If you want to know how to track AI mentions of your company, do not start with vanity searches. Start with the real questions buyers ask before they buy. Then test the same prompts across the main AI tools, log the answers in one scorecard, and repeat the checks by country and language. That gives you something you can compare month to month.

The hard part is not running one prompt in ChatGPT. The hard part is making the checks consistent enough that a change in results means something. AI answers vary by wording, location, language, account state and whether the tool shows sources. A useful process controls what it can, records what it cannot, and gives you a short list of reasons why your company was named, ignored or misrepresented.

Start with the questions buyers actually ask - how to choose realistic prompts that show whether AI tools mention your company during buying research

Most companies begin with their own brand name. That tells you very little. If someone asks ChatGPT for your company by name, the model may simply repeat what it already knows or what it finds in obvious places. That is not buying research. It is a brand recall check.

The better test is to use prompts that match the stage before a prospect has chosen a supplier. Think like a buyer who knows the problem but not the vendor. Good prompts usually fall into four groups.

First, category discovery. For example:

  • “Which accounting firms in Cork help small manufacturers with R&D tax claims?”
  • “Best payroll software for restaurants in Poland”
  • “Who installs commercial heat pumps in Manchester for warehouses?”

Second, comparison. For example:

  • “Compare CRM systems for estate agents in the UK”
  • “Which German translation agencies specialise in medical devices?”
  • “What are the alternatives to [competitor name] for Nordic e-commerce brands?”

Third, local shortlist building. For example:

  • “Recommend family law solicitors in Dublin for custody disputes”
  • “Top B2B PR agencies in Lithuania for fintech”
  • “Who are the best Shopify developers in Sweden for multilingual shops?”

Fourth, trust and fit. For example:

  • “Which providers are good for small businesses, not enterprise?”
  • “Who offers support in Polish?”
  • “Which firms have experience with regulated industries?”

Those prompts surface whether AI tools mention you when buyers are actually narrowing options.

A practical way to build your prompt list is this:

  1. Pull your top commercial pages and service pages.
  2. Pull the queries already bringing impressions in Search Console.
  3. Add the phrases your sales team hears on calls.
  4. Add competitor names only where buyers genuinely compare options.
  5. Rewrite each query as a natural question in the buyer’s language.

Aim for 15 to 30 prompts at first. Fewer than that and you will overreact to noise. More than that and the process becomes too slow unless you automate it.

Keep the prompts realistic. Do not write things no buyer would say, such as “best excellent affordable award-winning provider”. Do not over-specify if buyers do not over-specify. If your market is multilingual, make separate prompt sets. A buyer in Germany asking in German is not the same as a buyer in English asking about Germany.

If you serve several countries, include location in the prompt where buyers would. “Solicitor” in the UK and Ireland is not just a word choice issue. The underlying market differs. The same is true for regulated sectors, where local compliance language matters.

Test each AI tool in a consistent way - how to run checks across ChatGPT, Gemini, Perplexity, Claude and Grok so results are comparable

Consistency matters more than volume. If one test uses a logged-in paid account and another uses a free anonymous session a week later, you are not comparing like with like.

For each round of checks, decide and document:

  • Which tool you are using
  • Which plan or account type
  • Which model or mode, if visible
  • Date and time
  • Country you are testing for
  • Language of the prompt
  • Whether web browsing or live search is enabled
  • Whether you used a fresh chat

Then keep those settings as stable as you can.

A workable manual process looks like this:

ChatGPT

Use a fresh chat for each prompt set. If web search is available in your account, note whether it was used. Copy the exact answer, not your summary. Record whether your company appears in the main answer, only after a follow-up, or not at all.

Gemini

Again, start fresh where possible. Gemini often blends model knowledge with current web information. Note if it provides direct links or named sources. Record whether the answer is a ranked list, an unranked paragraph or a mixed format.

Perplexity

Perplexity is often the easiest to inspect because citations are prominent. That does not make it the most important. Buyers use different tools. Record not only whether you were named, but which sources Perplexity used to justify the mention.

Claude

Claude may be more cautious in some categories and may provide fewer live citations depending on the mode available. Check whether it avoids naming companies at all for certain prompts, because that affects interpretation.

Grok

Grok can be more conversational and may pull in current web context differently. Record whether the answer is direct, speculative or clearly source-backed.

