Who Pays When an Agent Reads the Web?
AI agents can use valuable information without sustaining the people who produce it. A forecast of the commercial web through bargaining power, substitutes and the price of tomorrow's reporting.
A publisher can put an advertisement in the page an AI reads. It cannot assume the advertisement will survive the journey to the person asking the question.
That gap is becoming a commercial dispute. Publishers want compensation when assistants use their work. Assistant providers want control over the answers and advertising they present. An advertiser buying space between them needs to know what, exactly, has been delivered.
In July, TIME began testing one answer to that problem. Working with ad-tech company Mobian, it placed sponsored, FAQ-style brand material in machine-readable versions of its pages. Ally Bank and Project Management Institute were among the early buyers. Ordinary visitors received the normal webpage; selected AI systems could receive text prepared for machines. The sponsored blocks were labeled. Pricing and evidence of commercial effectiveness were not disclosed in Digiday’s original report.
Less than two weeks later, Perplexity said it had blocked TIME’s Markdown advertising from influencing its search index and agents. It did not explain the implementation in the reporting. We have a documented experiment and a reported refusal, rather than proof of either reliable influence or comprehensive filtering. Digiday’s follow-up captures the disagreement.
The case is small, but the question is large: who pays for information when the system using it can bypass the interface that financed its production?
Our argument is that agent-mediated browsing separates activities that a page visit once brought together: obtaining information, encountering advertising and completing a transaction. The resulting payments will depend on what each participant can refuse or replace. Useful information can earn little when substitutes are easy to find; control over a customer’s task can command a fee even when someone else supplied the evidence.
That gives us a starting point for forecasting the market. It also leaves a harder question: will the people able to charge enough be the people whose work needs to continue?
What changes when the visit disappears?
A useful part of the advertising-supported web ran on a reciprocal arrangement. Publishers made information available. Search engines indexed it and sent visitors. Some visitors encountered ads, subscribed, followed an affiliate link or became customers. The exchange was uneven, but the publisher had a place in which attention could become revenue.
An assistant can interrupt that sequence. It may retrieve the useful material, combine it with other sources and answer inside its own interface. The source contributes to the task without receiving the same visit.
For a task such as choosing project-management software, a person might otherwise visit a comparison article, inspect a vendor’s features and check pricing. An assistant could collect that information across several sources and return a shortlist. The publisher’s reporting can help the decision even if the person never opens the article. A sponsored FAQ in the retrieved page might be ignored, used as a brand claim or repeated without its disclosure. Those are different outcomes.
A direct page visit
- PublisherCombines reporting and advertising on its page
- PersonVisits that page and may encounter its ads
An assistant-mediated answer
- PublisherSupplies information, possibly with sponsorship
- AssistantSelects evidence and controls what it presents
- PersonReceives the answer and any accepted, labeled placement
This extends the question in When Your Customer Stops Opening Your App: a service can remain useful while another system carries the customer’s task. For a publisher, losing the interface can also mean losing the place where its contribution was monetized.
What makes useful information worth paying for?
Suppose the software review uncovers a limitation that changes which product the customer should choose. Other accessible pages repeat the finding. The original reporting made the information available, but the assistant may now obtain an adequate account elsewhere. The fact can remain valuable while the original publisher loses leverage over this retrieval.
We need to distinguish the value created by a better decision, the price paid for an input and the return its supplier keeps after costs. A publisher’s reporting costs matter to its survival; they do not establish what every buyer will pay. The assistant’s computing bill creates the same problem on the other side. A costly answer is not necessarily an answer someone values enough to finance.
For a source, the relevant commercial comparison is what the assistant could accomplish without that particular deal. Exclusive reporting, current specialist data and reliable evidence can make the alternative worse. Easy substitutes make it better. Access rights and the ability to withhold material affect which alternatives are feasible; posting a price does not settle those rights.
The assistant’s maximum payment depends on the extra return it expects from using the source, after other affected costs. That could involve retained subscriptions or better task completion, rather than a separate charge for each answer. The publisher’s minimum depends on what supplying this use costs and what it gives up compared with refusing. If the minimum exceeds the maximum, that proposed bargain has no mutually acceptable price.
