The open web is a miracle.
Anyone can publish, learn, and collaborate. It is the closest thing to humanity's living memory, and the open commons that fuels what AI can do on our behalf.
The web was built for human eyes, human judgment, and human attention. Every page, every link, every search result, every business model follows from that premise.
Now the web's primary user is changing. AIs are becoming the web's second user, and they will use it far more than humans ever have. They can reason across sources we could never read in a lifetime, connect ideas across fields we would never think to bridge, and help us refine knowledge at a pace we are only beginning to understand. But that potential needs the right infrastructure and the right incentives. Everything from the way information is discovered to the business models that fund its creation must evolve.
Parallel was built for this transition. Index is where the mission becomes concrete: a market that tracks which sources shaped an answer, measures how much they contributed, and sends value back to them.
I. The Web We Inherited
The internet that shaped us was a marvel of openness, but its basic assumptions were tuned for a reader at a screen, not an agent reasoning over the full living memory of civilization.
The web's original gift was openness. Its original blind spot was that attribution, trust, and payment would be handled by the human at the edge.
The web did not begin as a media empire, an ad network, or a social substrate. It began as a coordination problem. At CERN, people were rotating through experiments, documents were scattered across incompatible systems, and institutional knowledge was evaporating faster than anyone could map it. Tim Berners-Lee's original proposal was modest in tone and radical in implication: if information could be linked across machines and addressed consistently, an institution could become legible to itself.
That choice of minimalism mattered. The web shipped with universal addresses, transfer rules, and a simple markup language. It did not ship with identity, native pricing, or built-in provenance. It assumed that readers would supply judgment and context at the moment of use. We built the web for eyes, because a person, not a model, was expected to click, compare, trust, doubt, and decide.
That is why the early web could stay so permissive. If something was misleading, a human could leave. If something was useful, a human could link to it. If something was worth money, a human could subscribe, buy, or tolerate the ad next to it. The protocols remained thin because the reader supplied the missing machinery. Trust lived in behavior. Payment lived outside the page. Attribution lived in footnotes, habits, and professional norms.
There is real beauty in that inheritance. It gave us a network where permissionless publication could outrun institutional gatekeeping. It produced personal pages, forums, wikis, journals, fan archives, independent criticism, and the vast uneven public record we now call the open web. But it also left a foundational omission untouched. The web could point to a source, yet it could not natively price the contribution that source made downstream.
II. Why Search Worked
Search engines succeeded not merely because they indexed pages, but because they learned to read the traces of human judgment embedded in links, clicks, and time.
The next leap was not publishing. It was retrieval. Once the web became large enough, abundance threatened to turn into noise. Search worked because it discovered that the web already contained a distributed record of what people found worth referencing. PageRank read links as signals of importance, and then the surrounding ecosystem learned to interpret clicks, dwell time, and linking structure as a map of relevance.
That map was never neutral, but it was alive. Every citation, every backlink, every moment of attention created a feedback loop between creators and readers. Search did not manufacture truth on its own. It inferred which documents were probably worth the next person's time based on how previous people had behaved around them. The machine could rank because people had judged first.
From there the web built an economy around the same assumption. If attention was scarce, it could be monetized. Ads, sponsorships, subscriptions, and recommendation systems all became different methods of converting human presence into revenue. Even when the business model changed, the user model stayed fixed. The page existed to be seen. The source existed to win a visit. The loop closed when the audience came back.
This is the history behind our present confusion. We often speak as though search, publishing, and monetization were separate layers. In practice they were one tightly coupled arrangement. Retrieval worked because the web was browsable. Monetization worked because retrieval returned traffic. Attribution stayed socially meaningful because a human could click through to the originating page and see the source in context.
III. The Bargain Is Breaking
AI changes the shape of demand. The web is still the substrate, but the dominant user no longer pays with attention.
The bargain was simple: let me crawl your pages and I will return readers. That bargain weakens the moment the reader is no longer a person with browsable time.
The tension we feel now is not abstract. It is architectural. A crawler used to be a prelude to a visit. Increasingly, a crawler is the visit. The model ingests, summarizes, and answers without sending the human back through the source. robots.txt remains a coordination rule rather than an economic one, and so publishers are using the tools they have: blocklists, paywalls, bot policies, and negotiated access.
This is why the old traffic bargain no longer clears the market. A model can extract value from a document without generating a pageview, an impression, or a subscription event. The source still does the work. The source no longer reliably receives the compensating flow.
You can watch the internet inventing defensive behavior in real time. Cloudflare now markets tools to identify, block, and even monetize AI crawler access. That is a rational response, but it is also evidence that the protocols we inherited are insufficient for the users we have now. Seventy-nine percent of major news sites block at least one AI crawler. The licensing deals that exist are flat fees, disconnected from actual usage, available only to publishers with the scale to get a meeting.
The bargain was traffic. But resolution has no native clearing system on the web today. That absence is what this prototype is trying to make visible.
IV. Retrieval Is the Binding Constraint
Models can reason, but reason without reliable retrieval is an elegant machine operating on stale or unearned memory.
