GEO Optimization Fact Check (Autumn 2026): What the Research Really Shows

In short: There are many promises about GEO optimization, that is, optimising for AI search tools such as ChatGPT. We checked eight common claims against the research: the basis is more than 90 scientific papers and over 30 trade articles, and we cite around 50 of them directly. The result: AI adopts what is on your website too willingly rather than too critically. Quick tricks hardly work. And many of the figures used in advertising have no evidence behind them.

Who this article is not for: If you are looking for step-by-step instructions on GEO optimization, you will not find a checklist here, but rather what is actually proven. If you want to check whether ChatGPT already mentions your practice today, our article Measuring AI visibility will help you more. If you are looking for the technical settings for AI crawlers, you will find them under AI crawlers and llms.txt.

Why we checked

Since this year, many providers have been promising places in ChatGPT, their own metrics for AI search and new files that are supposed to make a website “AI ready”. Practices and companies in medicine and life science ask us what is true. So instead of reading the advertising, we read the research.

We analysed scientific papers from universities in the USA, Europe and Asia, studies on when an AI trusts a source, and trade articles from the search engine industry over the last six months. As in medicine, not every source carries the same weight:

For each claim, we state the level of the evidence.

“Doctors ask about the evidence for every new therapy. Hardly anyone asks that about marketing promises. That is exactly what we have now done.”

The 8 claims at a glance

None of the eight claims holds up in its common form. The details and all evidence follow below.

What AI does with your website

Claim 1: Does AI know better than your website?

Not when it comes to information about your practice. Language models adopt a well-written text even when they know the correct answer themselves. In a Stanford University study, this happened in over 60 per cent of cases, and the questions examined included medication dosages (ClashEval, Stanford).

What matters is how the text is built. A coherent, well-reasoned paragraph prevails, a bare assertion much less often (Xie et al., ICLR 2024). And the less known a topic is, the less knowledge of its own the model has (Mallen et al., ACL 2023). About your opening hours, your team or how an initial consultation works, it knows almost nothing by itself.

What this means for you: What your website says about your practice determines the AI’s answer, for better or worse. A practice page is overruled almost only on general health questions: when it deviates from professional consensus and a large portal confirms the consensus.

For medtech and pharma: The same applies to old product pages and outdated data sheets. Whatever is still online, AI may present as current.

Claim 2: Is the answer correct if the AI cites a source?

No. A Stanford study in Nature Communications examined medical answers from seven AI models. Depending on the model, 50 to 90 per cent of answers were not fully supported by the sources cited. Even with web search, around 30 per cent of individual statements remained unsupported (SourceCheckup, Nature Communications 2025).

In addition, ChatGPT retrieves many pages during a search but cites only about one in eight of them (MPI-SWS, preprint).

And even a correct AI answer does not automatically help. In a British study, 1,298 people were given ten case descriptions with symptoms. On its own, the AI named the correct condition in almost 95 per cent of cases. People who chatted with the AI themselves reached less than 35 per cent. That was no better than the group that looked for an answer without AI (Bean et al., preprint).

What this means for you: Your patients arrive informed in advance, but not always correctly informed. You can check whether AI says the right things about your practice with our guide What ChatGPT says about your practice.

Claim 3: Does AI filter out exaggerated advertising language?

There is no evidence for this, and the research tends to point the other way. No study shows that language models downgrade advertising language. What is proven, however, is that AI adopts slightly exaggerated, plausible statements more readily than obviously false ones (ClashEval, Stanford).

A study by the University of Waterloo shows how strong the effect of a single document is. One misleading health document in the retrieval reduced GPT-5’s agreement with medical consensus from 81.6 to 43.4 per cent. A helpful document alongside it weakened the effect (Amirshahi et al., preprint).

What this means for you: A phrase like “guaranteed painless” can end up as a fact in someone else’s AI answer, with no sender and no context. Factual language in line with the Heilmittelwerbegesetz (German law on advertising for medicines and treatments) is therefore not just a legal matter. It also protects the quality of what AI says about your specialty.

That is why we do not see the Heilmittelwerbegesetz as a brake, but as a mark of quality for AI search.

What Google and AI search have in common and what they do not

Claim 4: Does whoever ranks high on Google appear in ChatGPT?

Only partly. AI search tools prefer to cite third parties such as portals, press and videos, and less often a provider’s own website. For dentists in the USA and Canada, the sources of AI search matched the Google results in only 11.9 per cent of cases (Chen et al., University of Toronto).

For Germany, there is an analysis by the software provider SE Ranking covering 50,807 health questions. According to it, only 36 per cent of the pages cited by Google’s AI Overview were also among the top ten Google results. The most cited source was YouTube, while specialist journals accounted for 0.48 per cent of references (Deutsches Gesundheitsportal via SE Ranking).

