In the view of the Compagne, GEO optimisation is for the most part good search engine optimisation with clear, well-sourced content. Google itself classifies optimising for generative AI search as SEO¹.

Of 25 company websites in the Ingolstadt region, Germany, not a single one blocked ChatGPT’s search crawler in its robots.txt. Three had an llms.txt, ten had fewer than 300 words on the home page. The data comes from a small sample taken by the Compagne on 8 October 2026. From the Compagne’s perspective, it matches what Google and OpenAI themselves ask for: no additional AI files for Google Search², and no blocked search crawler for ChatGPT³.

This article checks common GEO promises against these sources and against the study behind the term.

What is GEO optimisation?

GEO stands for Generative Engine Optimization. It means preparing content so that AI systems such as ChatGPT search, Perplexity or Google’s AI Overviews find it, understand it and cite it as a source. The term has nothing to do with geo-targeting or local SEO, i.e. targeting a location.

It was coined by a research paper published in 2023 and presented at the KDD conference in 2024⁴. Terms such as LLMO (optimisation for large language models), AIO or AEO are used in similar ways in practice. In the Compagne’s observation, no fixed distinction between them has taken hold.

Google describes how its AI answers are produced: AI Overviews and AI Mode can run several related searches on subtopics for a single question and build an answer from the results². A page can therefore also serve as a source for a sub-question.

Is GEO something other than SEO?

For Google, no. Google’s guide to generative AI search says verbatim that optimising for it is optimising for search “and thus still SEO”¹. For ChatGPT there is an additional technical requirement: OpenAI’s search crawler must not be blocked³.

For AI Overviews and AI Mode, Google states that there are “no additional requirements”². A page has to be indexed and eligible to be shown with a snippet (the short text excerpt in the search result). Even then, appearing is not guaranteed².

Which AI crawlers should be allowed on the website?

For ChatGPT search, OAI-SearchBot is what matters most. OpenAI lists four crawlers³. A crawler is a program that fetches websites automatically. It is controlled via the robots.txt, a text file in the root directory of the website.

Crawler	Purpose according to OpenAI	Effect of blocking it in robots.txt
OAI-SearchBot	Showing websites in ChatGPT’s search features	Page does not appear in ChatGPT search answers, at most as a navigation link
GPTBot	Training the AI models	Content should not be used for training
ChatGPT-User	Individual actions that users trigger in ChatGPT; no automatic crawling	According to OpenAI, robots.txt may not apply, because the action is initiated by the user
OAI-AdsBot	Checking the safety of pages submitted as ads in ChatGPT; only visits those pages	Not stated; according to OpenAI, data is not used for training

According to OpenAI, search takes changes to the robots.txt into account after around 24 hours³.

Perplexity makes a similar distinction: PerplexityBot shows and links websites in Perplexity’s search results and, according to Perplexity, does not collect content for training AI models. Perplexity-User fetches pages when a user asks a question and generally ignores the robots.txt, because the request is initiated by the user. Changes to the robots.txt take effect after up to 24 hours⁷. A site that wants to appear as a source in Perplexity should therefore not block PerplexityBot.

In the Compagne’s sample, none of the 25 websites blocked OAI-SearchBot, and one blocked GPTBot. The robots.txt was checked for the home page; 24 of the 25 websites had one. Blocks via a firewall or a CDN (a network of servers that delivers websites) are not visible to this check. Such settings change easily and unnoticed during a relaunch; the website migration checklist therefore lists the point separately.

What does the GEO study really show?

In the GEO study by Aggarwal et al. (KDD 2024), three things helped most: citing sources, adding quotations and adding statistics⁴. Keyword stuffing, i.e. repeating search terms excessively, performed worse than the unchanged text on the source’s share of words in the answer in both tests. On subjective impression, it was slightly above the unchanged text in both tests⁴. The often-quoted “up to 40%” is a lab value from this experimental set-up. In the Compagne’s assessment, it cannot be transferred to an individual website, not least because the effect varied by topic⁴.

The set-up: a simulated answer engine had gpt-3.5-turbo write an answer from the top five Google results. The researchers modified individual source texts using nine methods and measured how much of them ended up in the answer. One partial test ran with 200 queries in Perplexity⁴.

Three points belong to an honest assessment:

• The effect depended on the topic. Different methods worked best depending on the subject area⁴.
• Lower-ranked sources benefited most. For the source in fifth place, visibility rose markedly with cited sources, while the source in first place lost on average⁴.
• The authors name limitations themselves. Answer engines change, the methods have to adapt, and the effect on search rankings was not examined⁴.

