How AI assistants find and quote a website
An assistant knows about your company in two ways. The first is training: text collected months or years ago that shaped the model. The second is live search: when a question needs current information, the assistant searches, fetches a few pages and writes an answer with links to its sources. A citation with a link to your site almost always comes from the second route.
Each provider uses separate bots for these jobs, and the difference matters when someone edits your robots.txt file. These are the names and roles the providers publish:
| Provider | Bot | What it does | If you block it |
|---|---|---|---|
| OpenAI | GPTBot | Collects content that may be used to train models | Your content should not be used for training |
| OpenAI | OAI-SearchBot | Indexes pages for ChatGPT search | Your site is not shown in ChatGPT search answers |
| OpenAI | ChatGPT-User | Visits a page when a user's request needs it | OpenAI says robots.txt rules may not apply, since a user started the visit |
| Anthropic | ClaudeBot | Collects content for training | Your future content is excluded from training |
| Anthropic | Claude-SearchBot | Crawls to improve Claude's search results | May reduce your visibility and accuracy in search answers |
| Anthropic | Claude-User | Fetches a page when a Claude user asks a question | Claude cannot retrieve your page for that user |
| Perplexity | PerplexityBot | Surfaces and links websites in Perplexity search; not used for model training | You do not appear in Perplexity search results |
| Google-Extended (a robots.txt token) | Controls use for Gemini training and grounding | No effect on Google Search, which includes AI Overviews and AI Mode |
Two consequences follow. Blocking training bots and allowing search bots is a legitimate choice: you stay quotable without feeding training sets. And blocking everything "to protect our content" also removes you from the answers. Sources: OpenAI crawler documentation, Anthropic's crawler help page, Perplexity's bot documentation and Google's list of common crawlers, all read on 17 October 2026.
ChatGPT search also uses third-party search providers, and OpenAI's help center names Bing among them. Microsoft Copilot draws on Bing too. So a site that Bing has never indexed starts with a handicap in two assistants, which is why Bing Webmaster Tools belongs on the list next to Google Search Console.
What Google says, and what it means in Romania
Google's AI Overviews are available in Romania and in Romanian, according to Google's own help page. AI Mode, the conversational search mode, was announced for Romania and the Romanian language on 8 October 2025. For a Romanian company, AI answers are therefore already part of ordinary Google searches in its own language.
Google's guidance for site owners is unusually direct: "There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." It adds that you do not need new machine readable files, AI text files or special structured data. A page has to be indexed and eligible to show a snippet, and the usual basics apply: allow crawling, link pages internally, keep important content as text, make structured data match what is visible. Source: Google Search Central, AI features and your website.
What actually helps
1. Let the right bots in, and check that they really get in
Decide on purpose which bots you allow, then write robots.txt to match. Google's guidance also says to make sure the CDN or hosting allows crawling: bot protection at that layer can turn away crawlers that robots.txt welcomes. The server logs show whether OAI-SearchBot or Claude-SearchBot ever arrive.
2. Put the facts in the HTML, as text
Services, sectors, locations, contact details and prices or price logic belong in the page itself, not in an image, a PDF brochure or a widget that loads later. We cannot count on every crawler running scripts the way a browser does, so the safe choice is pages that arrive complete from the server.
3. One page per question your buyers ask
An assistant quotes the page that answers the question best. A single "Services" page with six bullet points answers nothing in particular. One page per service, opening with a definition (what it is, for whom, what it is not), gives the assistant a sentence it can quote and gives a reader a reason to stay.
4. Keep your company facts identical everywhere
Company name, legal entity, address, what you do and for whom should read the same on the website, the Google Business Profile, LinkedIn and the directories you appear in. When sources disagree, an assistant either picks one at random or leaves you out.
5. Add structured data that matches the page
JSON-LD markup (Organization, Service, Article, FAQPage) tells machines which entity the page describes and how pages relate. Google says it is not required for AI features, so treat it as a clarity aid, not a trigger. It must say exactly what the visible text says.
6. Write in both languages properly
If you sell in Romania and the EU, each language needs real pages with correct hreflang links, not a machine translation pasted under a flag. People ask assistants in Romanian, and the answer can only quote what exists in Romanian.
