As AI agents start to browse, compare and book on behalf of people, some of the most reliable profit engines on the internet are losing their grip. The deciding factor is whether a business can still collect money at the moment it delivers value. Companies that sell human attention through ad slots and sponsored rankings look exposed, while those with unique supply, trust guarantees or deeply embedded workflows appear far better protected. Meta Muse, a consumer agent that runs its own computer in the cloud, makes this shift concrete.
What Meta Muse changes
Meta Muse runs on its own cloud-based virtual computer. When a user hands it a task, it does not just look things up in an internal database. It opens a separate browser on the server side and drives a virtual mouse and keyboard to move through web pages, click links, check booking pages and work with complex site elements. It never needs access to the user’s own device or to a desktop browser such as Chrome or Safari.
Skilled users could already do similar work with Claude by granting the right tool permissions. ChatGPT and Claude Cowork, however, remain underused or poorly understood by non-technical audiences. Muse packages the same agentic capability into an app that mainstream consumers can pick up right away. It connects to apps such as Gmail, Google Calendar, OpenTable, DoorDash, Uber and Instagram, and asks the user to review sensitive actions like sending an email or posting to Instagram.
Mark Zuckerberg has described how Meta plans to make money from it. Muse is offered for free with generous token allowances (tokens are the units of text an AI model processes) to build a large user base. Later, Meta intends to take small transaction fees on purchases and bookings made through Muse. Once adoption is solid, other revenue layers seem likely to follow.
The point of monetization
A useful lens for judging which businesses are at risk is the point of monetization: when a company collects payment relative to the moment the customer actually receives value.
| Business type | When it gets paid | Why |
|---|---|---|
| Consumer products | At the point of sale | Ads, packaging and refund guarantees raise perceived value so buyers pay upfront |
| Legal services | Partly upfront, partly on success | Quality cannot be verified in advance, so contingency fees take a share of the payout |
| Dating apps | Before any value appears | A successful match means the user leaves |
Dating apps show how a weak point of monetization can undermine a business that looks ideal on paper. They need little capital, cost little to run, earn high gross margins and can benefit from network effects. Finding a life partner is one of the most important outcomes in anyone’s life, and a person might gladly pay $5,000 or $10,000 after meeting a spouse.
The apps cannot collect that money. Success sends couples off the platform, and nobody will pay $500 after a good first date. So the apps charge around $20 a month upfront, while users are still unsure the service will work. To make up the gap, they sell add-ons: Super Likes and location changes on Tinder, the option to hide one’s age, and digital roses on Hinge, priced at 3 for $9.99, 12 for $29.99 and 50 for $74.99. These features only nudge the odds of a match, so many users see them as gimmicks, and fatigue sets in.
The ad-funded web breaks first
For decades, the open web ran on a simple bargain. A site like Investopedia answered a reader’s question for free and placed banner and video ads around the article. Ad networks such as Google Ads made this work: readers got information, and publishers earned a fraction of a cent per impression.
AI search breaks that loop. Google AI Overviews and ChatGPT Search answer questions directly on the results page. Even when they cite a source, users rarely click through, so page views and ad impressions fall. Search referrals to publishers have dropped by 20% to 50%, and one data set shows a 40% year-over-year decline. Alphabet’s Google Network revenue, which comes from ads on third-party sites, peaked at $32.8 billion and has slipped to $29.5 billion over the last twelve months.
Agents that browse on a user’s behalf could push this further. Calling it the death of the open web may sound sensational, but the traffic data suggests the concern is real.
▲ Falling search traffic to the open web
Why sponsored listings lose power
Marketplaces earn commissions and fulfillment fees, but a growing share of their profit comes from advertising. Amazon’s advertising revenue grew from $12.6 billion in 2019 to $76.1 billion over the last twelve months. Much of that money comes from monetizing impatience. Clorox pays for the top slot in an Amazon search for its wipes so cheaper knock-offs do not take the sale, and Foot Locker buys a sponsored ad above its own store listings in Google search.
For shoppers, these ads add little. A sponsored Crest listing at the top of a search for Crest toothpaste is harmless but adds nothing. At worst, sponsored results push the real results far down the page. On Yelp, a search for restaurants in San Francisco can return a full screen of paid listings before a single organic one.
That matters because of consumer surplus, the value a customer gets beyond what they pay for. Someone choosing lunch checks a few basic thresholds such as price, distance and healthiness. If Chipotle meets them and then adds free chips or a drink, that extra value builds loyalty. Platforms that only charge for attention and create no surplus are structurally weak.
An AI agent does not get tired or annoyed by scrolling. It reads every listing and picks the one that fits the user’s criteria. It will not choose a product simply because it sits in one of the top four sponsored slots. That directly threatens marketplaces that charge sellers for higher ranking.
Apartments.com illustrates the risk. Renters browse for free, while apartment communities pay for Diamond, Platinum or Gold packages that promise 300%, 100% or 33% more exposure than the base Silver tier, along with top placement, Matterport 3D tours and high-resolution photos. Because leases are not signed on the site, an agent could gather listing details, match them with MLS or Zillow records and contact the property manager directly. Still, claims of total disruption look overstated, since rich media such as interactive 3D tours are hard for an AI to reproduce.
Who holds up
Not every platform faces the same danger. Airbnb draws skepticism, but it is better protected than a plain directory. Its photo carousels and map browsing are polished, and it offers a trust layer that matters when staying in a stranger’s home: guarantees, host vetting and a “Guest favorite” badge that signals verified quality. An agent could compare prices when the same property also appears on Vrbo or Booking.com, but Airbnb holds a large amount of inventory found nowhere else.
