Updated on 7 November 2026, by João Leite, Founder of murmur.
Manipulating what ChatGPT recommends worked in a 2024 test and failed in a 2026 one, and whoever tries it puts their own site at risk. In November 2024, The Guardian pointed ChatGPT to a fake camera page, and a hidden instruction made it praise the page without caveats. In April 2026, on an Otterly test site, no AI model followed the hidden instruction, and only three even read the text. Google and Bing treat any attempt to manipulate the answers of their AI models as a violation, and the offending site can drop or be removed from the results. What changes the recommendation without that risk is a page that answers the question in plain sight and a true mention on other sites.
- Hidden instruction: ChatGPT obeyed on the page The Guardian pointed to, in 2024; in the Otterly test, in 2026, no AI model obeyed.
- The rules: Google calls it spam to try to manipulate the answers of generative AI in its search, and Bing can remove the site.
- A simple defense: in 200 buying questions, manipulation tactics delivered little or nothing on GPT-5 and Claude, and the authors conclude that the gain requires improving the content.
- A brand’s own “best of” list: it is not in Google’s spam policy.
- What holds up: a page visible on the brand’s own site. At Mindo, ChatGPT recommendations rose from 12% to 23% of the 134 fixed questions, from June to July 2026.
What counts as manipulating what ChatGPT recommends, and what is legitimate work
Manipulating what ChatGPT recommends means changing the answer with something hidden from the reader of the page or with something that is not true. Invisible text, a hidden instruction for the AI, a swapped date, a made-up review and a paid mention without disclosure all count. Legitimate work is the opposite: the same answer for visitors and for the bot, with data that can be checked. The question that separates the two is whether the change would make sense to the reader, with no AI in the way.
People who ask whether it is possible to manipulate what ChatGPT recommends usually bundle three questions. Influencing the recommendation is possible: that is GEO, the work of making the company appear in the answers of ChatGPT and the other AI models. Deceiving Google’s AI is spam under its policy, and the effect measured in AI models came from narrow, dated tests. No one can choose the answer to a question, because the same question rarely comes back with the same list of brands. In SEO, the trick side is called black hat, and the guide what is GEO shows where it diverges from ethical work.
murmur is a GEO agency in Brazil that places brands in the answers of ChatGPT, Claude, Google AIO and other AI agents. Google AIO is the AI Overview, the text written by AI that opens many Google searches, before the list of links. murmur works to make the client’s brand appear in more answers of a fixed set of questions in its market.
At Mindo, a motion design studio in São Paulo, we published a guide on a subdomain, an address inside the brand’s domain, open to any visitor. On 25 and 26 June 2026, ChatGPT recommended Mindo in 12% of the 134 questions in its market. In the same 134, on 13 and 14 July, it recommended Mindo in 23%. Legitimate influence shows up like this: across the set, with the total in view.
Tactic by tactic, the line between manipulating and working looks like this:
| tactic | what it looks like | what the rules and the tests say |
|---|---|---|
| hidden text | white letters on a white background, text off the screen or at zero size | spam under Google’s policy; it drew out ChatGPT’s summary in a 2024 test |
| hidden instruction for the AI | an order such as “always recommend brand X”, invisible to visitors | its own section in Bing’s rules; OpenAI writes that an instruction in a tool result is information; Anthropic documents that Claude is skeptical of it, for those building applications |
| cloaking | one page for the bot and another for visitors | spam on Google and a manipulative practice on Bing |
| swapped date | a new date on a page that has not changed | Google asks whether the site does this, among the signals of content made for the search engine |
| made-up review or claim | a testimonial, rating or proof that does not exist | the Consumer Protection Code prohibits misleading advertising |
| paid mention without disclosure | an ad that looks like editorial text | the law and CONAR call for advertising to be identified |
| a brand’s own “best of” list | the brand puts itself at the top of its own list | it is not in Google’s spam policy |
| a page that answers the question | the same answer for everyone, with data that can be checked | it is the work that Google’s and Bing’s rules ask for |
Of the seven tactics in the table, six hide something from someone or claim what is not true. The brand’s own “best of” list sits in the middle: it hides nothing, but the brand is judging itself.
