Improve your AI visibility: how to get found on ChatGPT

How the AI platforms search for brands

When ChatGPT, Gemini, Perplexity and Claude search the web for an answer, Achtung.app stores their search queries and the sources they find. That shows how differently they go about it.

1. 1

   Only ChatGPT searches for brands it already knows

   ChatGPT often writes brand names into its very first searches, before any result is back, although the question names no brand. Gemini, Perplexity and Claude hardly ever do.

   In at least 1,066 of 1,721 answers, ChatGPT's first round of searches already contained vendor names. A search query contained a vendor name in 10 of 1,677 Gemini answers, in 21 of 1,722 Perplexity answers and in none of Claude's 240. ChatGPT also searches a lot: 7.2 search queries per answer on average, partly in English, and the question's exact wording in only 8 of 1,721 answers.

2. 2

   Only ChatGPT runs searches on the brands' own websites

   With searches like "site:vendor.com pricing", ChatGPT checks offer and prices, mostly for brands it had already searched for before the first result.

   "site:" searches on vendor websites ran in 1,210 of 1,721 ChatGPT answers. Of 4,916 such checks, 2,906 targeted a brand ChatGPT had already searched for before the first search result; for the rest, the brand name was not found in the first round of searches. The industry reports show a similar ratio (137 of 227). About a third of the checks asked for prices or plans. Gemini, Perplexity and Claude ran no "site:" searches, neither in self-tracking nor in the industry reports.

3. 3

   Perplexity, Gemini and Claude search for the topic of the question

   Perplexity uses the question word for word as its search query. Gemini and Claude write their own search terms on the topic.

   Perplexity searched the exact question in all 1,722 answers and added its own terms in only 25. Gemini ran 3.3 search queries per answer on average and used the exact question in 364 of 1,677 answers; Claude ran 1.2 on average and used it in 26 of 240. In the industry reports, Perplexity searched the exact question in 165 of 165 answers, Gemini in 65 of 164 and Claude in 49 of 165.

4. 4

   In the results: lists, comparisons and, for ChatGPT, Reddit and Wikipedia

   For ChatGPT, Perplexity and Claude, the search returns best-of lists and comparison pages in most answers. Reddit and Wikipedia turn up almost only for ChatGPT.

   At least one returned address with words such as "best", "top-10", "comparison" or "alternatives", which usually point to lists or comparisons, appeared in 1,545 of 1,721 ChatGPT answers, 1,473 of 1,722 Perplexity answers and 212 of 240 Claude answers. In the industry reports the figures were 111, 104 and 111 of 165 each. Reddit was among ChatGPT's results in 1,639 and Wikipedia in 1,454 of 1,721 answers, in none of Perplexity's, and for Claude only Wikipedia, in 7 of 240. Gemini only gives the domain of its sources: Reddit was among them in 104 and Wikipedia in 15 of 1,677 Gemini answers.

What you can do: four steps

The steps follow what the search queries show. Each one says which platforms it mainly targets.

1. All platforms

   1. Appear on the pages the search returns

   Make sure comparisons, guides, trade media and forums covering your customers' questions name your brand and its offer in the text.

   Ask typical customer questions in ChatGPT, Gemini, Perplexity and Claude yourself and look at the sources. Ask the publishers of lists and comparisons to include you, give them facts they can check, or offer an expert article with a real occasion. Make sure your brand name appears in the text with a clear description, not only as a link. On ChatGPT, Reddit and Wikipedia are also among the results in most answers. Genuine contributions are what counts there: Reddit can treat repeated self-promotion as spam, and Wikipedia discourages editing about your own company and requires disclosure for paid contributions. Because ChatGPT partly searches in English, English-language pages are worth a look too.

2. Mainly Perplexity, Gemini and Claude

   2. One page per customer question

   Perplexity, Gemini and Claude search for the topic of the question. A page whose title picks up the question and whose first paragraph answers it can be found there and cited as a source.

   Write one page for each important customer question, in the language of the question. That means different questions, not many near-identical pages for variations of the same one. Answer it in the first paragraph in your own words, with specifics instead of marketing language. A visible FAQ block covers related questions. According to Google, you don't need a special writing style for AI. Being cited as a source doesn't mean the answer names your brand, so state clearly on the page what your brand offers for that question.

3. Mainly ChatGPT

   3. Describe your offer so it can be checked

   What you offer, for whom and at what price should be in the text on your website, without a login and not only in a PDF.

   When ChatGPT looks up your website with "site:", it searches for your offer and facts, and in about a third of cases for prices. If that only exists in images, downloads or behind a form, it is harder for the search to find. Also keep your name, offer and location consistent across your website, profiles and other sites so the brand is identified correctly.

4. All platforms

   4. Let the search crawlers in

   Check in your robots.txt that OAI-SearchBot (ChatGPT), PerplexityBot (Perplexity), Claude-SearchBot (Claude) and Googlebot (Gemini) may fetch your pages.

   These crawlers collect pages for the platforms' web search; Gemini uses Google Search for it. On top of that, ChatGPT-User, Perplexity-User and Claude-User fetch pages when someone asks for them in a chat. If you block these crawlers, your pages can be missing as a source. GPTBot and ClaudeBot are different crawlers: they collect content for training models, and blocking them doesn't affect web search. Google-Extended is different: according to Google's crawler documentation, this robots.txt token controls training and also whether your site's content may be used to ground answers in the Gemini apps and on Vertex AI. It does not affect inclusion or ranking in Google Search. Important content should also be in the HTML and not loaded later by JavaScript.

What doesn't help: measurements and Google's guide

Plenty of shortcuts are sold around AI visibility. Achtung.app measured two of them itself; Google's guide for its own AI search features covers the rest.

• llms.txt: Achtung.app compared more than 4,000 brands and found no notable difference: brands with an llms.txt are named 2.5 times per industry report on average, brands without one 2.3 times. The small lead is explained by the brands' size, not by the file. The figures are in the llms.txt analysis.
• Schema markup as an AI lever: across two measurement waves, most recently 2,343 brands in seven industries, brands with schema markup on their homepage are not named more often in any industry. It remains useful for rich results on Google. The figures are in the schema markup analysis.
• Rewriting text for machines or chopping it up: According to Google, you don't need to rewrite text for machines or chop it into chunks.
• Artificially seeded mentions: According to Google, they help less than they seem to.
• Many near-identical pages for variations of the same question: Google treats them as spam when they mainly aim to influence rankings or AI answers.

The methodology page summarises Google's guide.

How to check whether it works

Record how often your brand is named before each change, for each platform separately, and compare after a few weeks.