not to start. The manual method above is enough to find out whether you have a problem. Tools become worth considering when you need hundreds of prompts, daily history or several brands. Vendors such as Similarweb, Scrunch, HubSpot, SE Ranking and TopCited all sell this kind of tracking, and each publishes content promoting its own product. SparkToro’s advice applies to any of them: ask the provider how many runs per prompt it uses, how it handles variability, and whether it shows its math. Be wary of any tool that sells a single rank position in AI answers.

What should you avoid?

  • Judging from one answer. Rerun it. The result may change.
  • Tracking exact list position. Research shows it is the least stable part of an answer.
  • Mixing conditions. Different locations, memory settings or plans make your comparisons meaningless.
  • Changing the prompt set every month. You lose the trend.
  • Chasing every gap. Some are not yours to win.
  • Manufacturing mentions or publishing near-duplicate pages. Google warns against both.
  • Treating a citation as traffic. Bing is explicit that a citation is not a click.

Get a brand entity and structured data audit

If your log shows competitors ahead and you are not sure why, the fastest next step is an audit of the foundations. Apex Web Zone’s Brand Entity and Structured Data Audit reviews how consistently your business is described across your website and public profiles, validates your Organization, Article and Service markup against the visible page, checks crawler access for Googlebot, OAI-SearchBot and PerplexityBot, and runs a baseline prompt log against your closest competitors. You receive a prioritised fix list. We will tell you what is worth fixing and what is not.

[Request a Brand Entity and Structured Data Audit]([CTA URL: confirm contact or audit page]) or start with our free SEO audit report.

Frequently asked questions

Run a fixed set of buyer-style prompts several times in ChatGPT, Perplexity, Copilot and Gemini, log the brands named and the sources cited, and compare mention rate, citation rate and share of mentions. A single prompt is not evidence because AI answers vary from run to run.

Compare the same metrics for each brand across the same responses. A large gap that appears on several platforms and repeats month after month is a real signal. Small differences can be noise. Also list the domains cited beside competitors but not beside you, since those show where the gap comes from.

The manual log shows which competitors are named and which URLs are cited. The “why” comes from the source-gap analysis: whether it is their own content, third-party reviews and roundups, community discussions, or directory profiles. Neither Google Search Console nor Bing Webmaster Tools names competitors, so you need the log.

At least three, and five is better, across every platform you track. SparkToro suggested 60 to 100 runs for a precise read on one prompt, so treat smaller samples as directional and compare them month to month rather than relying on one reading.

Not by itself. Google says structured data is not required for its generative AI features. Schema helps machines read your entity facts and supports rich-result eligibility, so keep it accurate, but spend equal effort on content quality, crawler access and third-party mentions.

It can give you a guess, but not a record of what it did. Use the sources that actually appear in real answers to your prompts as evidence instead.

Monthly with the same prompt set is a sensible default. Add a mid-month spot check on your top five prompts after a major change, such as a site migration or a campaign.

  • Share:

Leave a Reply