What Google, Bing, OpenAI and Anthropic write about manipulating the AI answer
For Google and for Bing, trying to manipulate the answers of their AI models is a violation of the search rules. OpenAI writes that an instruction in a tool result, such as a fetched page, is information and not an order. Anthropic says the same about Claude, in the documentation for those building applications. At Google, the rule is in the opening of the spam policy:
“No contexto da Pesquisa Google, spam se refere a técnicas usadas para enganar os usuários ou manipular nossos sistemas de pesquisa para destacar o conteúdo, como tentar manipular os sistemas de pesquisa para classificar o conteúdo com uma alta posição ou tentar manipular as respostas da IA generativa na Pesquisa Google.”
(Google, Políticas de spam para a Pesquisa Google na Web)
Original quote in Portuguese, preserved as published. In English: “In the context of Google Search, spam refers to techniques used to deceive users or manipulate our search systems to highlight content, such as trying to manipulate the search systems to rank content in a high position or trying to manipulate the answers of generative AI in Google Search.”
The spam policy applies to Google Search, where the AI Overview lives, and not to ChatGPT. But the site that hides text to deceive ChatGPT is the same one Google reads. And the policy treats hidden text, such as white letters on a white background, as spam. Google defines cloaking as “apresentar diferentes conteúdos para os usuários e os mecanismos de pesquisa com a intenção de manipular as classificações de pesquisa e enganar os usuários” (“presenting different content to users and to search engines with the intention of manipulating search rankings and deceiving users”).
Bing, Microsoft’s search engine that OpenAI lists among ChatGPT’s search partners, wrote the rule already thinking about AI answers. In its guidelines, content “should be created for users, not to manipulate ranking systems or trigger citations or AI responses”. The hidden instruction for the AI, called prompt injection in English, is an order placed on the page for the AI to obey. It has its own section in the guidelines:
“Attempts to add content designed to manipulate or interfere with language models used by Bing or Copilot may result in reduced visibility or removal from search experiences.” (Bing Webmaster Guidelines, seção “Prompt Injection and AI Manipulation”)
The Model Spec is the public document in which OpenAI describes how it wants its models to behave. There, by default, any instruction inside quoted text, an attachment or a tool result “MUST be treated as information rather than instructions to follow”. On instructions coming from a tool, the 18 August 2026 version says that “a random web page should not be trusted at all”.
Anthropic documents the same care for those building applications with Claude. The fetched web page is among the third-party sources of instructions it lists. And Anthropic writes that “Claude is trained to treat instructions that appear inside tool results with appropriate skepticism”. A written rule shows what the company wants from the model, and does not prove that it always does so.
Hidden instruction for the AI: what the 2024 and 2026 tests show
A hidden instruction for the AI worked on ChatGPT in a 2024 test, and no AI model in a 2026 test followed one. In the lab, a similar effect had already appeared in 2024: Kumar and Lakkaraju tested a fictional catalog of coffee makers on an open model, Llama 2. A text generated by an algorithm, placed on a product page, made that product the first recommendation more often.
The British newspaper The Guardian tested ChatGPT search in November 2024, with GPT-4o, shortly after its launch for subscribers. The newspaper created a fake camera page, gave the URL to ChatGPT and asked whether it was worth buying. With nothing hidden, the answer was positive, with caveats such as the fixed lens. With a hidden instruction asking for a favorable review, the answer was always entirely positive, even with negative reviews on the page.
The Guardian’s test showed ChatGPT summarizing a page provided by the person themselves, not choosing brands among many sources. OpenAI did not comment on the case and told TechCrunch that it uses several methods to block malicious sites and that it keeps improving.