Even so, a good position on Google is still worth having. AI search tools usually pick their sources from the results of a search engine. In a comparative study, it helped a page considerably more to rank higher in those results than to rewrite its text for AI (C-SEO Bench, NeurIPS 2025).

What this means for you: Classic search engine optimisation remains the foundation. In addition, it matters where your practice appears outside its own website: in portals, directories, the regional press or in videos. We describe how search is changing overall in the SEO and GEO trends 2026.

Claim 5: How many search queries does AI break a question down into?

The industry often talks of 8 to 12 sub-queries per question. The research finds considerably fewer. An analysis of over 170,000 real conversations shows: when ChatGPT searches the web, it sends 3 search queries on average, Claude just under 2 (MPI-SWS, preprint).

These search queries are 8 to 13 words long on average, shorter for factual questions and longer for analyses. More than half of the new search terms are taken by the AI from texts it has already found (Agentic Search in the Wild, SIGIR 2026). However, measurements were only made in English.

What this means for you: AI asks few but precise questions. A page that answers the patient’s whole question, with procedure, duration, starting prices and location, fits this better than ten similar pages.

Tools, files and tricks

Claim 6: Do GEO tricks get your practice into every AI answer?

The first study on GEO did measure gains from quotations and statistics in the text, but in a simulated search engine (Aggarwal et al., KDD 2024). A later review found most such methods largely ineffective, and the gain shrank as soon as several providers optimised at the same time (C-SEO Bench, NeurIPS 2025). A review of 45 studies finds no evidence of a lasting effect across different AI search tools (Martinez 2026, preprint).

What this means for you: Anyone who promises you a fixed place in ChatGPT is promising something nobody can prove. Hidden texts for machines work in the lab. For practices they are legally risky, because they can be considered misleading.

Claim 7: Does an llms.txt file or schema markup get you into AI?

In this year’s data analyses, llms.txt showed no effect on visibility. In a study of over 137,000 domains, 97 per cent of the files were not retrieved a single time in May 2026 (Ahrefs). An analysis of around 300,000 domains found no link with citation frequency (SE Ranking).

For structured data according to schema.org, there are only correlations, no proof of a cause (GEO-16, preprint).

What this means for you: Structured data remains useful because it helps Google identify your practice clearly. According to current knowledge, however, it is not a proven lever for AI answers. Which settings really matter is described in AI crawlers and llms.txt.

Claim 8: Does ChatGPT have a search volume like Google?

Real user questions to ChatGPT are not public. Every figure on “prompt volume” is therefore an extrapolation. Even one provider of such tools writes about the claim to know real user questions: “That’s just not possible.” (Conductor) Another itself labels its AI search volume as “an approximation” (SISTRIX).

One scientific paper describes the problem like this: a list of test questions reflects its own market, not necessarily real demand (Measuring GEO Visibility, preprint).

What this means for you: For every figure, ask where it comes from. The most reliable source for demand for your practice remains Google Search Console. How to monitor your visibility in AI answers sensibly is described in Measuring AI visibility.

What nobody knows

Research has gaps too. We name them because they are part of the picture:

• No study has examined German practice websites. Almost all measurements are in English.
• Many figures come from lab experiments with models from 2023 to 2025. They show directions, not exact values for today’s AI.
• How Perplexity and Google’s AI Mode process questions has hardly been studied scientifically.
• Many papers are preprints. We have marked this for every source.

Anyone who claims today to know exactly how ChatGPT selects practices knows more than the research does.

Our position

From this fact check, we draw three principles:

“AI believes almost everything on your website. That is an opportunity and a responsibility. Only write what you would also say in the treatment room.”

We review this fact check every quarter. If the evidence changes, we change the traffic light.

Two misconceptions when you use AI yourself

Many practices and companies now use AI themselves, for texts, analyses or internal processes. Here too, there are two common assumptions:

On the first point: with 500 simultaneous instructions, the best model tested fulfilled only 68 per cent of them (IFScale). In a study by ETH Zürich, long rule files for AI assistants brought no measurable benefit but cost around 20 per cent more (ETH Zürich). A clear goal and a few fixed rules work better.

On the second point: different AI models often make the same mistakes. In one study, nine AI judges from seven model families carried only as much weight as just over two independent votes, and unanimous verdicts were wrong in 9 per cent of cases (Nine Judges, Two Effective Votes). As with a second opinion, independence counts, not the number. The professional review remains with people.

Conclusion

GEO optimization is not a magic trick, but it is not an empty promise either. One thing above all is proven: AI adopts what it finds about your practice. That is why complete, honest and consistent information counts for more than any trick. The foundation remains a website that is easy to find on Google, and a practice that others write about too.

