How we build websites AI assistants can cite
When someone asks ChatGPT, Claude, Perplexity or Gemini about one of our projects, the answer should be right and should name the right studio.
This is the checklist every MarVikIT website follows, our own and our clients'.
What gets a site into AI answers
Besides what a model learned in training, an AI assistant answers from two places: search indexes its own crawlers build ahead of time, and pages it fetches live while someone is asking. To be cited, a site has to be reachable by those crawlers, say the same facts everywhere and answer real questions in plain text. No special file or markup does that for you; Google says the same about its own AI features.
Crawlers welcome, by name
Our robots.txt files allow every crawler, then list the AI crawlers by name so there is no doubt. Each AI company runs separate bots for separate jobs:
- Search and citation. OAI-SearchBot (ChatGPT search), Claude-SearchBot (Claude) and PerplexityBot (Perplexity) build the indexes those assistants cite from. Block them and you drop out of their answers.
- Live fetches. ChatGPT-User, Claude-User and Perplexity-User fetch a page when a person's question needs it.
- Training. GPTBot and ClaudeBot collect content for training future models. Google-Extended is a robots.txt token rather than a crawler: it controls whether Gemini may use your pages, and it doesn't affect Google Search.
On our own sites we allow all of them. For a client it's a decision we make together, because training and citation are separate choices. The vendors document their bots themselves: OpenAI, Anthropic, Perplexity.
robots.txt is only half of it. A hosting firewall or a "bot protection" switch can block the same crawlers without anyone noticing, so we check that too.
Plain facts, the same everywhere
Every page opens with sentences that can be quoted as they stand: what the product is, who it is for, what it does and what it costs. The same facts, in the same words, appear on the website, the store listings, the GitHub pages and llms.txt.
AI answers are put together from several sources at once. When the sources agree, the answer comes out right. When one page says "solo developer", another says "studio" and a third has an old phone number, the answer hedges or mixes them up. Consistent facts are the cheapest thing you can do for how assistants describe you.
Structured data that connects the sites
Every page carries structured data (Schema.org JSON-LD) that describes the studio as one Organization with a fixed ID, and each app with its own fixed ID. The product sites reuse exactly the same IDs, so search engines see one studio behind swapgenie.app, bigbands.app, exactcups.com and marvikit.com instead of four unrelated sites. The details match the stores too: an app listed under Food & Drink isn't described as a health app.
No markup guarantees a citation. Structured data earns its place because it removes doubt about who makes what, for search engines and assistants alike.
llms.txt, with honest expectations
Each of our sites has an llms.txt: a short plain-text file at the root that says what the site is, lists the key facts and pages, and ends with a "who makes this" section with example sentences of how to describe it correctly. You can read ours.
To be clear about what it does: llms.txt is a community proposal, not a standard, and none of the big assistants has said it reads the file when answering. We publish it because it costs almost nothing, it helps the tools and agents that do read it, and writing it forces us to state the facts in one place.
Light pages, real text
marvikit.com is plain HTML with one stylesheet and one script. On every site we build, everything that matters is in the page as text, not hidden in images or loaded later by JavaScript. Crawlers read it in one request, and people get a page that loads fast on a slow phone connection.
Clean URLs, clean 404s
Every page has one canonical URL and is listed in the sitemap. We verify each domain in Google Search Console with a DNS record, submit the sitemap, add Bing Webmaster Tools and read the reports.
When a page goes away, it redirects to a real replacement or returns a proper 404. We did exactly that here: when marvikit.com replaced an old WordPress theme, Google still had 31 old URLs on file, most of them the theme's demo pages. The five with a real equivalent on the new site now redirect to it, and the demo pages with no replacement return 404, which is what Google recommends. Recreating them, or sending them all to the home page, would only have added noise.
Measure, then adjust
We ask the assistants themselves, ChatGPT, Claude, Perplexity and Gemini, how they describe our projects. When an answer drifts from the facts, we find the page behind it and fix it. Search Console and the store numbers show the rest.
Want this for your site?
We build new websites this way and fix existing ones: Message on WhatsApp or write to hello@marvikit.com.