To keep results comparable, use the same prompt wording across all five tools in the same language and country batch. Do not “help” one tool by rephrasing the prompt until it gives you the answer you hoped for. If you want to test prompt sensitivity, do that as a separate exercise.

Two more rules help.

First, cap follow-ups. For example, allow one standard follow-up only: “Can you give me a shortlist?” or “Which companies fit best?” Then use that same follow-up everywhere. Otherwise you drift from measuring visibility into coaching the model.

Second, separate buyer research prompts from factual verification prompts. “Who are the best VAT advisers in Belfast?” is different from “What does company X do?” The first tests discoverability. The second tests brand understanding.

If you do this often, a platform that tracks mentions across ChatGPT, Gemini, Perplexity, Claude and Grok saves time. Seonis was built for exactly that, alongside local-language SEO work and reporting in the owner’s language. If you want to see how the platform is positioned, start with Seonis: SEO and AI visibility in your website’s language.

Do not rely on memory. Use a scorecard. A spreadsheet is enough at the start.

Your columns should include:

  • Test date
  • Tool
  • Country target
  • Prompt language
  • Prompt text
  • Follow-up used, yes or no
  • Your company mentioned, yes or no
  • Mention type
  • Ranking position
  • Exact wording
  • Link shown, yes or no
  • Link destination
  • Source or citation used
  • Confidence level
  • Notes on errors

For mention type, use a short controlled list:

  • Primary recommendation
  • Included in shortlist
  • Mentioned as alternative
  • Mentioned only after follow-up
  • Mentioned negatively
  • Not mentioned

For ranking position, keep it simple. If the answer is a numbered list, record the number. If it is a paragraph, use labels such as “first named”, “middle”, “last named” or “unranked”. Do not pretend a precise rank exists where it does not.

For exact wording, copy the sentence. This matters because AI tools may mention you with the wrong service, wrong geography or outdated description. “A software platform for retailers” is not the same as “an agency for manufacturers”. Both count as mentions, but one is commercially useful and the other is not.

For links, capture the destination URL if one is shown. It may link to your site, a directory profile, a review site, a news mention or a third-party article. That tells you where the model or search layer found enough confidence to name you.

For confidence, use your own internal scale. For example:

  • High - clearly recommended, accurate description, source-backed
  • Medium - named, but with thin support or minor inaccuracies
  • Low - vague mention, uncertain fit, weak or no visible support

That confidence score is not the model’s confidence. It is yours, based on the quality of the mention.

A simple scorecard lets you answer useful questions quickly:

  • Which tools mention us most often?
  • In which country and language do we disappear?
  • Are we named directly, or only after prompting?
  • Do citations point to our own site or to third parties?
  • Are we being described accurately?

Once you have a month or two of data, trends start to matter more than any single result.

Separate brand checks by country and language - why results differ across the UK, Ireland and non-English markets and how to track them properly

This is where many teams get misled. They run one English prompt from one laptop and assume the result represents Europe. It does not.

AI tools often behave differently by market because the underlying web evidence differs by market. Your English page may be strong, while your Polish pages are thin. Your Irish business citations may be good, while your UK ones are patchy. Reviews, directories, trade membership pages and local press mentions vary by country too.

The UK and Ireland are a good example. They share English, but not the same business landscape. Buyers use different terms. Local directories differ. Legal, financial and regulated services differ. A company visible in UK prompts may not appear in Irish prompts, even when the service is similar.

Now add non-English markets. If your website is in German, Swedish or Latvian, translated content often underperforms because it misses local phrasing. AI systems also pick up those weaknesses. If your pages sound machine-translated, lack local examples, or have weak internal linking, they are less likely to be surfaced confidently.

So split your checks by:

  • Country
  • Language
  • Service line
  • Buyer stage

For example, do not mix:

  • German prompts for Germany
  • English prompts for Germany
  • English prompts for the UK
  • English prompts for Ireland

Those are four different test groups.

Use native phrasing in each language. If you are not fluent, get a native speaker to review the prompt set. This matters because small wording differences can change which companies are named.

The same applies to your content and site structure. If you publish in the site’s own language, keep the internal linking and page intent native too. Generic translated clusters often look tidy in a spreadsheet and weak in the market. If you are fixing that side of the problem, this guide on internal linking for local language SEO pages is worth reading.

Find out why you are or are not being named - how to inspect citations, site content, third-party references and technical gaps that affect AI visibility

Once you know where you are and are not being mentioned, the next question is why.