Where there is room for a deal, several prices can leave both parties better off. Which one they accept depends on bargaining: competing suppliers and buyers, their information, their patience and their ability to walk away. The following invented example makes that allocation visible. It uses gains relative to disagreement, a central idea in Nash’s bargaining analysis, without estimating bargaining strength for this market.
Changing the bargaining share redistributes the available gains without increasing them. Better alternatives can change how much there is to divide. The equation exposes those assumptions; it cannot tell us what share is fair or predict a price without evidence about the participants.
This turns the question from which payment format is technically possible to why anyone would accept its terms. Publishers and assistants have at least three attempted answers.
First answer: carry the publisher’s advertising
The proposed bargain sounds familiar: pay for access without advertising, or accept advertising that subsidizes the information.
For human visitors, publishers can place content and ads together in an interface they control. With an assistant, another system decides what to retain and show. Including promotional text in a response cannot create a placement inside that system. The products therefore need different names and evidence:
| What is being sold? | What delivery would establish |
|---|---|
| A sponsored block supplied to a machine | An eligible system received the material |
| Paid entry into a comparison | The offer entered the declared consideration process |
| A labeled placement shown to a person | The agreed advertising exposure occurred |
| A qualified lead or transaction | The contracted action happened, subject to its definition and verification |
None automatically establishes the next. A transaction can also happen without being caused by the advertising. The novelty of the recipient does not excuse an old measurement mistake.
“Selling ads to agents” is therefore loose language. Advertisers buy the placement. Agents are the intended intermediaries or machine audience. What the advertiser ultimately values might be awareness, consideration or an additional customer, but a server log measures a much earlier event.
A sponsored-access agreement would need to specify whether the assistant must show a labeled offer, preserve attribution or record a particular placement. It would also need to address irrelevant offers, cached copies and answers combining several sources. Downloading a block and displaying an ad have different costs and value.
For the software comparison, one possible agreement would supply the review at a lower price if the assistant displayed a separate, labeled sponsor offer. The assistant could still judge the shortlist against the customer’s requirements. An agreement requiring the sponsor to rank first would purchase something else: influence over the comparison itself. This is an illustrative contract distinction, not a claim about TIME’s terms.
The assistant has a reason to choose subsidized access when the saving exceeds the additional presentation and integration costs, including any damage to user trust. The publisher needs access payments and sponsor receipts to beat the costs and its alternatives. A source offering an easily replaced review may find that the assistant prefers another source to either version of its offer.
The user’s rules impose a further boundary. An instruction to provide an independent comparison does not become permission to sell recommendation priority because a source offers cheaper access. A separate, disclosed placement can be compatible with that comparison. Changing the recommendation for a sponsor is a different arrangement.
Nor does a fetch establish that the assistant operator accepted a contract. Identification, agreement and enforcement are work the commercial system still has to perform.
That gives us a possible bargain, but it depends on the assistant agreeing to carry the placement. Perplexity’s reported refusal shows why a publisher cannot assume that agreement. If carrying the ad is the obstacle, the next answer is to sell the useful information on its own.
Second answer: pay for the information
Charging for information has a simple appeal: the assistant can use the reporting without carrying the source’s ads, and the publisher can be paid for contributing to the task. There is evidence that the old referral bargain needs another source of income.
In their working paper, Saharsh Agarwal and Ananya Sen randomized whether a Chrome extension hid Google AI Overviews. Their 17 June copy reports that AI Overviews reduced outbound organic clicks on searches where an overview would appear. This is a conditional finding from a short US desktop experiment, not a prediction of every publisher’s traffic loss. Sponsored-click differences were undetected, but the authors warn that those clicks were rare and the test had limited power. Full paper
That distinction matters to the financing question. A source can lose valuable visits before the intermediary loses equivalent advertising income. The costs of the transition need not arrive at the same place or at the same time.
The market for paid information is already taking shape. TollBit describes publisher-set usage pricing, authenticated requests and separate licenses for summarization and full display. Cloudflare’s Pay Per Crawl design uses priced requests and payment-intent headers around an HTTP 402 response. These mechanisms allow a source to ask for payment. They cannot make a crawler buy. TollBit, Cloudflare’s design
Microsoft announced its Publisher Content Marketplace in February, describing publisher-defined terms, usage reporting and licensed material to support AI-generated answers. Google’s June account also described paid news partnerships and a broader pilot for websites contributing to the freshness and factuality of generated answers. These are documented partnership mechanisms, rather than evidence that every publisher can replace its lost revenue. Microsoft’s announcement, Google’s account
These examples establish that companies can pay for information without carrying the source’s advertising. Whether a publisher can earn enough this way depends on which content commands payment, what use is covered and who determines its contribution.
