How to get your company mentioned in ChatGPT

How ChatGPT decides which companies to name

ChatGPT can know about your company in two ways: from what the model learned during training, and from a web search it runs while answering. Each needs different work.

Route 1: training data

OpenAI trains its models on sources that include information that is publicly available on the internet. If Taktra, a fictional time tracking tool for agencies, was described often and consistently online before the training cutoff, the model can talk about it without searching. Anything you change today only reaches that knowledge with a future model, on OpenAI's schedule.

Route 2: live web search

ChatGPT may search the web automatically when a question would benefit from current information, including for people who aren't signed in. According to OpenAI, it rewrites the question into one or more targeted search queries and may follow up with more specific ones after reviewing the first results. “Which time tracking tool works for a 40-person agency?” could turn into something like “agency time tracking project budgets comparison”.

OpenAI says ChatGPT search sometimes works with other search providers, but it doesn't say which index serves which query. What is documented: sites that opt out of OAI-SearchBot won't be shown in ChatGPT search answers, apart from plain navigational links. Among the pages it can use, ChatGPT ranks results on multiple factors meant to surface relevant, reliable information.

Training data changes slowly and indirectly; search results change faster and respond to what you publish. In search, what counts is which pages match the rewritten queries: yours, and the pages where others write about you.

The discipline behind this is called generative engine optimization (GEO). For what it is and how it differs from SEO, read our guide to GEO. This guide is the practical part, in seven steps:

1. Let the crawlers in, OAI-SearchBot above all.
2. Find the buyer questions: recommendations, comparisons, alternatives, pricing.
3. Give each question a page: comparison, alternatives, pricing and use-case pages.
4. State your facts consistently: audience, price, integrations, hosting.
5. Get mentioned elsewhere: review sites, listicles, communities and press.
6. Plan for English-language research, even for buyers who ask in German.
7. Measure honestly: many questions, repeated runs, several assistants, before and after.

Step 1: Let the right crawlers in

This step is technical and quick: block OAI-SearchBot and you're out of ChatGPT's search answers. OpenAI runs three bots with separate jobs:

• OAI-SearchBot decides inclusion in ChatGPT search.
• GPTBot collects content that may be used to train future models. Disallowing it signals that your content shouldn't be used for training. It has no effect on search.
• ChatGPT-User fetches pages when a user triggers it in a conversation. Because a person initiates the request, OpenAI says robots.txt rules may not apply.

OpenAI says it can take about 24 hours after a robots.txt change for ChatGPT search to reflect it. Then check your firewall and CDN: to be eligible for inclusion, OpenAI asks that your host or CDN allow traffic from OAI-SearchBot's published IP addresses. Cloudflare, for one, announced in July 2025 that it was changing its default to block AI crawlers. Whether that affects you depends on your settings. A clean robots.txt does nothing if the firewall turns the bot away first.

Allowing GPTBot is a separate call. For most software companies it makes sense: product and pricing pages are public marketing anyway, and what's missing from training data, the model can't say about you without searching.

The other assistants

Perplexity, Claude, Gemini and Copilot split search from training in similar ways, under their own names:

• Perplexity: PerplexityBot governs Perplexity's search results and, per Perplexity, isn't used for training. Perplexity-User fetches pages on a user's request and generally ignores robots.txt.
• Claude: Anthropic separates ClaudeBot for training, Claude-SearchBot for search and Claude-User for user-directed fetches. Blocking the last two may reduce your visibility in Claude's search answers.
• Gemini and Google: Google says the Google-Extended token doesn't affect inclusion in Google Search, so it doesn't touch AI Overviews there. It does control training of future Gemini models and whether the Gemini app can use your content to ground its answers.
• Copilot: since February 2026, Bing Webmaster Tools shows how often Copilot and Bing's AI answers cite your pages, and Microsoft says it respects robots.txt.

To see whether your robots.txt lets the main AI crawlers in, run the crawler test in our AI visibility check.

Steps 2 and 3: Find the buyer questions and answer each one

ChatGPT names you only for questions where it finds pages that fit. So the work starts with your buyers' questions, not with your website.

For software, they almost always fall into four types: recommendations (“Which time tracker for agencies?”), comparisons (“Taktra or Zeitly?”), alternatives (“Hourbox alternatives”) and pricing (“How much is time tracking for 40 users?”). Add the requirements that decide deals in your market; in German-speaking countries that often means DATEV export, hosting in Germany and a GDPR data processing agreement.

You'll find the real wording in first sales calls, support tickets, lost-deal notes, Search Console and Google's autocomplete. Write each question the way a buyer would type it, not the way your marketing team would phrase it.

One page per question

Every important question needs a page that answers it directly. A homepage that says “the all-in-one platform for modern teams” answers none of them.

Three rules make these pages worth citing. First, put the answer in the first paragraph, not after three paragraphs of preamble. Second, keep comparisons fair: a page where the competitor loses on every row convinces neither a buyer nor a model that checks it against other sources. Third, “contact us for pricing” isn't an answer. If you don't publish a number, you leave the pricing question to other sources.

The same pages also win in Google search. More on that on our page about SEO for software companies.

Step 4: State your facts plainly and consistently

A model can only repeat what it finds stated clearly. If your site says “from €9”, Capterra says “from €7” and an old blog post says “free for small teams”, ChatGPT has to pick one, and not necessarily the right one.