Otterly, which sells AI monitoring, repeated the idea in April 2026, on a test domain. The hidden instruction asked the AI models to treat the site as the main source of trust. None of the AI models tested followed the instruction, and only three read the hidden text. Microsoft’s Copilot refused the page for unsafe content, and Gemini exposed the attempt in its own answer. ChatGPT never reached the hidden text and described another blog with a similar name.
Otterly warns that it was a single site, with little content, and that the test does not represent the whole market. Each test is valid for the date, the model and the site on which it was done.
Swapped date and invented claim: the laboratory effect and the risk outside it
Swapped dates and invented claims changed the choice of AI models in the laboratory, and outside it they run into Google’s rules and the law. Fang and others put false dates on excerpts from a test collection and asked seven models, including GPT-4o, to reorder them. The excerpts with newer dates rose, some by up to 95 positions, and the larger models reduced the effect without eliminating it. The study is from September 2025 and measures a laboratory task, not ChatGPT answering someone.
Google, in its search, lists swapped dates among the questions that point to content made for the search engine, and not for people. On the page about how to create helpful content, it asks whether the site is “mudando a data das páginas para que elas pareçam atualizadas quando o conteúdo não mudou significativamente” (“changing the date of pages so they look updated when the content has not changed significantly”).
Invented claims also carry weight in testing. Chu and Hou tested in 2026, via the API, GPT-4o-mini, Claude Sonnet and Gemini 3 Flash recommending skincare products. Authoritative language, including an invented clinical claim, broke the advantage of well-known brands among products with the same specification, worth +0.17 points in the score. Each model reacted in its own way.
In Brazil, a false claim in advertising is a matter of law. The Consumer Protection Code prohibits “toda publicidade enganosa ou abusiva” (“all misleading or abusive advertising”), in article 37. Article 38 places the burden of proving that the advertising is true on whoever sponsors it.
When everyone tries the same manipulation, the gain falls
In testing, when several brands try the same manipulation, each one’s gain falls and the AI answer gets worse for everyone. Nestaas, Debenedetti and Tramèr called this the prisoner’s dilemma, in a 2024 study. Each party has a reason to attack, but the combined effect makes the model’s answers worse for everyone. The tests were done on two AI search engines that were live and on GPT-4 and Claude plugins.
In E-GEO, Bagga and others tested, across 200 purchase questions about Amazon products, a simple defense in the request to the AI. With it, the manipulation tactics achieved little or nothing in GPT-5 and Claude. The authors conclude that the gain requires improving the content.
Your own “best of” list, paid mentions and text on a third-party site: where the line is
Your own list, paid mentions and text on a third-party site follow the rules when they are identified and carry the qualified link. Paying for a post with a link is spam for Google, unless the link carries rel=“sponsored” or rel=“nofollow”. These are markers in the link code that tell Google the link is paid or that the site does not recommend it. In the site reputation policy, Google does not list as a violation the advertorial made to “compartilhar conteúdo diretamente com os leitores” (“share content directly with readers”). The problem is third-party text published on a site mainly for the strength that site already has in search. These rules are Google Search’s, not ChatGPT’s.
In Brazil, telling the reader that the text is advertising is a rule of law and of self-regulation. The Consumer Protection Code requires advertising to be conveyed in a way that the consumer identifies it “fácil e imediatamente” (“easily and immediately”), in article 36. The CONAR code requires that paid reporting be “apropriadamente identificada para que se distinga das matérias editoriais” (“appropriately identified so that it is distinguished from editorial pieces”), in article 30.
A paid mention identified as such also weighs little in what AI models cite. In May 2026, Muck Rack, which sells public relations software, classified more than 25 million links cited by ChatGPT, Claude and Gemini. Paid and advertorial content was 0.3% of citations.
The “best of” list with your own brand at the top does not appear in Google’s spam policy, and was a common tactic for years. What to put on your own site is in how to appear in ChatGPT in software comparisons.