Our earlier bargaining question now becomes concrete: is this licensed input better enough than the alternatives to justify its price? Reliable delivery can add value by reducing retrieval and checking effort. The publisher still needs actual paid demand; every blocked request is not a customer waiting to pay.
Would enough machines actually pay?
Paid access succeeds only if useful machine activity turns into paying demand. The traffic figures show why publishers are interested, but also why the distinction matters.
The infrastructure evidence shows a substantial change. Cloudflare reported in September that daily requests from AI agents on its network had grown more than 1,700% over the preceding year. Its July retrospective said non-human traffic had crossed 50%, and that training accounted for 52% of crawler requests classified by purpose in June. These are different categories and denominators. They do not establish that most browsing is now autonomous shopping. September account, July retrospective
A training crawler gathers material for a development pipeline. A search indexer prepares future discovery. A retrieval system fetches information for an answer. A delegated agent may compare offers or request an action. Other bots monitor services, collect prices or probe for vulnerabilities. Their requests cannot be priced as if they represented the same customer opportunity.
One person can generate many machine requests through one task. One cached retrieval can support several later answers. Request growth measures activity; its commercial significance depends on what that activity accomplishes. Cloudflare also sells tools to manage and monetize these interactions, so its diagnosis and its proposed business response deserve separate assessment.
An illustrative calculation shows the difference. Suppose 100,000 displaced human pageviews would have generated $10 per thousand in publisher receipts: $1,000. If 500,000 eligible machine retrievals produce 10% paid uptake at $5 per thousand paid retrievals, receipts are $250. Replacing the $1,000 would require $20 per thousand at the same uptake, before new servicing costs.
These are invented inputs, not observed publisher economics. Raising the price could also reduce paid uptake. The calculation’s purpose is to expose that dependency: a large machine audience does not establish a profitable market at a price the publisher needs.
Explore the revenue-replacement example
Would paid access replace the lost receipts?
Invented monthly inputs in US dollars. Compare receipts before new serving and contract costs.
Enter non-negative numbers, with whole counts for views and retrievals.
Monthly receipts ($)
- Displaced human receipts
- $1,000.00
- Paid access receipts
- $250.00
- Unreplaced receipts
- $750.00
Replacement price at this uptake: $20.00 per 1,000 paid retrievals.
With no paid retrievals, a replacement price cannot be calculated.
At the current access price, 40% paid uptake would replace the displaced receipts.
See the equations
H = D × h ÷ 1,000
Displaced receipts: human views lost (D) times receipts per thousand views (h).
R = L × (p ÷ 100) × z ÷ 1,000
Access receipts: eligible retrievals (L) times paid share in percent (p) times receipts per thousand paid retrievals (z).
Replacement price = 1,000 × H ÷ (L × p ÷ 100)
Defined only when there are paid retrievals.
This is a sensitivity model, not a demand forecast. Changing the price holds assumed uptake fixed; actual buyers may switch sources or stop paying. Receipts above the displaced amount are not profit.
The price and the number of paying requests have to work together. Access controls give a publisher something to negotiate; hard-to-replace information gives a buyer a reason to agree. That still leaves the assistant with a bill. Where could it find the money?
Third answer: let advertisers fund the assistant
One answer is for the assistant provider to acquire the information it needs and fund its own interface with advertising. That can pay for access without promising to reproduce each publisher’s sponsored block.
OpenAI already documents advertiser tools, programmatic campaign management and measurement infrastructure for ChatGPT Ads. Calling this only a future demand-side platform, or DSP, misses a central role. A DSP buys media for advertisers; a provider selling placements inside its own assistant is also the owner of that advertising surface. OpenAI’s Ads documentation
OpenAI says its ads run on separate systems, do not influence answers and are clearly labeled. Its FAQ describes advertising for eligible Free and Go users alongside ad-free paid plans. These are the company’s product commitments. They do not constitute an independent audit of every answer or every sponsored source the assistant might retrieve. Ads in ChatGPT
The Answer Independence Paradox examines how that promised separation could be tested. For the financing question, the significant point is that the assistant can sell a placement of its own while buying information upstream.
