How AI Agents Could Change the Economics of Online Information

The problem: content use and publisher value can become disconnected

An AI agent can retrieve a publisher's work, combine it with other sources, and answer a user's question. The user may finish the task without opening the source. The information was useful, but the interaction may never reach the event that supports its production: an advertising impression, a subscription, a purchase, or a licensing payment.

That is an economic design problem. A useful agent needs reliable inputs. The people producing those inputs need a workable way to fund the work. The question is how the value of agent-mediated consumption could reach the producer, and what that arrangement would cost to operate.

Separate use, attribution, traffic, and payment

These are different events. A system should not treat one as proof that the others occurred.

Attribution remains useful for traceability and inspection. It does not determine the economic return. A visible citation can produce no visit. A visit can produce no subscription. A licensing receipt can arrive without a visit. Each mechanism needs its own measurement. An agent may also introduce a source to readers who would not otherwise find it. The net effect depends on displaced visits, new visits, and the value of each.

What the available search evidence shows

Pew Research Center analyzed 68,879 Google searches from browsing data shared by 900 U.S. adults. Browsing occurred in March 2025; the corresponding result pages were collected in April. Its July 2025 analysis reported the following patterns. [1] (opens in a new tab)

These are observations about AI-assisted search, not autonomous agents. The analysis is not a randomized estimate of the effect of adding a summary. It also does not measure publisher revenue. Use it as evidence that AI-mediated information interfaces can be associated with different referral behavior. Do not convert the difference into a universal revenue-loss forecast. [1] (opens in a new tab)

Model the revenue channel before choosing a payment mechanism

For a simplified publisher supported by visits, define baseline receipts as visits multiplied by net receipts per visit. An agent-related arrangement adds another channel: counted paid uses multiplied by net receipts per use. It can also add accounting and support costs.

baseline_receipts = baseline_visits * net_receipts_per_visit
new_contribution = retained_visits * net_receipts_per_visit
                 + paid_uses * net_receipts_per_use
                 - additional_operating_cost

This deliberately excludes subscriptions, donations, and other revenue. Common editorial costs are assumed unchanged, so they cancel when comparing the two cases. It is a sensitivity model for one channel, not a valuation of a publishing business.

A worked example with explicit assumptions

Assume the following monthly values in U.S. dollars. Every input is hypothetical. The traffic change is not derived from the Pew study.

• Baseline source visits: 10,000.
• Retained source visits under the hypothetical agent scenario: 6,000.
• Net receipts per source visit: $0.02.
• Counted paid uses: 20,000.
• Additional fixed operating cost: $40. The per-use receipt is already net of variable transaction fees.

from decimal import Decimal

baseline_visits = Decimal("10000")
retained_visits = Decimal("6000")
net_per_visit = Decimal("0.02")
paid_uses = Decimal("20000")
additional_cost = Decimal("40")

baseline = baseline_visits * net_per_visit
retained = retained_visits * net_per_visit
required_per_use = (baseline - retained + additional_cost) / paid_uses

print(f"Baseline visit receipts: ${baseline:.2f}")
print(f"Retained visit receipts: ${retained:.2f}")
print(f"Required net receipt per paid use: ${required_per_use:.3f}")
print(f"Equivalent per 1,000 paid uses: ${required_per_use * 1000:.2f}")

Baseline visit receipts: $200.00
Retained visit receipts: $120.00
Required net receipt per paid use: $0.006
Equivalent per 1,000 paid uses: $6.00

The exact snippet was run locally to check the arithmetic. At these assumed values, $120 in net usage receipts would replace the $80 visit-receipt gap and cover $40 of additional operating cost. That requires $0.006 per paid use. It does not show that any agent provider would pay that amount.

The result is sensitive to the denominator. If only half the uses are billable, the required receipt per paid use doubles. If agent referrals have different commercial value, the retained-visit term changes. If additional costs grow with usage, the model needs a variable cost term. A quoted price is meaningless until these units are defined.

Compare mechanisms against their failure modes

The U.S. Copyright Office's May 2025 pre-publication report on generative AI training discusses why licensing can be more feasible for centrally controlled collections and harder when ownership is fragmented and transaction costs are high. That supports examining the transaction structure. It does not establish a pricing model for autonomous agents. [2] (opens in a new tab)

Micropayments describe payment size. They do not solve resource identification, rights coverage, metering, or settlement costs. Aggregating small charges may reduce payment overhead, but the underlying usage records still need to be correct.

Suggested design: define the paid-use event

I would make the transaction explicit before assuming that licensing revenue can support the content. A proposed agreement or accounting record should answer these questions:

• What resource and version were covered?
• Which action creates a charge: retrieval, reuse, distribution, or another defined event?
• How are failed requests, retries, cached copies, and repeated use counted?
• Who is the payer, and which verified recipient receives the payment?
• Which attribution and usage conditions accompany the transaction?
• What records support reconciliation without storing unnecessary user queries or personal data?

Charging only for network retrieval could reward persistent caching, even when later reuse creates the relevant value. Charging for every downstream appearance could create expensive tracking and privacy problems. The billable event should match the agreed use and a measurement the system can actually support.

Limits of the proposal

Different producers have different objectives. A subscription publisher, an independent technical writer, and a public information service may choose different arrangements. Some may prioritize reach. Others may need direct receipts. There is no basis here for assuming that one mechanism fits every source.

The measured evidence concerns search behavior. The financial example verifies arithmetic under stated assumptions. The proposed arrangements describe possible business mechanisms. None establishes that a particular use legally requires payment or that a quoted rate would sustain a producer. An economically useful design needs a defined event, a reliable count, an appropriate recipient, and costs that leave enough value to distribute.

References

[1] Pew Research Center, Google users are less likely to click on links when an AI summary appears in the results (22 July 2025) (opens in a new tab)

[2] U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI Training, pre-publication version (May 2025) (opens in a new tab)