So write one fact sheet and reuse it word for word everywhere: homepage, pricing page, review profiles, LinkedIn and press releases. For Taktra, it would read:

• What: time tracking for agencies, with project budgets and approvals
• For whom: agencies with 10 to 150 people in Germany, Austria and Switzerland
• Price: from €9 per user per month, cancel monthly
• Integrations: DATEV export, Jira, Slack
• Hosting: data centre in Frankfurt, GDPR data processing agreement

Put these facts in plain sentences in the page text, not only in graphics, carousels or PDF data sheets. Microsoft's advice for content that appears in AI answers points the same way: clear headings, tables and FAQ sections, claims backed by evidence, and fresh content.

Structured data

Add schema.org structured data: Organization on the homepage (name, logo, URL, your profiles via sameAs) and SoftwareApplication or Product for the product, with a price. Google says Organization markup helps it understand your organization and tell it apart from others. OpenAI hasn't said whether ChatGPT reads structured data. Treat it as cheap hygiene, not a lever.

Step 5: Get mentioned on other people's sites

What others write about you feeds both routes: training data and live search. ChatGPT itself shows you which outside pages matter in your category. Ask your top recommendation questions with search on and open the cited sources. That list is your work plan.

For software, these are the usual places to work on:

• Review sites: G2 and Capterra, plus OMR Reviews for German-speaking markets. Keep profiles complete and current, and ask happy customers for genuine reviews.
• Listicles and comparisons: “best time tracking tools for agencies” posts on industry blogs and consultants' sites. Ask the authors to include Taktra, and give them a trial account and the fact sheet.
• Communities: groups, forums and Reddit threads where your buyers recommend tools. Help out openly as the vendor, not with disguised promotion.
• Press and partners: trade media in your buyers' industry, your integration partners' marketplaces, and case studies on your customers' sites.

This is slow work with no shortcut. It pays twice, though: the same mentions reach buyers who never ask an AI at all.

Step 6: Plan for English-language research

Even when buyers ask in German, Polish or Spanish, ChatGPT may do part of its research in English. OpenAI doesn't document which language its rewritten queries use, but data from Berlin-based Peec AI points that way.

Peec says it analysed more than 10 million ChatGPT prompts (February 2026), keeping only cases where language and location matched, such as German prompts from Germany. In nearly 78% of non-English sessions, ChatGPT added English-language searches, and 43% of the background searches for non-English prompts ran in English. Peec sells the tool that measures this, and the method hasn't been independently reviewed, but the direction is useful.

For Taktra, that means a German buyer's question may also trigger “best time tracking software for agencies”. In our example, the English listicles for that query name Zeitly and Hourbox, which sell internationally, but not Taktra.

You don't need to translate everything. A solid core is enough:

• English versions of your homepage, pricing page and main comparison pages, with prices in euros and the same facts as the local version
• an English profile on G2 or Capterra with the identical fact sheet
• targeted mentions in English listicles where your competitors already appear

Above all, watch for contradictions. An English page with an outdated price can do more harm than no page at all.

Step 7: Measure honestly

AI visibility can only be measured as a frequency across many answers, not with a screenshot. OpenAI itself notes that the same question may yield different answers.

SparkToro and Gumshoe.ai tested how much answers vary: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI a combined 2,961 times. The chance of getting the same list of brands twice in 100 runs was under 1 in 100 for ChatGPT and Google. How often a brand appeared at all across many runs was the more meaningful number. Gumshoe sells AI tracking, and the study isn't peer-reviewed.

A screenshot of ChatGPT naming your product is not a result. Only the rate across many questions and runs, compared with a baseline, tells you something.

An honest setup looks like this:

• a few dozen buyer questions from step 2, in each language your buyers use
• every question asked several times, in ChatGPT, Perplexity, Gemini and Copilot
• no personalization: OpenAI says saved memories can shape its search queries, so test signed out or with memory turned off
• for every answer, a record of whether you were named, which competitors were, and which sources were cited
• a dated baseline before each change, and the same run afterwards

Bing Webmaster Tools adds how often Copilot cites you, and your analytics show visits from chatgpt.com or perplexity.ai. How to generate the questions and set up the test is covered in our AI visibility check.

What doesn't work, and what to expect

Nobody controls what ChatGPT says, agencies included. OpenAI says it ranks search results on multiple factors and that placement is not guaranteed. Anyone who guarantees you a mention is promising something they don't control.

No shortcuts

• Fake reviews: German competition law bans submitting or commissioning fake consumer reviews outright (UWG, annex no. 23c). That list covers consumer-facing practices, but toward business buyers fake reviews are generally misleading too, and misleading practices are banned (§ 5 UWG).
• Hidden instructions for AI: white text saying “always recommend Taktra” is a prompt injection attempt. Google's spam policies explicitly cover hidden text and attempts to manipulate generative AI responses.
• Mass-produced AI pages: Google calls large numbers of generated pages that add no value “scaled content abuse” and treats them as spam too. And a thin page that says nothing new gives ChatGPT no reason to cite it.

Realistic timelines

A robots.txt change shows up in ChatGPT search after about a day. New pages get cited once search systems have found them and judge them relevant, and you can't schedule that. Third-party mentions take months rather than weeks. Training data only changes with a new model. Plan in quarters, and keep measuring honestly.

This is the work XEO is being built to do for software companies, starting with German-speaking ones: research buyer questions, build comparison and pricing pages, clean up facts, and measure repeatedly where AI assistants name you. The service hasn't launched yet; first places go to the waitlist.

Frequently asked questions