There is a legitimate integrity reason to reject paid material that appears to be independent evidence. There is also a commercial advantage in controlling the final advertising surface. Both incentives can exist. Their coexistence does not prove why a particular provider rejects a format; it explains why the terms will be contested.
Alphabet’s October results expose a related tension. Its filed Q3 release reported 12% growth in Search & other revenue and the strongest year-on-year query growth rate in five years. In the available earnings-call transcript, management identified lower monetization in AI Mode alongside currency, sector and comparison effects. More searches and proportionately more advertising income are different achievements. Filed results, call transcript
The reported 12.6% share-price fall on 29 October brought those concerns into sharp focus. It does not tell us how much of the repricing came from AI or prove permanent damage to Search. AP’s market report
These results complicate the apparent solution. Owning the advertising surface does not guarantee that a growing mix of AI queries earns revenue at the previous rate. Both publishers and assistants are trying to finance information consumed through a different interface, with no guarantee that their returns improve together.
User subscriptions can fund information and assistance too, including ad-free services. Merchants can pay for qualified referrals or transactions where those events are verifiable. Google’s Direct Offers describes another direction: promotions within AI shopping experiences, with checkout integration for UCP merchants. Some spending could go into a better price or service offer rather than an ad slot. Google Direct Offers
These payments can coexist in one task, but they are not all new money. A source license is a cost to the assistant and a receipt to the publisher. Some commercial fees replace existing advertising or affiliate spending. A purchase that merely moves between merchants does not create its full value again.
The experiments establish ways to exchange money. They do not yet establish how the competing exchanges fit together, or whether the information supply can survive their terms.
Why would the market build another complicated stack?
Today’s programmatic advertising already has supply-side platforms helping publishers sell inventory, demand-side platforms buying it for advertisers, exchanges and measurement services. Industry standards coordinate parts of the trade. IAB Tech Lab traces OpenRTB to a protocol initiative assembled by supply-side and demand-side participants in 2010; its sellers.json and SupplyChain specifications identify sellers and selling paths. These are forms of private coordination, not a central authority deciding everyone’s price. OpenRTB history, supply-chain specifications
An agent creates more connected decisions. Which information is worth acquiring? Which model or tool should process it? Which commercial offers fit the user’s requirements? Who receives a fee if the task ends in a purchase? The user-side agent, retrieval supplier, model supplier and final presentation surface need not belong to the same company. One company could also integrate them.
A broker has a potential business if finding sources, negotiating rights, combining small payments and verifying use would otherwise cost more than its fee. A broad license may be cheaper to manage than thousands of individual deals. Automation could also make some direct exchanges cheap enough to bypass a broker. These competing effects follow the transaction-cost question in Coase’s account of firms and markets.
Two-sided market theory adds another complication: who pays can matter as much as the total charge. A service can subsidize users to attract advertisers or subsidize access to obtain supply. Adding participants changes those incentives; it does not prescribe a fixed number of intermediary layers. Rochet and Tirole
An intermediary can earn a fee by reducing real costs. It can also retain income because a supplier or customer has difficulty leaving it. A complex stack may contain both. Its existence alone tells us little about whether its participants are paid in proportion to their contribution.
Could today’s bargain underfund tomorrow’s reporting?
Once an investigation exists, providing another copy can be cheap. Producing the next investigation may still be expensive. Information is non-rival: one person’s use does not consume the fact and prevent another person’s use. That makes duplication and continued production different economic problems. Goldfarb and Tucker’s Digital Economics examines the implications for pricing, bundling and incentives to produce.
A publisher might rationally accept a small payment for an existing article while total receipts remain insufficient to fund its next year of reporting. Each assistant might rationally choose an inexpensive substitute while relying on a body of information someone still has to originate. All those individual choices can make sense without sustaining the collective supply.
There is a possible market response. A buyer that expects future information quality to affect its own returns could offer a longer license, commission specialist work or fund a source directly. Publishers could sell exclusive feeds, memberships or services. But a buyer may prefer to let competitors pay for reporting it can later obtain elsewhere. Sources may contract or exit before a replacement bargain emerges.
