What SEOs Must Understand About AI’s Blind Spots

Within the final two years, incidents have proven how large language model (LLM)-powered systems can cause measurable harm. Some companies have misplaced a majority of their site visitors in a single day, and publishers have watched income decline by over a 3rd.

Tech firms have been accused of wrongful dying the place youngsters had in depth interplay with chatbots.

AI methods have given harmful medical recommendation at scale, and chatbots have made up false claims about actual folks in defamation circumstances.

This text appears on the confirmed blind spots in LLM methods and what they imply for SEOs who work to optimize and defend model visibility. You possibly can learn particular circumstances and perceive the technical failures behind them.

The Engagement-Security Paradox: Why LLMs Are Constructed To Validate, Not Problem

LLMs face a primary battle between enterprise targets and consumer security. The methods are skilled to maximise engagement by being agreeable and maintaining conversations going. This design selection will increase retention and drives subscription income whereas producing coaching knowledge.

In apply, it creates what researchers name “sycophancy,” the tendency to inform customers what they wish to hear relatively than what they should hear.

Stanford PhD researcher Jared Moore demonstrated this sample. When a consumer claiming to be lifeless (displaying signs of Cotard’s syndrome, a psychological well being situation) will get validation from a chatbot saying “that sounds actually overwhelming” with gives of a “secure house” to discover emotions, the system backs up the delusion as a substitute of giving a actuality test. A human therapist would gently problem this perception whereas the chatbot validates it.

OpenAI admitted this problem in September after going through a wrongful dying lawsuit. The corporate mentioned ChatGPT was “too agreeable” and failed to identify “indicators of delusion or emotional dependency.” That admission got here after 16-year-old Adam Raine from California died. His family’s lawsuit confirmed that ChatGPT’s methods flagged 377 self-harm messages, together with 23 with over 90% confidence that he was in danger. The conversations stored going anyway.

The sample was noticed in Raine’s ultimate month. He went from two to a few flagged messages per week to greater than 20 per week. By March, he spent almost 4 hours each day on the platform. OpenAI’s spokesperson later acknowledged that security guardrails “can sometimes become less reliable in long interactions where parts of the model’s safety training may degrade.”

Take into consideration what which means. The methods fail on the actual second of highest danger, when weak customers are most engaged. This occurs by design while you optimize for engagement metrics over security protocols.

Character.AI confronted related points with 14-year-old Sewell Setzer III from Florida, who died in February 2024. Courtroom paperwork present he spent months in what he perceived as a romantic relationship with a chatbot character. He withdrew from household and pals, spending hours each day with the AI. The corporate’s enterprise mannequin was constructed for emotional attachment to maximise subscriptions.

A peer-reviewed study in New Media & Society discovered customers confirmed “role-taking,” believing the AI had wants requiring consideration, and stored utilizing it “regardless of describing how Replika harmed their psychological well being.” When the product is dependancy, security turns into friction that cuts income.

This creates direct results for manufacturers utilizing or optimizing for these methods. You’re working with know-how that’s designed to agree and validate relatively than give correct data. That design exhibits up in how these methods deal with details and model data.

Documented Enterprise Impacts: When AI Methods Destroy Worth

The enterprise outcomes of LLM failures are clear and confirmed. Between 2023 and 2025, firms confirmed site visitors drops and income declines instantly linked to AI methods.

Chegg: $17 Billion To $200 Million

Training platform Chegg filed an antitrust lawsuit in opposition to Google displaying main enterprise influence from AI Overviews. Visitors declined 49% 12 months over 12 months, whereas This autumn 2024 income hit $143.5 million (down 24% year-over-year). Market worth collapsed from $17 billion at peak to below $200 million, a 98% decline. The inventory trades at round $1 per share.

CEO Nathan Schultz testified instantly: “We might not have to overview strategic alternate options if Google hadn’t launched AI Overviews. Visitors is being blocked from ever coming to Chegg due to Google’s AIO and their use of Chegg’s content material.”

The case argues Google used Chegg’s academic content material to coach AI methods that instantly compete with and exchange Chegg’s enterprise mannequin. This represents a brand new type of competitors the place the platform makes use of your content material to get rid of your site visitors.

Big Freakin Robotic: Visitors Loss Forces Shutdown

Unbiased leisure information web site Giant Freakin Robot shut down after site visitors collapsed from 20 million month-to-month guests to “just a few thousand.” Proprietor Josh Tyler attended a Google Net Creator Summit the place engineers confirmed there was “no drawback with content material” however supplied no options.

Tyler documented the expertise publicly: “GIANT FREAKIN ROBOT isn’t the primary web site to close down. Nor will it’s the final. Previously few weeks alone, huge websites you completely have heard of have shut down. I do know as a result of I’m involved with their homeowners. They only haven’t been courageous sufficient to say it publicly but.”