From the Compagne’s perspective, the study’s three best methods are close to what Google recommends: verifiable, original content¹.

Caution with summaries of the study: what was examined were nine methods of rewriting texts⁴. Structured data, text length and update dates were not among them. Claims that attribute effects of FAQ markup, minimum word counts or updates to the study cannot be found in it.

GEO promises fact-checked

Many of the extra measures sold as GEO are explicitly not required by Google. The table compares common promises, summarised in substance, with the primary sources. The assessment is the Compagne’s.

Promise (in substance)	What the primary source says	Assessment
No AI visibility without an llms.txt	Google: no new machine-readable files, AI text files or markup required2; Google Search ignores such files, they neither harm nor help there1. llmstxt.org calls the file a proposal5.	No effect for Google Search. Unproven for other systems.
AI needs special schema	Structured data is not required for generative AI search, and there is no special schema.org markup1.	Structured data remains a sound standard.
FAQ markup brings extra visibility	Google has not shown FAQ rich results since 7 May 2026 and removed the documentation in June 20266.	Questions and answers remain useful as content; the markup no longer brings special display on Google.
Texts must be rewritten into chunks and an AI style	No chunking needed, no separate writing style just for generative AI search1.	Clearly written texts are enough.
Build mentions on the web, no matter where	According to Google, seeking inauthentic “mentions” is less helpful than it seems1.	According to Google, inauthentic mentions achieve little.
Tool X shows the AI ranking	“No third-party tool has access to our internal ranking or AI systems.”1	Tools can show samples, not Google internals.
None of this applies to ChatGPT	OpenAI names not blocking OAI-SearchBot as a requirement3.	Check the technical side; after that, the same content questions apply.

Is an llms.txt worth having?

An llms.txt is a text file intended to give language models an overview of a website’s most important content. llmstxt.org describes it as a proposal from 2024⁵. According to Google, Google Search does not evaluate it, but it does no harm either¹.

Whether other AI systems use the file when answering has not been demonstrated. The effort is small, the effect uncertain. the Compagne’s website also provides an llms.txt, as a small experiment. In the Compagne’s assessment, an existing file does not need to be deleted, and a site without one misses little as things stand.

Where ChatGPT gets its reasons from

In a small test by the Compagne, ChatGPT justified its recommendation of IT service providers for two of the three companies it named with the service descriptions on their websites.

The test on 7 October 2026: three questions asking for recommendations in Ingolstadt, Germany (IT service provider, tax adviser, engineering firm), ChatGPT not logged in, a new chat for each question. For the IT question, ChatGPT named three companies. All three were also among the top ten results on Google Maps, but were not the ones with the best-maintained profiles there. For two of the three, the reasons came from the service descriptions on their websites; for the third, there was hardly any justification. For the other two questions, ChatGPT named industry directories and a review portal alongside company websites. In all three answers, ChatGPT asked follow-up questions, for example about headcount or budget.

The limitations: three questions, one run each, all with a local reference; answers can vary, and a causal relationship has not been demonstrated. the Compagne’s hypothesis: a website that describes concretely what a company does and for whom gives an AI material for a justification.

What websites actually lack

In the Compagne’s assessment, the websites in the sample lack content and SEO fundamentals rather than AI files. Which pages carry this work for B2B providers is described in the article B2B SEO. 32 websites from a company list on OpenStreetMap in the Ingolstadt region, Germany, were requested. 25 could be evaluated, 4 blocked the check via robots.txt, 3 were unreachable. The home page was checked automatically.

Characteristic	Share
XML sitemap present	20 of 25
Meta description present	15 of 25
Schema markup as an organisation	12 of 25
Home page with fewer than 300 words (incl. menu and footer)	10 of 25
Exactly one main heading (H1)	10 of 25
Knowledge section (blog, guides, news) linked from the home page	8 of 25
llms.txt	3 of 25
OAI-SearchBot blocked in robots.txt	0 of 25

The sample is small and not representative. Only the home page was checked; a knowledge section can exist without being linked there.

A term from Google’s guide helps with the assessment: “non-commodity content”. It refers to content with original expertise or experience that goes beyond general knowledge¹. In the Compagne’s assessment, a home page with little text and no linked specialist articles offers little that can be cited.

Which measures are demonstrably worthwhile

What the primary sources name is worthwhile: being accessible, meeting SEO fundamentals, writing original content with cited sources, keeping business information up to date^{2, 1, 3, 4}.