Added value · In briefMeasuring AI visibility: how practices can see whether ChatGPT and Google mention themWhether ChatGPT, Gemini and Google’s AI already mention your practice today, and how you can check this yourself:

Want to build your practice's visibility in search and AI answers for the long term? That is what our SEO and AI visibility service is for.

Studies for further reading

All sources for this fact check, sorted by claim, with origin, level of evidence and key finding. One click opens the list.

Claim 1: AI knows better than your website anyway7 sources

1. Adaptive Chameleon or Stubborn SlothFudan University, Ohio State, Penn State; ICLR 2024 · Peer-reviewed studyA coherent, well-reasoned text is adopted, a crude contradiction is not.
2. ClashEvalStanford University · PreprintOver 60 per cent adoption of false sources, although the model knew the correct answer.
3. Studying LLM Behaviors Under Context-Memory Conflicts With Real DocumentsTübingen AI Center · PreprintWith real documents, the model sticks to its prior knowledge in only 0.1 to 6.2 per cent of cases.
4. When Not to Trust Language Models (PopQA)University of Washington; ACL 2023 · Peer-reviewed studyFor little-known facts, the model itself knows hardly anything, and the source it finds decides.
5. Context versus Prior Knowledge in Language ModelsETH Zürich; ACL 2024 · Peer-reviewed studyThe better known a topic, the more stable the model’s prior knowledge remains.
6. DeepSeek-R1 ThoughtologyMila, McGill University · PreprintEven “reasoning” models follow false context in 78 per cent of cases.
7. Memory vs. Context? Influential Factors of Factual Recall2026, 31 models · PreprintThe wording of the question alone strongly shifts whether the model follows prior knowledge or the source.

Claim 2: If the AI cites a source, the answer is correct6 sources

1. SourceCheckupStanford University; Nature Communications 2025 · Peer-reviewed study50 to 90 per cent of medical AI answers not fully supported by the sources cited.
2. Evaluating Verifiability in Generative Search EnginesStanford University; Findings EMNLP 2023 · Peer-reviewed studyOnly just over half of the sentences were fully supported by citations.
3. Clinical knowledge in LLMs does not translate to human interactionsUnited Kingdom, 1,298 participants · PreprintWith AI help, laypeople did not identify conditions any better than without.
4. Characterizing Web Search by Conversational LLM AgentsMPI-SWS, Saarland University · PreprintChatGPT cites only 13.3 per cent of the pages it retrieves during a search.
5. From Citation Selection to Citation Absorption2026 · PreprintBeing cited as a source does not mean the content was actually used.
6. We Compared Eight AI Search EnginesTow Center, Columbia University · JournalismOver 60 per cent incorrect source attributions across eight AI search engines.

Claim 3: AI filters out exaggerated advertising language5 sources

1. Evaluating the Robustness of RAG to Adversarial Evidence in the Health DomainUniversity of Waterloo, Toronto Metropolitan University · PreprintOne misleading health document halves agreement with consensus, a helpful one alongside it protects.
2. Faithfulness vs. Safety (MedCounterFact)Findings ACL 2026 · Peer-reviewed studyModels mostly adopt invented study evidence without checking it.
3. Dr ChatGPT tell me what I want to hearCSIRO, University of Queensland; EMNLP 2023 · Peer-reviewed studyAdditional evidence in the prompt can make health answers worse.
4. Evaluating search engines and LLMs for answering health questionsUniversity of Santiago de Compostela; npj Digital Medicine 2025 · Peer-reviewed studyThe accuracy of the inserted source determines the quality of the health answer.
5. ClashEvalStanford University · PreprintSlightly falsified, plausible values are the most likely to slip through.

Claim 4: If you rank high on Google, you also appear in ChatGPT5 sources

1. Generative Engine Optimization: How to Dominate AI SearchUniversity of Toronto · PreprintAI search tools prefer third-party sources; for dentists, 11.9 per cent overlap with Google.
2. C-SEO Bench: Does Conversational SEO Work?NeurIPS 2025 · Peer-reviewed studyA better ranking has a stronger effect than rewriting texts.
3. Google users are less likely to click on links when an AI summary appearsPew Research Center, 68,879 Google searches · Independent surveyOnly around 1 per cent click on a source in the AI summary.
4. Which sources does AI use?Deutsches Gesundheitsportal via SE Ranking, 50,807 German health questions · Vendor studyYouTube most cited source, specialist journals 0.48 per cent of references.
5. AI Search OverlapAhrefs · Vendor studyOnly about 12 per cent of links from ChatGPT, Gemini and Copilot are in Google’s top 10.

