The number that puts everything in perspective
AI answers fluctuate. In late 2025, SparkToro ran 2,961 queries against three AI assistants and published the results in January 2026: for an identical question, the same brand list came back in under one percent of cases (reports on it: Search Engine Land and Search Engine Journal). Anyone who sells you a fixed "position in ChatGPT" is selling you a number that looks different on the next call.
What can be measured is something else: whether your brand shows up at all in answers to purchase-relevant questions, how often compared to competitors, and which sources the systems cite while doing so. A share in a sample, not a position. Everything that follows stays within that boundary.
Tonight: Ask the way your customers ask
The first step costs nothing and takes ten minutes. Open a new chat and ask the way someone who does not know you yet would ask: for a provider of what you sell, in the market where you sell it. Three to five questions, in ChatGPT and in Perplexity.
You will then see three things: whether you show up, who shows up instead, and which pages are named as sources. That is your starting point. Note the date.
Tonight: Watch for whether a source is given
A pattern we noticed in a customer conversation: as long as the answer comes from the model's memory, no source is given underneath it. Sources only appear once the system triggers a web search. That is exactly where your lever is. What a model learned during training, you cannot change. What it finds during a search, you can.
That is why our measurements only count answers that triggered a web search. Everything else is memory, and nobody can edit that.
This week: Let the right crawlers in
The systems that cite sources send their own crawlers. They are called GPTBot and OAI-SearchBot, ClaudeBot and Claude-User, PerplexityBot and Perplexity-User, Google-Extended. Many sites block them without knowing it, because a firewall or a plugin blanket-blocks anything that looks like a bot.
Have someone check whether these names get through on your site. We measured exactly this case with our own web check: pages that a human sees without any problem answered the machine with a block page. Whoever does not get in does not get cited.
This week: State who you are, in a file
There are two things a machine can read without interpreting your page. An llms.txt under your domain: a heading, a short paragraph on what you do, then facts and links, no marketing. And an Organization schema in the head of your page: name, location, founders, contact, services. Whatever is in there has to also appear on the visible page.
Both take an hour to build. We only caught up on both for artheris.de last week, with person entries for both founders. It was overdue.
On llms.txt, as of 28 October 2026: for Google Search no benefit is documented, for agents it is intended. Trade press reports that Google's guidance lists it among the measures you can skip for Google Search, and that Lighthouse 13.3 checks it in an experimental category called Agentic Browsing (reports: TechWyse and Search Engine Journal).
Before your next relaunch: Give machines Markdown
This is something we have built for artheris.de ourselves. Cloudflare introduced a feature that lets a client request the Markdown version of a page on demand, instead of HTML. In February 2026, Cloudflare put the savings, using one of its own blog posts as an example, at around 80 percent fewer tokens (report on it: The New Stack). Coding agents like Claude Code already ask for pages this way today.
Why that matters: an agent that can read your page for a fifth of the effort is more likely to read it in full, and it reads the facts instead of the navigation. For artheris.de, this exists as a second version of the same address, for example /index.md for the home page. If your site sits behind Cloudflare, it is a toggle. If not, it is an afternoon for your developer.
From now on, always: Date everything
Every question you ask gets a date, the exact wording, the system, and the answer. A screenshot in a folder is enough at the start. In three months you will want to know whether anything has moved, and without a date and exact wording you are comparing two things that have nothing to do with each other.
And do not draw a line from a single run. In our product, we deliberately do not show a trend curve from one data point, because a horizontal line drawn from a single measurement claims a stability that was never measured.
What I am not claiming
That your brand will be recommended after these steps. Nobody can promise that, because the systems do not belong to any of us. What I am claiming: after this, the systems can find you, read you, and cite you. Before, they often could not. The rest is measuring, changing, remeasuring, with a date.
And now, the self-promotion
There is no getting around this part, and I will keep it short. You do not have to build two of the steps above yourself. The starting point, that is, the question of whether your brand shows up in AI answers, runs on artheris.de as a basic check with no email required: enter your domain, choose a market, pick a buyer question, wait half a minute. The check asks AI models from OpenAI and Perplexity with web search and gives you a dated sample. And anyone who requests the report on top of that gets a ready-made change proposal for their site delivered with it, in the simplest case the llms.txt, ready to review and add.
If you get stuck rebuilding any of this, write to me: mail@artheris.de.