Parallel's founder wants AI firms to pay publishers by how much each page improves answers
Parag Agrawal wants to price each page by what the AI loses without it.
Training Data · Parallel’s Parag Agrawal: Building a New Web for AI AgentsParag Agrawal, the former Twitter chief executive now running the search startup Parallel Web Systems, was explaining what happens when software agents, not people, show up to read the web. A publisher that turns a thousand human visits into 20 subscriptions has no idea what a thousand agent visits are worth, he said. “You can’t monetize that visit.”
His argument is that the old bargain breaks. Ads worked, in his telling, because they subsidized free content like search and social feeds. But an agent does not click ads or sign up for subscriptions, so publishers have reason to block agents entirely and save their pages for humans.
Agrawal said his answer borrows from two past jobs: running ads, and selling Twitter data to OpenAI. He wants what he calls efficient differential pricing, with pay varying both by how good the content is and by how valuable the work is that uses it. A banker doing expensive work with a page would pay more for it than his retired father reading it.
The way to set that price, he said, is to ask how much worse the agent’s answer gets without a given source. Take a page out of the corpus, run the agent, and see how far quality falls. Then ask whether a little more computing power could win that quality back. If a cent of extra computing restores what was lost, he said, “this source is worth close to a cent.”
The formal version of that test, he said, is Shapley values, a game theory idea for splitting a larger pie created by collaboration so everyone still wants to collaborate. Doing the full math might take several dollars to justify paying a publisher one dollar, so Parallel uses estimated values built from test data.
His macro math, he argued, could work: if 2 to 10% of spending on AI for knowledge work went to web data, that would dwarf every web content business today outside what Agrawal called “walled gardens like Facebook and LinkedIn.” He predicted that within 12 to 24 months, as agent traffic grows, that math could send meaningful dollars to a wide range of publishers, at a time when sites like Stack Overflow have seen human traffic plummet.
Questions people ask
What exactly is a Shapley value?
It is a formula from game theory for dividing the gains from teamwork. If three people together create more value than they would alone, it assigns each person the average extra value they add across all possible combinations of teammates.
Why can't publishers just run ads for agents?
Today a site makes money when a person visits, sees ads, or buys a subscription. An AI agent reads the text and carries the answer back to the user, so the visit produces no ad view and no signup. Without a new payment, the publisher pays to host content that earns nothing.
How would this money actually reach publishers?
Agrawal pitches Parallel as the middleman between AI systems and publishers. When an AI company pays Parallel for search and grounding, he says Parallel would use its quality tests to estimate each source's contribution and share part of that payment with content owners.