The Effective Download: A New Metric for AI Consumption

For two decades, COUNTER has been the common language of academic publishing analytics. Investigations and requests gave publishers and librarians a shared vocabulary for usage — one that drives purchasing decisions worth billions of dollars a year.

Then AI agents arrived. And the numbers stopped meaning what they used to mean.

A quick refresher on COUNTER

COUNTER tracks two item-level signals:

• Investigations — any interaction with an item or its metadata. Viewing an abstract counts. Viewing the full text counts.
• Requests — accessing the content itself. A PDF download. A full-text view. The user has decided this item is worth their attention.

Every request is also an investigation. The ratio tells a story: a journal with high investigations and low requests is being browsed but not used. A journal where most investigations convert to requests is essential.

This framework assumed something quietly fundamental: a person, sitting at a computer, deciding whether or not to read an article.

What AI breaks

A human reads one item at a time. An agent might pull twenty in a single query. A human downloads a PDF and reads it whole. An agent ingests a few thousand tokens — three paragraphs from the methods section, the conclusion, two figure captions — and moves on.

A single AI-mediated literature review might generate hundreds of "investigations" with almost no "requests" — or, depending on how the platform classifies API calls, hundreds of requests where a human would have made one. The ratio that used to mean something now means whatever the plumbing happens to mean.

A new unit: the Effective Download

We propose a unit that maps to how AI agents actually behave: the Effective Download.

Effective Downloads = Total tokens consumed from an item / Total tokens in the item

The "item" is whatever the publisher counts as a unit of content — an article, a chapter, a book, a journal issue, a database record. If an item contains 10,000 tokens and an agent reads 5,000 of them, that's 0.5 Effective Downloads. If the same item is referenced again and again across many queries and accumulates 100,000 tokens of consumption, that's 10 Effective Downloads.

Why tokens, not chunks? Chunks are a retrieval-time choice — they vary in size, overlap, and segmentation strategy across platforms. Two systems serving the same article might produce four chunks or forty, and a chunk from one system isn't comparable to a chunk from another. Tokens are model-native and deterministic: every model consumes content in tokens, and the token count of a given item is fixed. They're the one measurement that survives translation between systems.

The unit works in both directions, and that's the point.

Sub-unit consumption — under one full read. An agent pulls three paragraphs from the methods section. A citation references a single figure. A literature review consults only the conclusion. None of these would register as a meaningful "request" today, but together they describe how AI actually engages with content. The Effective Download makes the partial read visible.

Over-unit consumption — beyond one full read. The same item gets referenced again and again. A researcher returning to a chapter ten times shows up today as one download. Under AI workloads — where a single item can be reread across hundreds of queries — the gap only widens. The content that's load-bearing in someone's research looks identical, in current metrics, to content that was opened once and abandoned. The Effective Download separates them.

The Effective Download isn't trying to replace COUNTER. It's trying to give COUNTER something to count.

Mapping back to investigations and requests

Two events bridge the Effective Download into existing librarian workflows — and any infrastructure-level publisher can observe both.

Content delivered to the agent → Investigation-equivalent. When an agent pulls tokens from an item, whether or not it ends up using them, the retrieval is observable. Counted in tokens, it composes cleanly into Effective Downloads.

Sources cited in the response → Request-equivalent. When the agent surfaces a citation to the user — and especially when the user clicks through — the item has crossed the same threshold a request used to mark. The modern version of "they came back for the full text."

Together these reconstruct the investigations-to-requests ratio in a form that means something again. Heavy delivery, few citations: the agent is browsing your corpus but not relying on it. Heavy delivery, many citations: your content is load-bearing in someone's workflow.

Why this only works at the infrastructure layer

If you depend on the AI application to report usage, you'll wait a long time, and what comes back will be inconsistent across vendors. None of them owe you a faithful account of how your content was used inside their product.

But if the publisher operates the MCP server through which agents access the content, the measurement happens upstream of the model. Every retrieval, every token, every citation event is observable directly — across any AI agent, any MCP client, any product built on top. The agent doesn't have to opt in. It already sent the request.

In the human era, publishers measured clicks because clicks were what reached them. In the AI era, tokens delivered and citations served are what reach the publisher's infrastructure — and those are exactly the right things to count.

The proposal

The Effective Download is what we believe usage should be counted in from here forward. Three things are already settled:

1. Agents consume in tokens, not whole items. The unit has to match the behavior.
2. The events that matter are observable at the infrastructure layer. There's no need to wait for application vendors to define a reporting standard they have no incentive to define.
3. Investigations and requests are still the right two questions. What was looked at, and what was actually used? The Effective Download restates those questions in a unit that survives contact with AI.

The next chapter of COUNTER will be written by the publishers who can already produce this data. The metric is ready. The infrastructure is ready. The conversation is the part that's still beginning.