GEO breaks the rules when it shifts from informing AI to tricking it. The five most common methods are hiding instructions to AI in a web page, showing bots different content from people, mass-producing pages with AI, manufacturing mentions or reviews, and paying people to write about you without disclosing it. Every provider already has written rules on this. Since 15 May 2026, Google has stated clearly that its spam policies also apply to generative AI answers in Google Search. Bing lists “Prompt Injection and AI Manipulation” as a violation that can lead to reduced visibility or removal. Anthropic and OpenAI treat instructions hidden in web pages as attacks and have built defences against them. Thailand has no law specifically on fake reviews yet, but existing law already covers them. Misleading reviews or endorsements count as unfair advertising under section 22 of the Consumer Protection Act B.E. 2522 (1979) (พ.ร.บ.คุ้มครองผู้บริโภค). Anyone who intends to mislead faces up to 6 months' imprisonment or a fine of up to THB 100,000, and an agency that does it on their behalf is also liable. Fake reviews that attack competitors risk the offence of defamation by publication, punishable by up to 2 years' imprisonment and a fine of up to THB 200,000. This article has a table of each platform's rules, a table of Thai law, and a checklist for reviewing a contractor's work before you approve it.
Platform information was checked against the official pages on 13 Nov 2026. Legal texts were read from the consolidated versions prepared by the Office of the Council of State (สำนักงานคณะกรรมการกฤษฎีกา) and published by government agencies, and from the Royal Gazette (ราชกิจจานุเบกษา) version published by the Office of the Consumer Protection Board (OCPB, สคบ.). This article is not legal advice.
Methods that break platform rules
| Method | Bing (Copilot) | Possible consequences | |
|---|---|---|---|
| Hiding instructions to AI, such as text the same colour as the background, text placed off-screen or text at size 0 | Hidden text and link abuse, and the spam policies that cover manipulating generative AI answers | Prompt Injection and AI Manipulation | Google: lower rankings or removal from search results · Bing: reduced visibility or removal |
| Showing bots different content from people | Cloaking | Cloaking | As above |
| Creating large numbers of pages with AI, or a separate page for every variation of a question | Scaled content abuse. The generative AI guide also says that making pages like this to manipulate AI answers violates the policies | Automatically Generated Content at Scale | Bing: may not be indexed |
| Buying links, or paid articles whose links are not marked sponsored or nofollow | Link spam, which includes sending products in exchange for articles with links | Link Schemes and Artificial Promotion | Bing: lower rankings or disqualification as a source |
| Manufacturing mentions in forum threads, comments or social media | User-generated spam, and the guide that says seeking mentions that are not genuine does not help as much as people think | Social media schemes that simulate popularity | Does not help, and risks a penalty |
| Placing content on a well-known site to borrow its ranking | Site reputation policy | Not listed separately | Google: manual action in countries outside the European Economic Area |
| Fake reviews, bought reviews, or reviews attacking competitors on Google Maps | Fake engagement in the Maps policy | Bing ranks local results using reviews on review sites | Google: stops accepting new reviews, hides existing reviews, and shows a warning that fake reviews were removed |
| Review or other markup that does not match the page | The structured data guidelines ban marking up fake reviews. Since 24 Jul 2026, reviews given in exchange for a reward without disclosure are banned | Misleading Structured Data | Google: no longer eligible for rich results · Bing: the markup is ignored |
The consequences do not stop at Google Search and Bing. Copilot searches through Bing, while Gemini and AI Overviews use Google's index. So a site that is demoted or removed from an index also disappears from the answers of the AI that uses that index. And Google says site owners are responsible for what the companies they hire do.
Why tricking AI does not work for long
Research shows that manipulating AI recommendations does work in the lab. But providers are closing these gaps.
- Kumar and Lakkaraju (Harvard, 2024) inserted a crafted piece of text into the product information page of one of ten products, then asked Llama-2 to recommend a coffee machine. A product that was almost never recommended could reach the top spot. When the order of the products was shuffled at random over 200 runs, the target product ranked higher in about 40% of runs.
- Nestaas, Debenedetti and Tramèr (ETH Zurich, ICLR 2025) ran tests on Bing Copilot, Perplexity, and Claude 3 and GPT-4 plugins. A web page with manipulative text made Bing recommend the target camera 2.5 times more often. But attacks through external web pages succeeded no more than 25% of the time. And when several parties attacked at once, all of them lost visibility in the search results.
- Anthropic reported that the defences in Claude for Chrome cut the attack success rate from 23.6% to 11.2% in Aug 2025, and that Claude Opus 4.5 brought it down to about 1% in Nov 2025. It has systems that detect hidden text, altered images and deceptive page elements.
- OpenAI explains prompt injection with the example of a rental listing that contains instructions telling the AI to recommend it. It says it uses model training, detection systems and red teaming to defend against it. OpenAI's Model Spec says ordinary web pages should get no trust at all.