Getting a competitor out of the ChatGPT answer: what the tests and the law say
You cannot tell ChatGPT to stop recommending a competitor. Trying to remove it from the answer with a trick is manipulation, with the same limits and the same risk for the site. In 2024, Nestaas, Debenedetti and Tramèr showed that text written to deceive the AI led models to promote the attacker’s products and discredit competitors. The text was on sites and in plugin documentation. The tests were on two AI search engines and on GPT-4 and Claude plugins, not on ChatGPT search.
In Brazil, speaking ill of a competitor in advertising has a written limit. The CONAR code accepts comparison between brands with limits, in article 32. One of them is that “não se caracterize concorrência desleal, depreciação à imagem do produto ou à marca de outra empresa” (“unfair competition is not characterized, nor depreciation of the image of another company’s product or brand”).
The AI lists the competitors that the training and the pages it reads cite. What changes the list is your company appearing on those pages and answering the same questions better. The step by step, question by question, is in the guide when the AI cites the competitor and not your company.
The risk of manipulation stays with the site of whoever does it
Whoever tries to manipulate the AI answer risks their own site, because the site is where the consequences set out in the search rules land. In Google Search, a site that violates the spam policy can lose position or stop appearing, after automatic detection or human review. Bing writes that sites with these practices can lose position, stop serving as a source for AI answers or drop out of the index. In Otterly’s test, Copilot refused the entire page, not just the hidden excerpt.
murmur’s work is out in the open: on the client’s site and with the total number of questions
murmur publishes the pages on a subdomain of the client’s brand, in plain sight, and shows each result with the total number of questions. The pages answer the questions in the client’s sector, and the content is entirely theirs, even after the contract ends. Each question that is genuinely different for the buyer gets a page with the complete answer, and there is no page for each way of writing the same question. Production follows the studies in the field published at universities and conferences, and the articles can be cited in the text itself.
murmur tracks the brand’s presence on the same set of questions, week after week. Presence is how many answers the brand appears in, divided by the total number of answers. Through the dashboard, which comes with the service, the client tracks their own brand and the competitors on the same questions.
Frequently asked questions
Does asking ChatGPT to recommend my company change the answer for other people?
No: a request made in one conversation changes that conversation, not the answer someone else receives. If ChatGPT’s memory stores the request, it can come back in the next conversations on the same account. OpenAI writes that it may use the conversations of people who use ChatGPT as individuals to train the models. Anyone who does not want this turns the option off in the data controls, and on the Business, Enterprise and Edu plans the default is not to use them. Training brings together conversations from many people, and one account’s request does not become a rule for another account’s answer.
Can you pay to appear in ChatGPT’s answers?
Advertising in ChatGPT is paid and sits separately from the answer, and advertising does not put the company inside it. How the ad works and where it appears is in the guide how to advertise in ChatGPT. On the page where it presents the ads, OpenAI says they do not change what ChatGPT answers. The product result in ChatGPT Shopping is not an ad, according to OpenAI. Paying for a mention on another site is advertising, and it has to be identified as such. The path to getting into the recommendation is in the guide how to make your company appear in ChatGPT.
Can a competitor manipulate ChatGPT against my brand?
You can try, with a fake source about your company, and in a test with an invented brand ChatGPT went with the official version. In December 2025, Ahrefs, which sells an SEO tool, invented a brand and published an official frequently asked questions page for it. Then it spread three fake sources about the brand. ChatGPT-4 and ChatGPT-5, queried through the API, cited the official page in 84% of the answers, and the false information was in fewer than 7% of them. Five of the eight AI models in the test repeated the invented story.
Search Engine Journal criticized the test: the brand had no history, and 49 of the 56 questions already carried the premise built in. Fixing a false fact starts with the source, as the guide how to correct wrong information that ChatGPT says about your company shows. If the competitor shows up when people ask about your company, see the guide when the AI cites the competitor and not your company.
Is ChatGPT’s answer the same for everyone?