Claim 5: AI breaks every question down into a dozen search queries6 sources

1. Characterizing Web Search by Conversational LLM AgentsMPI-SWS, Saarland University · Preprint171,264 real conversations: ChatGPT sends 3.07 search queries on average, Claude 1.80.
2. Agentic Search in the WildCarnegie Mellon University; SIGIR 2026 · Peer-reviewed study14.4 million queries, 7.6 to 12.7 words on average; 54 per cent of new search terms come from texts already found.
3. A Picture of Agentic SearchUniversità di Pisa et al. · PreprintAI search queries have 6 to 14 words and use question words less often than human ones.
4. Question Decomposition for Retrieval-Augmented GenerationHumboldt-Universität zu Berlin; ACL 2025 SRW · Peer-reviewed studyBreaking questions down only helps together with sorting against the overall question.
5. When Should Queries Be Decomposed?Eastern Institute of Technology Ningbo, Shanghai Jiao Tong University · PreprintSimple questions are better not broken down at all.
6. LLM-based Query Expansion Fails for Unfamiliar and Ambiguous QueriesUniversity of Tsukuba; SIGIR 2025 · Peer-reviewed studyWith unfamiliar technical terms, AI drifts towards the best-known meaning.

Claim 6: With the right GEO tricks, you will appear in every AI answer7 sources

1. GEO: Generative Engine OptimizationIIT Delhi, Princeton University; KDD 2024 · Peer-reviewed studyThe original study measured gains, but only in a simulated search engine.
2. C-SEO Bench: Does Conversational SEO Work?NeurIPS 2025 · Peer-reviewed studyMost methods largely ineffective, and the gain shrinks when everyone optimises.
3. Optimizing Visibility in Generative Engines: A Critical Survey2026, 45 studies · PreprintNo lasting effect proven across different AI search tools.
4. What Evidence Do Language Models Find Convincing?UC Berkeley; ACL 2024 · Peer-reviewed studyWhat matters is relevance to the question, not tone or professional references.
5. What Generative Search Engines Like (AutoGEO)Carnegie Mellon University · PreprintModels prefer comprehensive, verifiable and neutrally written content.
6. Ranking Manipulation for Conversational Search EnginesUC Berkeley; EMNLP 2024 · Peer-reviewed studyManipulation works in the lab, including with Perplexity.
7. Adversarial Search Engine Optimization for Large Language ModelsETH Zürich · PreprintWhen everyone manipulates, all answers get worse.

Claim 7: An llms.txt file or schema markup gets you into AI6 sources

1. llms.txt studyAhrefs, 137,210 domains · Vendor study97 per cent of the files were not retrieved a single time in May 2026.
2. llms.txtSE Ranking, around 300,000 domains · Vendor studyNo link with citation frequency.
3. GEO myths put to the data testSISTRIX, 22.10.2026 · Vendor studyllms.txt without measurable effect, schema only as a correlation.
4. AI features and your websiteGoogle Search Central · Manufacturer’s claimGoogle Search does not use llms.txt.
5. How to Optimize for AI SearchAhrefs · Vendor studyIn the test, AI systems used the visible text, not the structured data.
6. AI Answer Engine Citation Behavior (GEO-16)2025 · PreprintStructured data only as a correlation, no proof of a cause.

Claim 8: ChatGPT has a search volume just like Google5 sources

1. Measuring GEO Visibility: Prompt Corpora Define the Answer Market2026 · PreprintA list of test questions reflects its own market, not real demand.
2. The Dice Roll Method2026 · PreprintAI answers vary greatly, and only become reliable across many different questions.
3. Measuring AI Search like SEOMoz · Vendor studyThe most frequent “prompts” are filler phrases; only the topic level is useful.
4. AI Prompt Generation for AEO and GEOConductor · Vendor studyKnowing real user questions is said to be impossible.
5. Prompt Research: keyword research for AI searchSISTRIX · Vendor studyThe AI search volume is explicitly an approximation based on Google data.

Two misconceptions when you use AI yourself7 sources

1. How Many Instructions Can LLMs Follow at Once? (IFScale)USA · PreprintWith 500 instructions, the best model fulfils 68 per cent.
2. AgentIFChina · PreprintLong instructions to AI agents are fully followed in only 27 per cent of cases.
3. Evaluating AGENTS.mdETH Zürich · PreprintRule files without measurable benefit, 19 to 23 per cent higher costs.
4. Probe-and-Refine Tuning of Repository Guidance2026 · PreprintCounterpoint: tested and refined rules help considerably.
5. Nine Judges, Two Effective VotesUSA · PreprintUnanimous AI verdicts were wrong in 9.1 per cent of cases.
6. Great Models Think AlikeTübingen · PreprintThe stronger the models, the more similar their mistakes.
7. Debate or VoteUniversity of Wisconsin · PreprintSimple voting beats discussion between AI models.

As of 26 October 2026. Next review: January 2027.