Calling the distribution fair does not resolve this. Payment proportional to contribution, recovery of production costs, voluntary participation and reasonable consumer prices are different tests. Keeping every historic business model alive is not a definition of fairness either.
Mathematics can help measure contribution. Shapley methods average what an input adds across different combinations of inputs. In machine learning, the result depends on the dataset, algorithm and performance measure, rather than discovering a universal monetary value. Ghorbani and Zou’s Data Shapley
Turning a contribution score directly into a payment introduces another problem: participants may change what they report to improve their return. Fan, Rotello and Karimireddy examine this in Do Data Valuations Make Good Data Prices? Their theory and controlled RAG experiments distinguish contribution measures from incentives for truthful cost reporting. The main theory assumes known buyer valuations; it does not establish a general solution for private, competing web businesses. The paper and its limitations
There is therefore no automatic step from a useful source to a fair price, or from an acceptable price today to a viable information supply tomorrow. Nor does a financially viable bargain guarantee a good answer for the person using it.
But can the customer still trust the answer?
The assistant transforms what it receives. That creates a problem every one of these bargains has to solve: can the customer still tell which claims were paid for?
It may select passages, combine accounts and generate a shorter answer. A sponsorship label can be dropped while the claim beneath it survives. If the answer then cites the publisher, a brand’s paid message can acquire the appearance of independent reporting.
That is a hypothetical failure mode, not a finding that TIME’s campaigns produced it. It is also the practical extension of The Phenomenology of Search: finding relevant language does not establish the reasons to trust it. Several pages repeating a paid claim are not several independent observations.
Ordinary retrieval should also be distinguished from model training. Giving a model text for one answer does not automatically update its parameters or change every user’s future answers. Retrieval-augmented generation explicitly combines a learned model with an external source of passages; the roles are different. Lewis and colleagues’ RAG paper makes that architecture concrete.
The same care is needed with the phrase prompt injection. A clearly labeled factual sponsorship is not automatically an instruction to override an agent’s rules. An embedded command to disregard the user is a more direct attempt to redirect behavior. Between them sits a wider problem: evidence selected or fabricated to distort a decision.
An October preprint, Welcome to the Age of Agent Ads, demonstrates steering through fictitious concepts and harmful associations supplied as search material. Its controlled experiments use listing datasets and mocked search results; they do not measure publisher campaigns or real purchases. The study shows why instruction defenses and evidence integrity are separate concerns. Davidson and colleagues
For the customer’s benefit, we can ask for more than “the facts were brand-approved”: the answer should preserve who supplied a claim, what supports it and what was paid for. This gives us an integrity condition for judging a service. It is not a prediction that every participant will adopt it. An advertiser can fund visibility without becoming an independent authority on its own superiority.
If the money can cross these boundaries while its disclosure disappears, a deal can succeed commercially and still fail the person using it. We can judge an arrangement by whether it funds useful information while keeping selection, sponsorship and outcomes open to inspection. A forecast also has to admit arrangements that persist while failing that test.
What we can expect competition to produce
The cases above establish experiments and stated mechanisms. From the incentives we have examined, we can build a conditional forecast for the next two years. Our base case is concentrated control over consumer-facing advertising and selective payment upstream: general-purpose assistants retain much of the downstream advertising relationship, while source payments concentrate where freshness, distinctiveness or reliability cannot be cheaply replaced. We can expect machine use to spread faster than a broad, dependable income stream for ordinary publishers.
That is a directional judgment from the incentives above, not an estimated market share. Five pressures explain it and the conditions under which it would change.
When information is interchangeable and users stay with one assistant, the assistant can press source payments down. It can buy selected licenses, use adequate alternatives and sell placements on its own surface. More machine consumption need not produce proportionately more publisher income. This is the strongest case for assistant-controlled advertising, but it depends on both replaceable inputs and retained users.
When current, distinctive information materially improves outcomes a buyer can monetize, suppliers gain room to bargain. They could sell feeds, broader licenses, direct services or negotiated sponsorship. Publisher sponsorship might resemble syndication if a source supplies information with an agreed labeled placement. Its scale remains uncertain: the public evidence does not establish reliable delivery, repeat advertiser demand or attractive publisher returns. Being original helps only if the buyer cannot obtain an adequate substitute on better terms.