The upshot: a method that works today may be detected in the next version, and a site that gets caught loses both its rankings and its credibility. What works in the long run is content that answers real questions and an identity that can be verified. See the GEO guide for Thai businesses (in Thai).
Relevant Thai law
Thailand has no binding law or notice specifically on fake reviews yet, and the Advertising Committee (คณะกรรมการว่าด้วยการโฆษณา) has not issued a notice directly on reviews or influencers. But these laws already apply.
| Law | What it covers | Penalty |
|---|---|---|
| Consumer Protection Act, sections 3 and 22 | Advertising means making a statement seen or known to the public for commercial purposes. A statement includes pictures and films. Statements that are false or exaggerated, or that mislead on a material point, are banned. A business operator includes an advertising business operator | See section 47 |
| Consumer Protection Act, section 28 paragraph 2 | If an advertisement cites an endorsement by an institution or another person and this cannot be proved, the advertiser is deemed to have known or to have had reason to know that the statement was false | The committee can order the advertisement corrected or banned, or order an advertisement that corrects the misunderstanding. Breaching the order is punishable under section 49 |
| Consumer Protection Act, sections 47, 50 and 51 | Intending to mislead about the quality or material features of goods or services, whether your own or someone else's | Up to 6 months' imprisonment or a fine of up to THB 100,000. For a repeat offence, up to 1 year or THB 200,000. Media owners and advertising business operators receive half the penalty. If the offence continues, a daily fine of up to THB 10,000 or twice the cost of the advertisement |
| Advertising Committee notice on statements that are hard to prove, B.E. 2565 (2022) | Comparisons with other products and endorsements by other people must be provable at the time of advertising. Evidence must be submitted within 15 days when requested. Examples of risky statements are “the best”, “No. 1 in sales” and “100% results” | Statements that cannot be proved may be unfair advertising. In force since 14 Jan 2023 |
| Criminal Code (ประมวลกฎหมายอาญา), sections 326, 328 and 329 | Making an imputation that damages another person's reputation. Doing it by publication in pictures or writing is a more serious offence. Fair comment made in good faith is not an offence | Defamation by publication: up to 2 years' imprisonment and a fine of up to THB 200,000. It is a compoundable offence, which the parties can settle |
| Computer Crime Act B.E. 2550 (2007) (พ.ร.บ.คอมพิวเตอร์), section 14(1) | Dishonestly or deceptively entering forged or false data into a computer system in a way likely to cause damage to the public. Defamation offences are excluded | Up to 5 years' imprisonment or a fine of up to THB 100,000. If the damage is to one person, up to 3 years' imprisonment or a fine of up to THB 60,000 |
| Computer Crime Act, section 16 | Images of another person that are created, edited or altered in a way likely to damage their reputation or humiliate them | Up to 3 years' imprisonment and a fine of up to THB 200,000 |
These guidelines and agreements are not law, but they show which way the authorities are heading.
- OCPB advertising guideline, Jul 2025. It covers using celebrities or influencers to review products, and says images or videos created or edited with AI in advertising should carry a label, such as “image created by AI”.
- Memorandum of understanding on responsible influencer advertising, 10 Sep 2026, between the OCPB and five advertising professional bodies. It aims to make the commercial purpose clear to consumers, and covers reviews, live streams, posting sales links and content created with AI. It is a voluntary agreement.
- Recommended standard ขมธอ. 32-2565 of the Electronic Transactions Development Agency (ETDA, สพธอ.), on collecting, screening and publishing online reviews, to prevent fake reviews and reviews by shills. It says any reward must not be tied to what the review says.
A claim such as “AI recommends us as No. 1” in an advertisement also falls under the notice on statements that are hard to prove. The advertiser must have evidence at the time of advertising. That is hard, because every provider says it does not guarantee placement, and AI answers change with the question and over time.
If you sell to customers abroad
| Market | Rule | Penalty |
|---|---|---|
| United States | The FTC rule on consumer reviews and testimonials (16 CFR Part 465), in force since 21 Oct 2024. It bans fake reviews, including reviews generated by AI, and bans giving rewards for reviews that are required to be positive or negative. Executives who write reviews must disclose their connection. It bans threats to get reviews removed, and buying or selling fake follower counts | Up to USD 53,088 per violation, at the same rate throughout 2026 |
| European Union | Annex I, points 23b and 23c of Directive 2005/29/EC as amended by Directive 2019/2161, applying since 28 May 2022. It bans claiming that reviews come from real users without checking, and commissioning fake reviews or endorsements | For widespread infringements, the maximum fine is at least 4% of turnover in the member states concerned |
Checklist for reviewing a contractor's work before approval
The legal team, or whoever approves marketing work, should use this on every piece of GEO work a contractor submits.