The mythology of AI often centers on model capability, as though intelligence were primarily a matter of bigger weights and better inference. But most meaningful work does not begin with pure generation. It begins with contact: the current law, the latest filing, the original paper, the conflicting source, the buried edge case, the record that sits just outside the model's parametric memory.
That is why retrieval is not a utility feature. It is the binding constraint on whether machine reasoning can remain grounded in the living world. The web is still the most expansive continuously updated public memory humans have built, and when an agent cannot reach that memory with enough precision, every downstream claim gets weaker.
There is a long intellectual lineage here. Vannevar Bush imagined a memex because abundance without navigability was already a problem in 1945. Ted Nelson imagined a linked publishing system with transclusion and micropayments because he saw that references should not only connect documents but preserve lineage. The dream of retrieval-augmented thought is older than the modern web. We are not inventing the desire. We are finally colliding with the scale that makes it unavoidable.
A useful agent is therefore not just a model. It is a reasoning surface connected to memory, tools, and verifiable source material. Hutchins gives us a language for this: cognition distributed across people, artifacts, and environments. The web plus retrieval plus agent interfaces is a distributed cognitive system. If the substrate is open, current, and attributable, the system compounds knowledge. If the substrate fragments, the system compounds confidence without grounding.
V. What a Price Signal for Knowledge Could Do
If attention was the old signal, contribution has to be the new one: a measurable record of which sources actually helped resolve which questions.
A price is a compression of context. Attribution can play the same role for knowledge if it can trace not just citation but contribution.
F. A. Hayek described prices as a system for transmitting dispersed knowledge. No buyer sees the whole field; no seller needs to. The price carries enough of the system's state forward for coordination to happen. That is the analogy worth stealing for the next web. Prices compress local knowledge into shared action. Attribution, if done well, can compress intellectual contribution into a usable public signal.
Imagine a network in which every serious answer emitted not just a confidence score, but a contribution map: these sources carried the synthesis, these passages supplied the crucial edge cases, this dataset anchored the answer, this archival page established lineage, this publisher earned the largest share because their reporting changed the conclusion.
That contribution map could do what raw traffic never could. It could tell creators which topics are undersupplied, which documents remain indispensable long after publication, and which work is structurally load-bearing inside new machine-mediated workflows. Prices would stop tracking eyeballs and start tracking usefulness.
History suggests that this matters. Dittmar's work on the printing press argues that communication infrastructure changed the rate of urban and economic growth. The point is not that every protocol innovation yields instant prosperity. The point is that when a society learns to disseminate and recombine knowledge more efficiently, entirely new kinds of coordination become possible.
VI. Why Index by Parallel
The next web needs more than crawl permissions and closed deals. It needs a public-facing interface for provenance, contribution, and payout.
Index by Parallel starts from a straightforward conviction: if agents are going to read on behalf of people, then the web needs infrastructure that makes those readings legible to the sources that made them possible. The missing layer is not only search quality. It is public accounting. Who contributed? How much? To which outcome? With what confidence? And what value should flow back?
We already know that modern research systems can emit detailed basis objects, citations, and confidence metadata. Parallel's own deep research stack demonstrates that provenance can be structured, machine-readable, and useful in the output itself. The question is what happens when that trace becomes a user-facing product primitive rather than an implementation detail hidden in logs.
This page is a proposal disguised as a reading experience. Hover a sentence and you see a miniature ledger. Expand a phrase and the document reveals another layer of reasoning, along with the sources that made that elaboration possible. The essay behaves like software because writing itself is becoming programmable. But programmability without attribution would only accelerate extraction.
Attribution is the clearing layer.
VII. A Document That Knows It Came From Somewhere
The future of publishing may not be a static page or a chat answer, but a responsive document that can reveal its lineage as it unfolds.
For most of the web's history, a document was assumed to end at its visible boundary. You read the paragraph, maybe opened the footnote, maybe clicked away. The new interface possibilities are stranger and more promising. A document can now expand where your curiosity concentrates. It can surface sources at the sentence level. It can expose confidence and ambiguity instead of pretending every paragraph arrived from nowhere with equal certainty.
That is why telescoping text matters here. It is not a novelty flourish. It is a clue about how machine-written and machine-augmented documents may behave in public. A sentence can be the compact form of a deeper argument. A reader can ask for more without leaving the page. A source can remain visible even as the wording becomes adaptive.
If we do this badly, the web turns into a smooth surface of answers detached from origin, context, and compensation. If we do it well, every good synthesis becomes a new reason to sustain the reporting, scholarship, documentation, and public writing beneath it. The document becomes an interface to the commons rather than an extraction layer on top of it.
That is the invitation behind Index by Parallel. Not a plea to preserve the old web exactly as it was, because what got us here will not take us there.
The invitation is to build the next retrieval layer so that intelligence can compound without severing itself from the people and institutions that make knowledge worth retrieving in the first place.
Built from the Demand Side
Parallel was founded in 2023 by Parag Agrawal with the mission to keep the web open, competitive, and transparent.
Backed by Sequoia Capital • Kleiner Perkins • Khosla Ventures • Index Ventures • First Round • Spark Capital • Abstract • Terrain
About Parallel