When users can move their task context and permissions, assistants face stronger competition too. An agent could switch models or sources while preserving the user’s workflow. Some gains could reach users through lower fees or better offers. That does not guarantee higher publisher receipts, because information suppliers may still compete intensely with one another. Portability changes who can refuse which deal.
When coordination is expensive, bundles, brokers and integrated firms have an advantage. If automation lowers the cost of direct contracting and verification enough, some of those layers could disappear. We can be more confident in these opposing incentives than in a forecast that the old SSP and DSP roles will reappear unchanged. Merchant-funded referrals or transactions offer another source of income where performance can be verified and margins support the fee; they do not remove the input-pricing problem.
When production remains expensive but individual sources are easy to replace, the open information supply can deteriorate. More reporting could move into exclusive products or subscriptions; some could disappear. Longer-term buyer funding is a possible response, not an inevitable correction. A market can continue processing a growing stock of existing information while weakening the incentives to produce its next valuable additions.
These pressures can operate inside the same task. A subscribed user could receive licensed information and a separate commercial offer before making a purchase through a paid referral. A rival assistant could choose a different combination. Some sources will refuse the terms, some buyers will bypass them, and some proposed markets will fail. Economically persistent arrangements can also be opaque or damaging to people outside the bargain.
The architecture made these separations foreseeable. It did not predetermine the distribution of returns. Information theory gives no price for trustworthy reporting, and network centrality alone gives no right to collect a fee. Contracts, standards and public rules affect the alternatives people have; decentralized competition does not make those conditions disappear.
There is a slower scenario too: people continue using search, social feeds and rich websites while agents handle portions of the journey. Entertainment, taste, exploration and a direct relationship with an author still give people reasons to visit. A forecast about delegated tasks should leave room for those activities.
What would make the forecast worth acting on?
The useful indicators are paid retrievals at stated prices, publisher receipts after costs, renewed sponsorship contracts, retained disclosures and additional fulfilled purchases. Each answers a different question. A rising request count cannot stand in for all of them.
We would weaken the forecast of platform advantage if portable agents captured substantial repeat use and users could switch without rebuilding their task context. We would strengthen the forecast of supplier leverage if distinctive sources won renewals at prices reflecting demonstrated improvements over substitutes. We can test whether paid machine use sustains original reporting by examining access receipts, production costs and continued investment in new work. Rising paid use alongside shrinking reporting would weaken that financing hypothesis, although we would still need to examine other causes of the decline. If direct deals became cheaper than brokered bundles, the expectation of another large intermediary layer would need revision.
For people building products, one useful experiment would compare a licensed source with its best feasible replacement on a fixed set of tasks. Does it improve correctness, freshness, decision quality or completion enough to matter to the buyer? The test would then need to connect that improvement to a commercial return. A technical quality gain alone does not establish willingness to pay, and a citation does not establish the source’s incremental contribution.
A complementary audit could compare a publisher’s human and machine representations, trace supported claims through repeated test answers and check whether sponsorship remains visible. It would need authorized access and clear limits on which systems it can observe. These are proposed products and experiments, not services whose effectiveness has already been demonstrated.
Commercial reconciliation would connect permitted advertising events to the merchant’s orders, margin, cancellations and returns. Who Owns the Pixel? explains why platform attribution and independent evidence of business value are different records. A successful product would help someone make a better spending decision, rather than add another attractive count.
Publishers have information to supply. Assistants have choices about what to use and present. Advertisers have outcomes they are willing to fund. Their bargains will determine who earns money from the next answer. The crucial question is whether an assistant’s cheapest adequate answer today leaves anyone willing to pay for the reporting its next answer will need.
Sources and scope
This analysis uses public reporting, provider documentation and research inspected on 2 November 2026. Company descriptions establish stated mechanisms and policies, not independently audited commercial results. No private campaign contracts, bids or publisher revenue records were inspected. The Agarwal/Sen finding is described qualitatively from the 17 June working-paper copy. The later revision has not been compared, so we do not reproduce a precise effect size or sample counts. The forecast and interactive models are the author’s analysis, not estimates fitted to market data. The bargaining example illustrates assumed alternatives and an assumed division of gains; it neither calculates fairness nor recovers the full cost of future reporting.