Start with citations and visible sources. Perplexity often makes this easiest, but check all tools that show links. Ask:

  • Are they citing your homepage, service page or blog article?
  • Are they citing a directory, marketplace or review site?
  • Are they citing a competitor comparison page?
  • Are they citing a news article or association listing?

If your company is being named from third-party sources rather than your own site, that is not bad. It tells you external references are doing useful work. But it may also mean your own pages are not clear enough.

Then inspect your site content.

Look at the page that should earn the mention for that prompt. Does it clearly state:

  • What you do
  • Who it is for
  • Where you operate
  • What makes you relevant to that use case
  • Proof points, examples or sectors served

If not, the model has little to work with.

Next, check whether the page matches the buyer question. A page titled “Solutions” with vague copy is unlikely to support mentions for “commercial heat pump installers for warehouses in Leeds”. A specific service page is more likely to.

Then inspect third-party references. For local and B2B businesses, useful signals often include:

  • Industry directories
  • Chamber or trade association listings
  • Marketplace profiles
  • Review platforms
  • Local press coverage
  • Partner pages
  • Case study mentions on client sites

If competitors are being named and you are not, compare their footprint. You may find they have stronger category wording on-site, more consistent local citations, or more third-party references in the right language.

Do not ignore technical basics. AI visibility still depends heavily on what can be crawled, understood and corroborated on the web. Check:

  • Important pages are indexable
  • Canonicals are correct
  • Language targeting is clear
  • Internal linking points to key service pages
  • The site is not hiding key copy in scripts or tabs that render poorly
  • Titles and headings reflect the actual service and place
  • Contact and company details are consistent

Also check whether your strongest content is trapped on a platform page with weak crawlability or poor structure. That can happen on WordPress, Shopify, Webflow and Ghost if the site has been assembled without much thought for discoverability. The CMS itself is not the issue. The implementation usually is.

Finally, watch for false positives. Sometimes an AI tool names your company because your name resembles a category term, or because it misread an old article. That is not durable visibility. Log it, but do not celebrate it.

Set up a monthly process that saves time - how to turn one-off checks into a repeatable reporting routine using spreadsheets or tools such as Seonis

A monthly cadence is enough for most small and medium businesses. Weekly checks usually create noise unless you publish heavily or operate in a fast-moving market.

A practical monthly routine looks like this:

Week 1 - run the prompts

Use the same scorecard and the same prompt sets by country and language. Test your priority services first. If time is tight, rotate lower-priority categories every other month.

Week 1 - review changes

Compare results with last month:

  • New mentions gained
  • Mentions lost
  • Rank position changes
  • New citations appearing
  • Description accuracy improving or worsening

Week 2 - diagnose causes

For each gain or loss, look for likely reasons:

  • New service page published
  • Better internal linking
  • Fresh third-party mention
  • Competitor article outranking yours
  • Technical issue on key pages
  • Prompt wording drift
  • Tool behaviour change

Week 2 - assign actions

Keep the action list short. For example:

  • Rewrite two service pages in native language
  • Add internal links from five related pages
  • Claim or improve two directory listings
  • Publish one comparison article
  • Fix indexing on one country page

Week 3 and 4 - implement and document

Update the scorecard with what changed on your side. That gives context next month. Otherwise you will forget whether a visibility gain followed a content change, a citation gain or nothing obvious.

If you manage this in a spreadsheet, create separate tabs for:

  • Prompt library
  • Monthly results
  • Citation review
  • Actions taken
  • Country and language coverage

If you manage several client accounts as an agency, standardise the template across all of them. Agencies serving multilingual SMEs may also want a system that handles the checks, content workflow and reporting together. Seonis does that with native-language keyword research, article production, publishing and AI mention tracking in one place. Agencies can also review the Partner programme for agencies if they need a cleaner operational model.

One final point. Tracking AI mentions is not the same as owning the result. These systems change. A mention can disappear even when your site improves. That is why the process has to be grounded in evidence you can inspect - prompts, wording, citations, links and page quality - not just a single headline score.

The companies that get value from this are not the ones obsessing over one screenshot from ChatGPT. They are the ones running the same realistic buying prompts every month, in the right countries and languages, then fixing the pages and references that the answers reveal. That is how you move from curiosity about AI visibility to something you can actually manage.