Search engines were designed to do several things at once: Set up a field of options, move the user to one of them, and keep the person inside the decision so that the engine never owns the option. That last part was no accident. It was a debt structure. Major language forms were developed without them. They are designed to answer a question directly, which is a completely different task, and the design choices that follow from it change what visibility looks like, what vulnerability looks like, and what the name means. ROI It can literally mean that when something is sending you traffic it wasn’t designed to send traffic in the first place.
Two Systems, Two Functions
The job description of a search engine is long. It scans the web, identifies it, compares a number of candidate results against the query, presents it as a ranked list, and waits for someone to make a decision to click. SERP itself has been pulling in retention for years now, with galleries, rich snippets, response boxes, location maps, video carousels, and AI Overview all layer on features that keep the user on the page for longer and drive a few of them to third party sites. But the basic contract remained the same. The engine offers options. The user selects one. The user owns the option.
The LLM does not offer electives. It reveals the answer. The quote, if it appears, does not work as a routing tool. It is close to the underlying artifact produced by the retrieval pipeline, or in some contexts, the privacy fence, or both at the same time. Whichever reading you choose, none of them describe a system designed to send traffic elsewhere. The system is designed to solve the query on the spot.
That difference sits at the bottom of every metric discussion in this space. When practitioners ask what the LLM referral rate is, what the referred traffic number looks like, what the clicks are from the AI response, they are asking questions that take a non-architectural route. Whatever traffic comes in is a product, not a design goal, and confusing both is the first mistake in almost every conversation about AI is ROI visibility.
Credit Area Moved
The person in the click decision was a SERP shield. If the link we have selected leads to some dangerous, misleading, or offensive site, the engine can point to a list of options and the user’s agency in choosing one. The engine has yet to publish the claim. It presented 10 candidate sources, the user selected one, and whatever happened next was not the engine’s programming output. That’s no small feat. This is why the protections of Section 230 were drafted the way they were, and why the algorithmic standard has traditionally been regarded as something different from direct speech.
LLMs do not have equal standing shields. The program generates the answer directly, in its own voice, without an option field or source selected by the user. Most of the debt that SERPs were designed to remove resides with the output generation model, and the lawsuits that have already gone through the courts are beginning to draw the edges of that area.
Walters v. OpenAI was dismissed on summary judgment in May 2025, and the decision rested largely on OpenAI’s self-examination and a professional reader who reasonably knew that a chatbot could dream. That learning protects general-purpose chatbots from the consumer from some type of lawsuit. It does not support all products that use the language model. In a separate case, Air Canada was found guilty of false statements by a customer service chatbot about its bereavement payment policy, because a customer could rely on an airline support agent for accurate information about that airline’s policies. Reasonable reliance is a key legal term, and the more specialized and authoritative a chatbot becomes, the more difficult it is to defend against disclaimers.
The active case is still drawing the line. OpenAI is currently facing multiple lawsuits related to allegations that ChatGPT has led users to commit suicide or dangerous disappearances, many involving children. The New York Times’ copyright lawsuit against OpenAI was allowed to proceed by a federal judge in March 2025, and Anthropic settled with the authors in August 2025 for a reported sum of billions. European GDPR complaints continue to flow through Noyb. Battle v. Microsoft is still live. None of these results have been settled, and some will be overturned for the same reasons of overruling that settled Walters. The point is not that LLM operators will lose all cases. The point is that the credit area now sits with the system that produces the output, whether each claimant wins or loses, and each product structure against the LLM gains a specific version of that area when it uses the output of the system in its customer-facing work.
The Denominator Problem
The most common argument against investing in AI visualization work sounds presumptuous until you take a closer look at what it measures. The argument goes roughly: ChatGPT and others send little referral traffic, somewhere in the low single digits of total inbound traffic, so why reallocate budget to a channel that doesn’t move the needle? Candidate research pegs the combined share of AI referrals at around 1% of publisher traffic. That number is real. At first reading, it seems to close the ROI question cleanly.
It doesn’t cover anything. The problem is the denominator.
