Notes from the work.

Documented by the platforms

• Google’s guidance for its AI features says site owners don’t need new machine-readable files, AI text files or special schema markup; existing search best practices apply.
• Crawler controls separate AI training from search. OpenAI’s OAI-SearchBot is for ChatGPT search, while GPTBot collects content for model training, and OpenAI says each setting is independent. Anthropic separates Claude-SearchBot from ClaudeBot in the same way. Google-Extended controls use in Gemini training and grounding and, Google says, does not affect inclusion or ranking in Google Search.
• So blocking a training crawler is a different decision from blocking a search crawler. Making one without understanding the other can quietly remove a site from AI search answers.

• Google Search Central: AI features and your website (updated 10 Dec 2025) (external site)
• Google: common crawlers, Google-Extended (updated 15 Oct 2026) (external site)
• OpenAI: overview of OpenAI crawlers (external site)
• Anthropic: how site owners can control Anthropic’s crawlers (external site)

Found in one study

• In a 2024 paper on generative-engine optimisation, the authors’ methods raised visibility in generative-engine responses by up to 40% on their own benchmark, and how well each method worked varied by subject area. “Up to” is the best case on that benchmark, not a forecast for any particular site.

• Aggarwal et al., GEO: Generative Engine Optimization (arXiv 2311.09735, accepted to KDD 2024) (external site)

What we do

• On Strange Forms we record every search and AI-discovery choice with the strength of its evidence: established practice, a study result, or a hypothesis. Facts are stated in text, structured data is checked against the visible page by the build, and there is no llms.txt and no FAQ markup.
• Hypotheses we act on only because they also help readers: consistent names for the same thing everywhere, and sections that make sense when read on their own. Whether they change AI answers is untested.

What we built, and why

• Ask only what applies. A brief starts from the project type, which decides the sections suggested first, and conditional questions appear only when an earlier answer makes them relevant.
• Never guess. A question is shown only when its rule is definitely true. Rules may only refer to earlier questions, and broken rules are flagged in Review before anything is exported.
• Remove friction for the client. The brief goes out as a fillable PDF that opens in any PDF reader: no account, no portal.
• Keep it private by default. Briefs stay in the freelancer’s browser unless they choose the optional cloud storage; backups are files they keep themselves.

Limits

• PDF viewers can’t reliably hide fields, so conditional questions appear in the PDF with a short note instead of disappearing. And because local briefs live in one browser, backups matter.

An example from Małgosia’s work

• For Messageworks, Małgosia automated a 5,000-page catalogue using Illustrator scripting and signage colour logic, reducing the workflow from weeks to hours. This is individual project experience, not an AX Designers company result.

How we decide (our approach, not a research finding)

• Rules: structured inputs and explicit logic, such as layouts, colour rules, naming and file conversion.
• AI: varied or unstructured inputs, such as free text, mixed documents and first drafts, with a person reviewing the output.
• People: decisions with consequences, exceptions, and anything a customer sees first.