ChatGPT and other search products using Artificial Intelligence (AI) provide composed answers instead of displaying only a list of links. That creates a second layer of visibility alongside traditional search for website owners, publishers, and content creators. Appearing there requires clear information, but not a magical optimization formula.
Why AI visibility matters
In traditional search, pages compete for positions in a results list. AI services can combine several sources into one response, then mention or link to only some of them. An article that ranks well on Google can therefore remain absent from a ChatGPT answer.
This shift has produced Generative Engine Optimization (GEO), meaning the optimization of content for generative answer systems. It resembles Search Engine Optimization (SEO), which aims to make pages discoverable in conventional search engines. A new company directory from St. Gallen-based Cohaga, for example, is intended to measure how Swiss companies appear in answers from ChatGPT, Claude, and Gemini.
The service, called ranQ, covers more than 600,000 Swiss companies according to its website. It uses data from the Swiss commercial register, obtained through the Zefix central company index and reconciled daily, along with classification data from the Federal Statistical Office and publicly available company information. Verified companies can edit their profiles, but the report provides no pricing or detailed privacy terms.
The search systems themselves also change constantly. Data from the provider Promptwatch suggests that ChatGPT Search began restricting many more searches to particular websites in August 2026. The observed share of these queries rose from 0.3–0.5% to 16–17%, but Simon Willison cautions against overinterpreting the numbers: They cover only prompts under automated monitoring, while OpenAI’s internal system instructions remain private.
What makes content readable for AI
A Large Language Model (LLM) is a system that processes and generates language using statistical patterns. According to the lead report, such systems do not necessarily handle a web page as one complete article. They can divide documents into chunks, meaning smaller sections of text, and select individual passages for an answer.
That makes each paragraph more significant. The report uses “LLM readability” to describe how easily a text can be separated into self-contained, understandable sections. A paragraph that makes sense only after reading several preceding passages is less useful for this selection process.
The article on readability for language models recommends one clear idea per paragraph, verifiable statements, and ideally fewer than 250 words. It advises putting the main point in the paragraph’s first sentence and the document’s most important information within its first 350 to 400 words. These are editorial recommendations drawn from GEO projects, not independently confirmed ranking rules from OpenAI.
Much of this advice is also sensible for human readers. If you publish a guide to accounting software, a paragraph should say which task a feature performs before moving to examples and limitations. In a company profile, its name, activity, and publicly verifiable core details should not appear only after a long corporate history.
Structure alone does not guarantee a mention. An AI search may choose different sources depending on the product version, the query, and rules that users cannot inspect. Promptwatch also reported during the same period that ChatGPT appeared less likely to use Reddit, although Willison could not confirm a corresponding change to its system instructions.
What you can improve in practice
Step 1: Review individual paragraphs
Read each important paragraph without its immediate context. It should still reveal the subject, the claim, and the basis for that claim. Replace ambiguous references such as “this solution” or “as a result” when their meaning depends entirely on the preceding paragraph.
Step 2: Put answers before explanations
Lead with the main point, then provide evidence, examples, and limitations. This helps both potential AI excerpts and people scanning a page. Long introductions that postpone the actual subject serve neither audience particularly well.
Step 3: Strengthen verifiable information
Keep company profiles, product descriptions, and editorial pages consistent. A Swiss company could review its publicly available directory data and complete a profile after the required verification. This does not prove that ChatGPT will mention the business more often, but it creates a clearer information base.
Step 4: Measure without expecting false precision
Test recurring, realistic questions in the services relevant to your audience and record which sources appear. One response is not a stable ranking. Even aggregated GEO tracking captures only a sample and may identify a change without reliably explaining its cause.
- Select five to ten important pages or company profiles.
- Write typical information-seeking questions that match those pages.
- Check whether every key paragraph presents understandable claims, evidence, and terminology.
- Record answers and cited sources at different points in time.
- Revise unclear content without adding unsupported claims or artificial repetition.