7. Treat llms.txt and Markdown versions as cheap extras
llms.txt is a proposal from 2024 for a Markdown file that summarizes a site for AI agents. It costs little to add and can help agents that look for it. Google says no AI text file is needed, and we have no evidence that it changes whether an assistant cites you.
How this site is built
The site you are reading follows the list above. robots.txt allows the search bots and user fetchers of OpenAI, Anthropic and Perplexity, and also the training bots, by our choice. Every page has a Markdown twin: add .md to its address. We publish llms.txt and llms-full.txt, generated from the same source as the sitemap, and each page carries JSON-LD for the organization, the author, the article and its FAQ. We do not yet have citation data to report, so we describe the setup, not results.
What does not work
| Tactic | Why it does not work | Risk |
|---|---|---|
| An "AI SEO" package that guarantees citations | No one outside OpenAI, Anthropic or Google controls their answers, and answers change between runs and users | Money spent on a promise nobody can keep |
| Keyword stuffing for prompts ("best ERP consultant Romania" twenty times) | Assistants paraphrase meaning, they do not match strings; Google treats keyword stuffing as spam | A page people cannot read and a possible ranking penalty |
| Hidden text or a separate version for bots | Google's spam policies name hidden text and cloaking explicitly | Removal from Google, which also feeds AI Overviews and AI Mode |
| Hundreds of generated pages, one per possible question | Google calls this scaled content abuse; thin pages are rarely the best source | A site that looks less trustworthy to readers and machines |
| llms.txt on its own | It is a proposal; Google says it is not needed; it cannot fix weak pages | False confidence |
| Blocking every AI bot, then expecting citations | OpenAI says sites opted out of OAI-SearchBot are not shown in ChatGPT search answers | Invisible in the answers you wanted to be in |
What can and cannot be promised
Generative engine optimization (GEO) sold without clear expectations does two kinds of damage. It burns budget on a promise of citations, and when it does bring visibility for vague prompts, it brings people who do not fit: the wrong size, the wrong sector, the wrong problem. Being precise about who you serve, and saying who you do not serve, filters for assistants and readers alike.
An honest supplier can promise that the site is crawlable by the bots you chose, indexed in Google and Bing, built from pages that each answer a real question, consistent in its facts, and measured. An honest supplier cannot promise that ChatGPT will name you for a given question, a position in an answer, or a number of leads.
How to measure it
- Write down 10 to 20 questions your buyers actually ask, in Romanian and English. Once a month, ask them by hand in ChatGPT, Claude, Perplexity and Google, and note whether you are named, whether there is a link, and whether the description is correct. Answers vary from run to run, so watch the trend, not one result.
- In Google Analytics 4, look at referral traffic from chatgpt.com, perplexity.ai, claude.ai, copilot.microsoft.com and gemini.google.com, and group them into one channel. Some assistant visits arrive without a referrer and show up as direct traffic.
- In Google Search Console, clicks from AI Overviews and AI Mode are counted inside the Performance report, under the Web search type. Google does not report them separately.
- In Bing Webmaster Tools, the AI Performance report (public preview since February 2026) shows how often your pages are cited in Microsoft Copilot and Bing's AI summaries.
- Ask every new lead where they first heard of you, and record the answer in the CRM.
Why a redesign is the right moment
During a redesign you already review every page, decide the URL structure, set up redirects, choose how pages are rendered and pick a CMS. Making the site quotable adds little to that work; retrofitting it later means opening all of it again. The CMS choice decides whether your team can keep pages accurate without calling a developer, and the week-by-week redesign plan shows where the content inventory fits.
The hard part is rarely code. It is deciding what the company does, for whom, and with what proof, then writing it down so a stranger, human or machine, gets it right on the first read. That is business work before it is web work, which is why we start with a diagnosis and treat the website as one entry point into a wider change, built through our Product Development service. If you are still choosing who should do it, see freelancer, agency, SaaS or studio and how to choose a digital transformation partner.
Summary
- There is no ChatGPT optimization separate from a site that is easy to crawl, read and quote.
- Allow search bots on purpose; blocking OAI-SearchBot removes you from ChatGPT search answers.
- Google says AI Overviews and AI Mode need no special files or markup, and both are available in Romanian.
- Index the site in Bing as well: ChatGPT search and Copilot draw on it.
- Nobody can guarantee a citation. Measure with hand-tested questions, GA4 referrals, Search Console and Bing Webmaster Tools.