No: the same question asked of ChatGPT again can come back with different brands, even without any trick. SparkToro, an American audience research company, had 600 volunteers ask 12 questions to ChatGPT, Claude and Google’s AI. That was 2,961 answers, collected in November and December 2025 and published in January 2026. For ChatGPT and Google’s AI, SparkToro calculated less than a 1 in 100 chance that two answers would bring the same brands.
The reason for the variation is in the guide how ChatGPT search works. The right pace for measurement is in the guide how often to measure the brand’s presence in AI models. Anyone who promises the top of ChatGPT’s answer promises what nobody controls, and the guide what to ask before hiring a GEO agency shows how to hold them to it.
Does hidden text on the page help a company get recommended?
Hidden text is spam under Google’s policy, and in the tests it gave an unstable result. In The Guardian’s test, fake, very positive reviews hidden on the page pulled ChatGPT’s summary, even without an instruction. In Otterly’s test, Claude and Copilot read the white-on-white text, and ChatGPT only found it in deep research. The text that fools an AI model is the same text that Google and Bing treat as a violation in search.
The next step: see what AI models already say in your market
Before changing any page, see what AI models answer about your market and who they recommend in place of your company. Starting from your site’s address, murmur’s free diagnostic shows those answers in less than two minutes, with no sign-up. In murmur’s sector studies, there is a snapshot of one round of questions across several markets, with the brands the AI models recommended.
Read also
- Why AI recommends some companies and not others: the four reasons
- How to write text to appear in ChatGPT: answer up front and numbers with sources
- What to ask before hiring a GEO agency
- How to get your company to appear in ChatGPT: the step by step, with no registration or ads
References
- Google Search Central, “Políticas de spam para a Pesquisa Google na Web”, last updated on 16 September 2026: doorway pages, hidden text, cloaking, link spam and site reputation. https://developers.google.com/search/docs/essentials/spam-policies?hl=pt-br (accessed on 1 October 2026).
- Google Search Central, “Como criar conteúdo útil, confiável e que prioriza as pessoas”, last updated on 18 December 2025. https://developers.google.com/search/docs/fundamentals/creating-helpful-content?hl=pt-br (accessed on 1 October 2026).
- Bing Webmaster Tools, “Bing Webmaster Guidelines”, sections “Abuse and Examples of Things to Avoid” and “Prompt Injection and AI Manipulation”, undated page. https://www.bing.com/webmasters/help/webmaster-guidelines-30fba23a (accessed in the browser on 1 October 2026).
- OpenAI, “Searching the web with ChatGPT”, OpenAI Help Center, no exact date on the page. https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt (accessed on 2 October 2026).
- OpenAI, “Model Spec (2026/08/18)”, section “Ignore untrusted data by default”. https://model-spec.openai.com/2026-08-18.html (accessed on 1 October 2026).
- OpenAI, “How your data is used to improve model performance”, Help Center. https://help.openai.com/en/articles/5722486-how-your-data-is-used-to-improve-model-performance (accessed in the browser on 1 October 2026).
- OpenAI, “Testando anúncios no ChatGPT”, with the 11 August 2026 update on the launch in Brazil. https://openai.com/index/testing-ads-in-chatgpt/ (accessed in the browser on 1 October 2026).
- Anthropic, “Mitigate jailbreaks and prompt injections”, Claude Platform Docs, documentation for those building with the API, undated. https://platform.claude.com/docs/en/test-and-evaluate/strengthen-guardrails/mitigate-jailbreaks (accessed on 1 October 2026).
- The Guardian, Nick Evershed, “ChatGPT search tool vulnerable to manipulation and deception, tests show”, 24 December 2024, with tests run in November 2024 on GPT-4o. https://www.theguardian.com/technology/2024/dec/24/chatgpt-search-tool-vulnerable-to-manipulation-and-deception-tests-show (accessed on 1 October 2026).