Technical side: do not lock crawlers out

Googlebot must be able to index the pages², and OAI-SearchBot must not be blocked³. At OpenAI, training and search are handled separately: according to OpenAI, blocking GPTBot does not affect ChatGPT search³. Besides the robots.txt, firewall and CDN settings are worth checking, because crawlers can be blocked there too.

the Compagne’s working rules for citable texts

Points 3, 5 and 7 are supported by sources. Points 1, 2, 4 and 6 are the Compagne’s working rules; none of the sources evaluated measures their effect.

1. One question per section. The subheading asks it, the first sentence answers it.
2. Answer first, then the reasoning. The first sentence must be able to stand on its own.
3. Figures with source and date. In the study, citing sources and adding statistics were among the most effective methods⁴.
4. Concrete statements. “Set-up usually takes a few days” is usable. “Fast time-to-value” says nothing verifiable.
5. Make own experience recognisable. Own data and observations are content that goes beyond general knowledge¹.
6. Answer real questions. The questions prospective clients actually ask, in their words, with short answers.
7. Write for people. According to Google, a separate AI style is not necessary¹.

Keep business information up to date

Google recommends keeping the information in the Business Profile up to date². According to the guide, the profile can help with visibility in AI answers and other search results¹.

Structured data: a sensible baseline

Schema markup such as Organization or Article describes a page in machine-readable form. According to Google, it is not required for AI features¹. the Compagne still considers it sensible, as an assessment.

What does GEO optimisation cost?

In the Compagne’s assessment, Google’s AI features do not require a separate GEO budget, because Google sets no additional requirements^2, 1. Costs arise from clean technical foundations and well-sourced content. Four questions help with any proposal:

• Which measure exactly will be implemented?
• Which primary source supports it?
• How is it measured, and what does the measurement not show?
• Is a mention in AI answers being promised? Then caution is advised, because Google does not guarantee appearance either².

How can GEO visibility be measured?

In the Compagne’s assessment, only indirectly: via Search Console, server logs, spot checks in AI assistants and asking new enquirers. According to Google, no external tool has access to its internal systems¹. The sources support only the technical basis of these approaches.

• Search Console: Traffic from AI Overviews and AI Mode is included in the Performance report under Search type: Web². The source does not describe a separate breakdown.
• Server logs: Visits from OAI-SearchBot or ChatGPT-User can be recognised by their name³. They show which pages are fetched.
• Spot checks: The most important questions are asked regularly in the common assistants and the sources named are noted. Because answers can vary, the trend across several tests is what counts.
• Asking: New enquirers are asked where the name first came up.

In what order does GEO pay off?

In the Compagne’s assessment, the sources evaluated suggest this order: first the technical side and SEO fundamentals, then content with original expertise, and finally monitoring AI visibility through spot checks. the Compagne considers packages that promise mentions, and extra files as the main measure, to be dispensable.

the Compagne promises no rankings and no mentions in AI answers, and reports openly on what has been measured and what has not. The visibility audit checks the technical side, content, the Google profile and AI visibility, and ends with a prioritised list of measures. How Design and Visibility work together is described in the practice; the starting point is an initial conversation.

Updated on 21 October 2026: Perplexity crawlers and the end of FAQ rich results on Google added, along with an assessment of widely circulated summaries of the GEO study.

Sources

1. Google Search Central: Optimizing for generative AI search (last updated 2026-07-10): https://developers.google.com/search/docs/fundamentals/ai-optimization-guide (accessed 2026-10-08)
2. Google Search Central: AI features and your website (last updated 2025-12-31): https://developers.google.com/search/docs/appearance/ai-features (accessed 2026-10-08)
3. OpenAI: Overview of OpenAI Crawlers: https://developers.openai.com/api/docs/bots (accessed 2026-10-08)
4. Aggarwal et al.: GEO: Generative Engine Optimization (KDD 2024): https://arxiv.org/abs/2311.09735, https://arxiv.org/html/2311.09735v3 (accessed 2026-10-08)
5. llmstxt.org: The /llms.txt file (Jeremy Howard, 2024-09-03, modified 2026-08-10): https://llmstxt.org/ (accessed 2026-10-08)
6. Google Search Central: structured data changelog, entries of 2026-05-08 and 2026-06-15 on FAQ rich results: https://developers.google.com/search/docs/appearance/structured-data/faqpage (accessed 2026-10-21)
7. Perplexity: Perplexity Crawlers: https://docs.perplexity.ai/docs/resources/perplexity-crawlers.md (accessed 2026-10-21)