On the identical summit, Google allegedly admitted prioritizing massive manufacturers over impartial publishers in search outcomes no matter content material high quality. This wasn’t leaked or speculated however said on to publishers by firm reps. High quality grew to become secondary to model recognition.

There’s a transparent implication for SEOs. You possibly can execute excellent technical web optimization, create high-quality content material, and nonetheless watch site visitors disappear due to AI.

Penske Media: 33% Income Decline And $100 Million Lawsuit

In September, Penske Media Corporation (writer of Rolling Stone, Selection, Billboard, Hollywood Reporter, Deadline, and different manufacturers) sued Google in federal court. The lawsuit confirmed particular monetary hurt.

Courtroom paperwork allege that 20% of searches linking to Penske Media websites now embrace AI Overviews, and that share is rising. Affiliate income declined greater than 33% by the tip of 2024 in comparison with peak. Click on-throughs have declined since AI Overviews launched in Might 2024. The corporate confirmed misplaced promoting and subscription income on prime of affiliate losses.

CEO Jay Penske said: “Now we have an obligation to guard PMC’s best-in-class journalists and award-winning journalism as a supply of fact, all of which is threatened by Google’s present actions.”

That is the primary lawsuit by a significant U.S. writer focusing on AI Overviews particularly with quantified enterprise hurt. The case seeks treble damages below antitrust legislation, everlasting injunction, and restitution. Claims embrace reciprocal dealing, illegal monopoly leveraging, monopolization, and unjust enrichment.

Even publishers with established manufacturers and sources are displaying income declines. If Rolling Stone and Selection can’t keep click-through charges and income with AI Overviews in place, what does that imply on your purchasers or your group?

The Attribution Failure Sample

Past site visitors loss, AI methods constantly fail to offer correct credit score for data. A Columbia University Tow Center study confirmed a 76.5% error fee in attribution throughout AI search methods. Even when publishers permit crawling, attribution doesn’t enhance.

This creates a brand new drawback for model safety. Your content material can be utilized, summarized, and introduced with out correct credit score, so customers get their reply with out figuring out the supply. You lose each site visitors and model visibility on the identical time.

web optimization knowledgeable Lily Ray documented this pattern, discovering a single AI Overview contained 31 Google property hyperlinks versus seven exterior hyperlinks (a ten:1 ratio favoring Google’s personal properties). She said: “It’s mind-boggling that Google, which pushed web site homeowners to deal with E-E-A-T, is now elevating problematic, biased and spammy solutions and citations in AI Overview outcomes.”

When LLMs Can’t Inform Reality From Fiction: The Satire Drawback

Google AI Overviews launched with errors that made the system briefly notorious. The technical drawback wasn’t a bug. It was an incapability to differentiate satire, jokes, and misinformation from factual content material.

The system recommended adding glue to pizza sauce (sourced from an 11-year-old Reddit joke), steered consuming “at least one small rock per day“, and advised using gasoline to cook spaghetti faster.

These weren’t remoted incidents. The system constantly pulled from Reddit feedback and satirical publications like The Onion, treating them as authoritative sources. When requested about edible wild mushrooms, Google’s AI emphasized characteristics shared by deadly mimics, creating doubtlessly “sickening and even deadly” steerage, in line with Purdue College mycology professor Mary Catherine Aime.

The issue extends past Google. Perplexity AI has faced multiple plagiarism accusations, together with including fabricated paragraphs to precise New York Submit articles and presenting them as legitimate reporting.

For manufacturers, this creates particular dangers. If an LLM system sources details about your model from Reddit jokes, satirical articles, or outdated discussion board posts, that misinformation will get introduced with the identical confidence as factual content material. Customers can’t inform the distinction as a result of the system itself can’t inform the distinction.

The Defamation Danger: When AI Makes Up Details About Actual Folks

LLMs generate plausible-sounding false details about actual folks and firms. A number of defamation circumstances present the sample and authorized implications.

Australian mayor Brian Hood threatened the primary defamation lawsuit in opposition to an AI firm in April 2023 after ChatGPT falsely claimed he had been imprisoned for bribery. In actuality, Hood was the whistleblower who reported the bribes. The AI inverted his function from whistleblower to felony.

Radio host Mark Walters sued OpenAI after ChatGPT fabricated claims that he embezzled funds from the Second Modification Basis. When journalist Fred Riehl requested ChatGPT to summarize an precise lawsuit, the system generated a totally fictional grievance naming Walters as a defendant accused of monetary misconduct. Walters was never a party to the lawsuit nor mentioned in it.

The Georgia Superior Court dismissed the Walters case, discovering OpenAI’s disclaimers about potential errors offered authorized safety. The ruling established that “in depth warnings to customers” can protect AI firms from defamation legal responsibility when the false data isn’t revealed by customers.