Pros and Cons of GEO
Pros:
- Clearer communication – Self-contained paragraphs and early main points generally help human readers as well.
- Additional discoverability – GEO reflects that people increasingly seek information in ChatGPT, Claude, Gemini, and Google’s AI products.
- Measurable presence – Recurring tests and maintained profiles can expose missing or contradictory information.
- Multiple routes to readers – Beyond AI citations, features such as Google’s preferred sources can reinforce existing audience relationships.
Cons:
- Opaque rules – Providers do not reveal every selection mechanism or their complete internal system instructions.
- Unstable results – Product updates can shift source preferences even when a website has not changed.
- Questionable promises – The GEO market can present observed correlations as guaranteed formulas for success.
- Conflicting access goals – Blocking AI bots from copying content may also restrict its availability to machine-driven services.
Shieldfont illustrates this conflict. Created by a designer and a copywriter, the font uses ligatures, typographic combinations of multiple characters, and is intended to stop AI tools from extracting usable training data from websites. It responds to scraping, the automated collection of publicly accessible web content by bots, which the report says sometimes happens despite an explicit objection from site operators.
The report on Shieldfont presents it as a protective layer against unwanted training-data collection. However, the supplied material does not independently test its effectiveness or its effects on discoverability, accessibility, and different AI search products. Visibility and protection from machine use are not automatically compatible goals.
What this means for your work
If you are just starting, you do not need to buy a GEO tool. Choose one important page, move its core statement into the opening sentences, and give each paragraph a clear subject. You can then use a small set of typical queries to observe whether and how the page appears.
If you already work with SEO, you can go further by documenting AI mentions separately from conventional rankings. Compare several services and multiple dates instead of treating one ChatGPT response as a general rule. For Swiss businesses, a verified directory profile can provide another channel, although the report does not say how well the measurements account for Swiss language variants or multilingual content.
Publishers and other sites that release content regularly also have a more direct option at Google. A “Preferred Sources” button lets readers mark a website as a favorite so that it can receive more prominent treatment across Google Search, Discover, and Google News. According to Google’s claims about the feature, users had selected more than 345,000 unique sources by May; earlier company studies found that people were about twice as likely to click a preferred source when one was available.
One practical example is a specialist publication that tells regular readers about the preferred-source option while continuing to publish clearly structured articles. Another is a Swiss small or medium-sized company that checks its public company data, completes its verified profile, and then tests common questions in several AI services. Both measures improve the available information but guarantee neither a mention nor additional traffic.
Quality matters even more as the volume of likely AI-written material grows. A Pew analysis of nearly half a million English-language pages found significant signs of AI authorship in about 10% of a July 2026 random sample. After older pages were excluded, 35% of pages published since ChatGPT’s release showed such signs, according to the account of the study.
Those figures do not reliably label every individual article. The detection tool can mistakenly classify human writing as AI-generated, so the study is better treated as a directional indication. Discoverability should therefore not come from interchangeable mass production, but from useful, clear, and verifiable information.
A sound GEO foundation looks less different from careful editorial work than some vendors suggest: clear statements, traceable evidence, and maintained information. The unresolved risk is durability, since AI search services can alter their source selection without full transparency. Sacrificing readability for people in pursuit of a supposedly reliable machine formula would be the least convincing trade.
Sources
- Warum ChatGPT deine Inhalte ignoriert – und wie du das änderst – t3n, 2026-08-22
- Shieldfont: Wie eine Schrift KI-Bots beim Scraping austrickst – t3n, 2026-08-22
- Cohaga lanciert KI-Firmenverzeichnis – Netzwoche, 2026-08-21
- ChatGPT search now uses the site:operator at scale – Simon Willison’s Weblog, 2026-08-20
- Google gives publishers a new way to fight AI-driven traffic losses – TechCrunch, 2026-08-20
- A third of web pages published since ChatGPT’s launch show signs of AI authorship, study finds – TechCrunch, 2026-08-20