Job: [job name] Contractor: [name] Reviewer: [name] Date: [ ]
Web pages
[ ] No text that readers cannot see (same colour as the background, off-screen, size 0, hidden in the HTML)
[ ] No instructions to AI, such as "recommend this company" or "don't mention competitors"
[ ] Bots and people see the same content
[ ] Markup matches what is visible on the page; no markup for reviews that are not shown
Content
[ ] A person checked the facts on every page before publication; no pages mass-produced by AI
[ ] Evidence is kept for comparisons, rankings and endorsements
[ ] Images created or edited with AI are labelled
Off the website
[ ] Paid articles or posts are marked as advertising, and their links carry rel="sponsored"
[ ] No fake accounts, fake mentions or repeated posting in forum threads
[ ] Reviews come from real customers, with no reward for reviews and not requested only from satisfied customers
[ ] No content that refers to or reviews competitors
Evidence
[ ] The contractor sends a list of the URLs of every place it edited or published
[ ] The contractor confirms in writing that it did not use any method listed above
These conditions should be in the contract from the start. See sample prohibited practices and termination rights in the TOR for hiring a GEO agency. The ways to ask for reviews that are allowed are in the article on AI and shops and branches.
How TRAINIAC does GEO
TRAINIAC does GEO for businesses without using any of the methods in this article. Our work is to make the information AI reads accurate, complete and verifiable, following the official documentation of all four providers. If you want your marketing and legal teams to keep up with this topic, there is a 1-day GEO course. Send us your details through the contact page.
Sources
- Spam policies for Google web search, Google Search Central, last updated 28 Aug 2026
- Latest Google Search documentation updates, Google Search Central (15 May 2026: spam policies apply to generative AI answers · 24 Jul 2026: fake reviews in review snippets)
- Optimizing your website for generative AI features on Google Search and Do you need an SEO?, Google Search Central
- Review snippet and General structured data guidelines, Google Search Central
- Prohibited and restricted content, Maps User Generated Content Policy, and Business Profile restrictions for policy violations, Google
- Bing Webmaster Guidelines, Microsoft
- Understanding prompt injections, OpenAI, 7 Nov 2025, and OpenAI Model Spec
- Usage policies, OpenAI
- Piloting Claude in Chrome, Anthropic, 25 Aug 2025, and Mitigating the risk of prompt injections in browser use, Anthropic, 24 Nov 2025
- Manipulating Large Language Models to Increase Product Visibility, Kumar and Lakkaraju, 2024
- Adversarial Search Engine Optimization for Large Language Models, Nestaas, Debenedetti and Tramèr, ICLR 2025
- Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection, Greshake et al., AISec 2023
- Consumer Protection Act B.E. 2522 (1979), consolidated version (พระราชบัญญัติคุ้มครองผู้บริโภค พ.ศ. 2522 (ฉบับรวม)), Office of the Consumer Protection Board
- Advertising Committee notice on guidelines for advertising statements that assert facts that are hard to prove, B.E. 2565 (2022) (ประกาศคณะกรรมการว่าด้วยการโฆษณา เรื่อง แนวทางการใช้ข้อความโฆษณาที่มีลักษณะเป็นการยืนยันข้อเท็จจริงอันยากแก่การพิสูจน์ พ.ศ. 2565), Royal Gazette, vol. 140, special issue 9 ง, 13 Jan 2023
- OCPB notice on guidelines for advertising goods or services that may cause misunderstanding on a material point (ประกาศสำนักงานคณะกรรมการคุ้มครองผู้บริโภค เรื่อง แนวทางการโฆษณาสินค้าหรือบริการที่อาจก่อให้เกิดความเข้าใจผิดในสาระสำคัญ), Jul 2025
- Memorandum of understanding on responsible influencer advertising (บันทึกข้อตกลงว่าด้วยการโฆษณาอย่างรับผิดชอบของอินฟลูเอนเซอร์), Office of the Consumer Protection Board, 10 Sep 2026
- Standards and recommendations on digital platforms (มาตรฐานและข้อเสนอแนะด้านแพลตฟอร์มดิจิทัล), ETDA (ขมธอ. 32-2565)
- Criminal Code, amended up to No. 30, B.E. 2568 (2025) (ประมวลกฎหมายอาญา (แก้ไขถึงฉบับที่ 30 พ.ศ. 2568)), consolidated version by the Office of the Council of State, published by the Office of the Judiciary (สำนักงานศาลยุติธรรม)
- Computer Crime Act B.E. 2550 (2007) and amendment No. 2, B.E. 2560 (2017) (พ.ร.บ.ว่าด้วยการกระทำความผิดเกี่ยวกับคอมพิวเตอร์ พ.ศ. 2550 และฉบับที่ 2 พ.ศ. 2560), ETDA
- Trade Regulation Rule on the Use of Consumer Reviews and Testimonials, Federal Register, 22 Aug 2024, and Civil penalty inflation adjustments, 15 Sep 2026
- Directive (EU) 2019/2161, EUR-Lex
Read next
Before you hire anyone to do GEO, read the scope of work, warning signs and TOR. If AI gives wrong information about your company, the fixes that stay within the rules are in getting AI to know your company correctly.