Enterprise software looks even sturdier. Salesforce and ServiceNow are protected by three barriers:
- Habit: employees are trained on specific interfaces, and companies avoid retraining them.
- Fear of loss: moving core data and systems carries real operational risk.
- Mission-critical integration: the software is woven into a company’s daily workflows.
AI agents are more likely to work through these systems’ APIs than to replace them. New companies are a different story: a startup may use AI to build internal tools from scratch instead of buying a legacy suite.
All of this comes down to two questions for any business:
- Does the service still provide distinct value?
- Can it collect money at the exact moment it delivers that value?
The danger grows when those two points come apart. If an AI answers a finance definition directly, the reader never visits Investopedia, and its ad revenue never materializes.
A hotel search test shows the limits
Aggregators are not about to vanish. Muse was asked to find hotel rooms in Florence for April 15 to 26, an 11-night stay, at no more than $300 a night. It returned five options, all pulled from Booking.com. Asked to compare those rates with each hotel’s own website, it spent 14 minutes checking just five hotels.
| Result of the price check | Hotels |
|---|---|
| Booking.com significantly cheaper | 3 |
| Same price | 1 |
| Direct booking cheaper | 1 |
The one cheaper direct rate came only from a 10% discount for stays longer than ten days. Crawling individual websites in real time takes a lot of computing power and time. That is why aggregators, which pull scattered supply into one database, remain the path of least resistance even for agents. Booking.com also keeps the payment and booking step on its own platform, and its Genius loyalty program, with levels for account holders, guests with five or more stays and guests with 15 or more, accounts for more than half of its room nights.
▲ An agent comparing hotel prices
Shipping fast versus shipping perfect
How companies launch products has also shaped the AI race. A minimum viable product, or MVP, is a bare-bones version that early users can test and critique so the team can improve it quickly. The approach fails in high-stakes fields such as medical testing, but it suited consumer generative AI. OpenAI shipped early ChatGPT despite frequent hallucinations, meaning confident but wrong answers, and kept improving it based on user feedback.
Large incumbents struggle to do the same because of legal liability, compliance and brand risk. Google acquired DeepMind about 15 years ago and held advanced conversational models internally, yet waited for years over fears of errors and bad press. OpenAI captured the first wave, and Anthropic followed.
The opposite approach is a maximum possible product, or MPP: the most polished product technology allows at launch. Apple works this way. Steve Jobs pushed Corning to produce scratch-resistant glass for the 2007 iPhone that had been considered impossible, and Dyson turned the vacuum cleaner into a desirable premium device with the same mindset. Apple Intelligence features have taken a long time to arrive because Apple customers expect things to just work.
Today’s standalone agents still break often. Muse regularly fails or repeats the same output. If Apple eventually ships a polished agent at the operating system level, with broad system permissions across apps, it could threaten standalone agent apps. The MPP approach carries its own danger, though: in a fast platform shift, falling behind on user habits can be fatal.
Products beat benchmarks
Leadership in AI has changed hands quickly, from Anthropic to Google and back to Anthropic, with Meta’s Muse now in the spotlight and SpaceX pushing Grok. Most people, however, use products, not models. Benchmark scores matter mainly to researchers. Even if Muse scores lower than Claude on standard tests, it can complete practical tasks that Claude cannot, which may make it more valuable to everyday users. Early ChatGPT spread largely because of its simple interface with a single input bar.
Model quality alone is a weak moat. Frontier capabilities are typically matched by free open-source models or rival labs within six to nine months. Unless a model is something like three times better than the next best, distribution and workflow integration carry more weight. That pushes model makers into vertical products such as Claude Code and OpenAI Codex, and possibly into accounting, marketing and finance tools for small businesses. This is a projection, not a settled outcome.
Cursor shows where loyalty sits. After SpaceX acquired the AI code editor, OpenAI announced on August 28, 2026 that it would wind down Cursor’s API access by November 12, 2026. Cursor switched its default models to Anthropic and Google models, and its users stayed. Loyalty appears to belong to the product and its workflow, not the model vendor.
Cash flow is the final factor. Meta’s Family of Apps has produced quarterly operating income as high as $30.8 billion, so it can offer Muse for free without fees or ads. Google and Apple enjoy similar advantages. Pure-play AI labs have to earn money from consumers directly. In-chat ads in ChatGPT’s free tier, such as a dog food ad shown during a question about a vegetable’s nutrition, and credit settings that make Claude Fable users manage token use, may put them at a disadvantage against rivals that give away more for free.
What to do with this
The more AI agents stand between customers and businesses, the weaker attention-based models become, and the stronger businesses that turn distinct value into revenue on the spot look. If you run or are building a business, these checks can help:
- Map where you charge customers compared with when they actually get value, and look for a gap.
- Check how much of your revenue depends on sponsored placements or add-on upsells rather than core value.
- If you earn money from content, track whether AI search and summaries are cutting your referral traffic.
- Build things agents cannot easily replace: unique supply, trust guarantees and perks customers feel as extra value.
- If you build AI products, create user data and habit loops so customers stay even if you switch the underlying model.
- Match your launch style to your field: fast MVPs suit consumer software, while hardware and regulated sectors need polish first.
These frameworks are a way to read business structures, not a recommendation to buy or sell any company’s stock, and the forecasts above are interpretations of current trends rather than certainties.