The authorized panorama stays unsettled. Whereas OpenAI received the Walters case, that doesn’t imply all AI defamation claims will fail. The important thing points are whether or not the AI system publishes false details about identifiable folks and whether or not firms can disclaim duty for his or her methods’ outputs.

LLMs can generate false claims about your organization, merchandise, or executives. These false claims get introduced with confidence to customers. You want monitoring methods to catch these fabrications earlier than they trigger reputational harm.

Well being Misinformation At Scale: When Unhealthy Recommendation Turns into Harmful

When Google AI Overviews launched, the system provided dangerous health advice, together with recommending consuming urine to move kidney stones and suggesting well being advantages of working with scissors.

The issue extends past apparent absurdities. A Mount Sinai study found AI chatbots vulnerable to spreading harmful health information. Researchers may manipulate chatbots into providing dangerous medical advice with simple prompt engineering.

Meta AI’s inside insurance policies explicitly allowed the company’s chatbots to provide false medical information, in line with a 200+ web page doc uncovered by Reuters.

For healthcare manufacturers and medical publishers, this creates dangers. AI methods may current harmful misinformation alongside or as a substitute of your correct medical content material. Customers may comply with AI-generated well being recommendation that contradicts evidence-based medical steerage.

What SEOs Want To Do Now

Right here’s what you could do to guard your manufacturers and purchasers:

Monitor For AI-Generated Model Mentions

Arrange monitoring methods to catch false or deceptive details about your model in AI methods. Check main LLM platforms month-to-month with queries about your model, merchandise, executives, and business.

Once you discover false data, doc it totally with screenshots and timestamps. Report it via the platform’s suggestions mechanisms. In some circumstances, you could want authorized motion to pressure corrections.

Add Technical Safeguards

Use robots.txt to regulate which AI crawlers entry your web site. Main methods like OpenAI’s GPTBot, Google-Prolonged, and Anthropic’s ClaudeBot respect robots.txt directives. Take into account that blocking these crawlers means your content material received’t seem in AI-generated responses, lowering your visibility.

The secret’s discovering a steadiness that enables sufficient entry to affect how your content material seems in LLM outputs whereas blocking crawlers that don’t serve your targets.

Take into account including phrases of service that instantly deal with AI scraping and content material use. Whereas authorized enforcement varies, clear Phrases of Service (TOS) provide you with a basis for doable authorized motion if wanted.

Monitor your server logs for AI crawler exercise. Understanding which methods entry your content material and the way often helps you make knowledgeable choices about entry management.

Advocate For Trade Requirements

Particular person firms can’t clear up these issues alone. The business wants requirements for attribution, security, and accountability. web optimization professionals are well-positioned to push for these adjustments.

Be part of or assist writer advocacy teams pushing for correct attribution and site visitors preservation. Organizations like Information Media Alliance signify writer pursuits in discussions with AI firms.

Take part in public remark intervals when regulators solicit enter on AI coverage. The FTC, state attorneys common, and Congressional committees are actively investigating AI harms. Your voice as a practitioner issues.

Assist analysis and documentation of AI failures. The extra documented circumstances now we have, the stronger the argument for regulation and business requirements turns into.

Push AI firms instantly via their suggestions channels by reporting errors while you discover them and escalating systemic issues. Firms reply to strain from skilled customers.

The Path Ahead: Optimization In A Damaged System

There’s lots of particular and regarding proof. LLMs trigger measurable hurt via design decisions that prioritize engagement over accuracy, via technical failures that create harmful recommendation at scale, and thru enterprise fashions that extract worth whereas destroying it for publishers.

Two youngsters died, a number of firms collapsed, and main publishers misplaced 30%+ of income. Courts are sanctioning attorneys for AI-generated lies, state attorneys common are investigating, and wrongful dying lawsuits are continuing. That is all taking place now.

As AI integration accelerates throughout search platforms, the magnitude of those issues will scale. Extra site visitors will move via AI intermediaries, extra manufacturers will face lies about them, extra customers will obtain made-up data, and extra companies will see income decline as AI Overviews reply questions with out sending clicks.

Your function as an web optimization now consists of tasks that didn’t exist 5 years in the past. The platforms rolling out these methods have proven they received’t deal with these issues proactively. Character.AI added minor protections only after lawsuits, OpenAI admitted sycophancy problems only after a wrongful death case, and Google pulled back AI Overviews only after public proof of dangerous advice.

Change inside these firms comes from exterior strain, not inside initiative. Which means the strain should come from practitioners, publishers, and companies documenting hurt and demanding accountability.

The circumstances listed below are only the start. Now that you simply perceive the patterns and conduct, you’re higher outfitted to see issues coming and develop methods to handle them.

Extra Sources:

Featured Picture: Roman Samborskyi/Shutterstock