While the share of AI traffic is almost stable, the overall volume of search-driven traffic has fallen for most publisher categories. The same web data shows organic traffic to news publishers dropped from about 2.3 billion visits in mid-2024 to less than 1.7 billion in May 2025, with a loss of more than 600 million visits in less than a year. Business Insider searches fell 55% between April 2022 and April 2025, HuffPost lost nearly half of its search traffic, and The New York Times saw its share of desktop search and mobile traffic slide from 44% to 37%. Zero-click searches increased from 56% to 69% between May 2024 and May 2025 as AI Overview was expanded throughout the SERP. A Reuters Institute survey of 280 media leaders as of late 2025 found they expect a 43% decline on average over the next three years.
Read against that backdrop, a stable percentage of the shrinking pie is unstable. It is a loss. Skeptics who point to the 1% number measure the relative share of traffic that falls under them, and treat the overall decline as if it were a steady state. The real question is not whether LLMs are sending meaningful traffic yet. The real question is whether the channel that was sending meaningful traffic is still doing what it was doing, and the answer is clearly no. The denominator is moving, and any ROI figure focused on the old denominator is the previous figure, not the current one.
What Billions Mean
If the contradiction of the purpose of the design and the responsibility of the debt and the value still leaves room for doubt, the last place to look is expressed in preference. Which companies with the most complete internal data about user behavior actually make their money?
The answer is clear. The five largest US cloud infrastructure and AI providers have committed between $660 and $690 billion in 2026 revenue, nearly double 2025 levels. Alphabet alone is expected to generate between $175 and $185 billion by 2026, more than doubling its 2025 spending of $91 billion. Microsoft, Amazon, Meta, and Oracle all use aggressive curves in the same way. The most significant number, and one that ends the general debate, comes from Bank of America credit strategists who estimate that AI capex will reach 94% of operating income in 2025 and 2026, up from 76% in 2024.
That is not the case for a protective fence. Hedging is part of the income, used to avoid being caught flat-footed if a competitor’s bet pays off. Companies don’t put 94% of working capital into a division two years in a row unless leadership truly believes the division is a business. And those leadership teams have access to data that the rest of us don’t. They can see inside their products, their user behavior shifts, their team analysis, their business pipeline conversations. They are legally bound to spend their shareholders’ money in a way that reflects what they really see, and what they convey to them is a structure that produces specific answers rather than a limited list of options. To believe that search-as-we-know-it remains the gold standard, you have to believe that dozens of CEOs, boards, and senior leadership teams with only decades of internal data are reading their numbers wrong, while the outside industry with none of that data is reading the market correctly. That’s not a pencil.
The human behavior side of the equation makes the same point in a different register. Every labor-saving technology ever introduced has reshaped the landscape faster than skeptics predicted, because mental efficiency is not a popular commodity. It is a survival behavior, linked during long periods when calories were scarce, and shortcuts were essential. When a new tool comes along that makes a task easier, adoption is not a matter of whether. It’s a matter of how fast it goes and which curve it follows. ChatGPT now has around 900 million weekly active users, up from 200 million 18 months earlier, and the full category has surpassed one billion active users across all platforms. Behavior has already changed. Money has already changed. The only thing that hasn’t fully evolved is the measurement framework that many doctors still use to evaluate the channel.
Which brings the question back to the one you should really be asking. What do you do when there is no ROI with the old definition, and you can’t ignore the channel? The honest answer is that brands will need to invest in visibility work where their return can be expressed in clicks or referral traffic, because clicks and referral traffic are artifacts of previous design. Like a cited source, a grounded source, a reliable source within the response is a different type of visibility, and will require a different type of measurement. The teams that get that first won’t do it because they got the ROI case that convinced their CFO. They will be doing it because they look at capex curves, behavior curves, and credit curves, and conclude that the channel is the future, regardless of whether the spreadsheet knows how to score it now.
If this comes up somewhere in your work, or if it reads poorly where you sit, I’d love to hear about it. The change that is happening now is too big for any one doctor area, and the best signal I get comes from the conversations that start after the article ends.
Additional resources:
This post was originally published on Duane Forrester Decodes.
Featured Image: Krot_Studio/Shutterstock; Paulo Bobita/Search Engine Journal