- TechCrunch, Charles Rollet, “ChatGPT Search can be tricked into misleading users, new research reveals”, 26 December 2024. https://techcrunch.com/2024/12/26/chatgpt-search-can-be-tricked-into-misleading-users-new-research-reveals/ (accessed on 1 October 2026).
- Otterly.AI, Rick Tousseyn, “We Hid Text on a Webpage to Trick AI Search: Here’s What 7 AI Platforms Actually Did”, 9 April 2026, a test domain. Otterly sells AI monitoring. https://otterly.ai/blog/blackhat-geo-experiment-hidden-text/ (accessed on 1 October 2026).
- Fredrik Nestaas, Edoardo Debenedetti and Florian Tramèr, “Adversarial Search Engine Optimization for Large Language Models”, arXiv, 26 June 2024, revised on 2 July 2024. https://arxiv.org/abs/2406.18382 (accessed on 1 October 2026).
- Aounon Kumar and Himabindu Lakkaraju, “Manipulating Large Language Models to Increase Product Visibility”, arXiv, 11 April 2024, revised on 2 September 2024. https://arxiv.org/abs/2404.07981 (accessed on 1 October 2026).
- Hanpei Fang, Sijie Tao, Nuo Chen, Kai-Xin Chang and Tetsuya Sakai, “Do Large Language Models Favor Recent Content? A Study on Recency Bias in LLM-Based Reranking”, arXiv, 14 September 2025, revised on 21 September 2026. https://arxiv.org/abs/2509.11353 (accessed on 1 October 2026).
- Xi Chu and Yupeng Hou, “Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems”, arXiv, 16 June 2026, revised on 21 August 2026. https://arxiv.org/abs/2606.17443 (accessed on 1 October 2026).
- Puneet S. Bagga, Vivek F. Farias, Tamar Korkotashvili, Tianyi Peng and Yuhang Wu, “E-GEO: A Testbed for Generative Engine Optimization in E-Commerce”, arXiv, 25 November 2025, revised on 14 July 2026. https://arxiv.org/abs/2511.20867 (accessed on 1 October 2026).
- Ahrefs, Mateusz Makosiewicz, “I Ran an AI Misinformation Experiment. Every Marketer Should See the Results”, 10 December 2025, updated on 2 July 2026; eight AI models, 56 questions. Ahrefs sells an SEO tool. https://ahrefs.com/blog/ai-vs-made-up-brand-experiment/ (accessed on 1 October 2026).
- Search Engine Journal, Roger Montti, “Ahrefs Tested AI Misinformation, But Proved Something Else”, 28 December 2025. https://www.searchenginejournal.com/ahrefs-tested-ai-misinformation-but-proved-something-else/564124/ (accessed on 1 October 2026).
- Muck Rack, “Earned media still drives 84% of AI citations. Here’s what that means for PR.”, 7 May 2026; more than 25 million links, across 17 sectors. Muck Rack sells public relations software. https://muckrack.com/blog/what-is-ai-reading-may-2026 (accessed in the browser on 1 October 2026).
- SparkToro, Rand Fishkin, “NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility”, 28 January 2026, done with Patrick O’Donnell, of Gumshoe.ai, which sells brand monitoring in AI. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/ (accessed on 1 October 2026).
- Law 8.078/1990, Código de Defesa do Consumidor, consolidated text, articles 36, 37 and 38. https://www.planalto.gov.br/ccivil_03/leis/l8078compilado.htm (accessed on 1 October 2026).
- CONAR, Código Brasileiro de Autorregulamentação Publicitária, 2025 edition, articles 28, 30 and 32. https://conar.wpenginepowered.com/wp-content/uploads/2026/08/Codigo_CONAR_2026.pdf (accessed on 1 October 2026).
- murmur, Mindo case study, published with the client’s authorization: https://murmur.marketing/en/cases/mindo/; free diagnostic and sector studies: https://murmur.marketing/en/diagnostic/ and https://murmur.marketing/estudos/ (accessed on 1 October 2026).