In 1976, a University of Utah ecologist named Eric Charnov published an eight-page paper in Theoretical Population Biology called “Optimal foraging, the marginal value theorem.” It solved a specific problem: when should a foraging predator leave a patch of diminishing prey? The paper required calculus and an optimization condition, and it has been cited more than 7,000 times as of June 2026, per Google Scholar. Twenty-three years later, two researchers at Xerox PARC cited it in a paper about web navigation.
Peter Pirolli and Stuart Card were not borrowing the metaphor. When their 1999 Psychological Review paper applied Charnov’s theorem to the problem of when a website visitor should navigate away from a page, they imported the mathematical structure — the same optimization condition, with variables renamed. “Information foraging theory” is what they called the result. This piece reads both formalisms in parallel to show what the translation kept, what it adapted, and what it cost.
The hawk and the patch
The problem Charnov formalized had been stated a decade earlier. MacArthur and Pianka’s 1966 paper in The American Naturalist addressed a related question: which prey types a predator should include in its active diet. [3] Their treatment was largely graphical. The core decision rule, as described in secondary accounts: add a prey type to the diet when its profitability — energy gained per unit of handling time — exceeds the expected return rate from the currently optimal prey set. [6]
The diet breadth model’s structure is a ranking and a cutoff. Rank prey types by profitability. Begin with the most profitable. Add the next type when its profitability exceeds the average rate of return currently achievable, given encounter rates with everything already in the diet. [6] The rule is indifferent to any particular prey item’s absolute value; what matters is whether including it raises or lowers the overall return rate. Predators are not loyal to prey types. They are optimizing a rate.
Charnov’s theorem addressed a different dimension of the same problem: not which prey to pursue within a patch, but when to leave the patch entirely. A patch is a location where prey concentrates but depletes — a tree, a field, a stretch of riverbank. Returns diminish as prey is consumed. Moving between patches takes time and costs energy. The theorem asks: given a patch that is running low, when does it become worth stopping and traveling to a new one?
The answer is precise: leave the current patch when the instantaneous rate of energy gain there drops to the average rate of energy gain across the full habitat, including travel time between patches. [8] This is the marginal value condition. Not “when the patch is empty” and not “when the patch falls below some fixed threshold,” but when the marginal return from staying equals the average return from the full foraging strategy.
The graphical representation makes the structure visible. Draw the cumulative energy gain curve for a patch over time — steep at first as prey is abundant, flattening as it depletes. On the time axis, mark the travel time to the left of origin. Draw a line from that point tangent to the gain curve. The slope of that tangent is the maximum achievable average rate; where the tangent touches the curve is the optimal departure time. [8] Earlier departure wastes yield; later departure accepts below-average returns. The theorem identifies the exact point.
Two properties of the MVT formalism are relevant to the translation that came after.
First, departure timing is relative to habitat quality, not absolute patch quality. If the overall habitat is poor, the threshold is low and the predator stays longer in depleted patches, because the alternative is worse. If the habitat is rich, the threshold is high and the predator leaves sooner. Optimal departure time is not a fixed rule about the patch; it is a comparison between the patch and everything else.
Second, the currency is energy — measurable in calories, conserved across interactions, objective. Two ecologists can in principle agree on the energy content of a field mouse. This property underwrites the theorem’s empirical testability: you can measure the relevant quantities independently of the predator’s behavior, derive the prediction, and check whether departure times match.
The import
The translation Pirolli and Card performed in 1999 was systematic. [1] The primary paper was not directly accessible; the following account draws from secondary descriptions. [7]
The structural substitutions:
The patch becomes a webpage or website. The habitat becomes the broader information environment — the web, a database, a document collection. Energy becomes information value: the relevance of content to the user’s current goal. Different prey types become different links or information sources available from a given location. Handling time — the time to catch and process a prey item — becomes the time to read and process information within a source. Travel time between patches becomes the navigation time between web pages or sites.
Both decision rules transfer through these substitutions.
The diet breadth rule, translated: follow a link when its expected information value per unit of processing time exceeds the current average rate of information gain from available alternatives. The patch departure rule, translated: leave a webpage when the instantaneous rate of gaining relevant information falls below the average rate available across the broader information environment, accounting for the cost of navigating to a new source — the direct application of Charnov’s departure condition to information environments, following from [1].
These are not analogies in the loose sense — a structural resemblance noted for illustration. They are the same optimization conditions, stated in different variable names.
The key new concept Pirolli and Card introduced to make this work is “information scent.” [1, 7] The problem is that information value is not directly observable before a source is visited. Predators in ecology similarly cannot know the energy content of a patch before arriving — they estimate from proximal cues: the shape of a burrow entrance, the density of visible prey, the presence of competitors. Information scent is the name for the analogous proximal cues in information environments: the text of a link, its surrounding context, the snippet visible in a search result, the page’s headline and above-the-fold content. Users estimate expected information value from scent the way foragers estimate patch richness from observable indicators. [7]
This concept does more than name an existing phenomenon. It provides a theoretical account of why information architecture decisions matter: good interface design makes scent accurate, so that users can efficiently estimate information value before committing navigation time. Poor design corrupts scent, so that users follow low-value links and fail to follow high-value ones — a departure from optimal foraging behavior that can be predicted and measured.
What the substitution costs
Energy and information value are not equivalent as theoretical currencies, and tracking the difference matters for understanding what the formalism can and cannot do.
Energy is measurable. A calorie is a calorie; the energy content of a prey item can be determined independently of the predator’s preferences or state. Information value is not measurable in this sense. The same webpage contains different information value for a user researching tax law and a user researching bread recipes. Information value is goal-relative and subjective in a way that calories are not.
Energy is conserved. A prey item consumed by one hawk is gone from the environment. The hawk’s gain is real depletion. Information is not depleted by consumption — reading a page does not remove information from it. The “patch depletion” dynamic in information foraging theory is something different: it is the rate at which a specific user, on a specific visit, continues to encounter novel, relevant information on a page. The user’s marginal gain diminishes; the page itself is unchanged. This is a functional equivalent of depletion, but it is not the same mechanism.
Energy is cardinal. The energy from a field mouse and a vole can be added to get a total intake. Information value from two sources, even if both are relevant, does not add in the same way. What is the sum of two pieces of relevant information? The question does not have a clean answer.
The approach to these limitations follows the same pattern economists use for utility: treat information value as a theoretical construct and use observable proxies in experimental settings. Information scent, already introduced as a design concept, serves as the observable proxy for expected information value. Relevance ratings from experimental participants operationalize realized information value. [7] This is a standard move in behavioral modeling — it is exactly how utility theory in economics works, and it is defensible for the same reasons. But it changes what the formalism can be tested against.
In Charnov’s ecology, verifying the MVT requires measuring energy content, tracking depletion curves, measuring travel times, and checking whether observed departure times match the prediction. [8] The predictions are exact. In information foraging theory, information value is not independently measurable. What can be tested is whether a computational model implementing the IFT equations produces navigation patterns that resemble human behavior.
That is what SNIF-ACT, published by Fu and Pirolli in 2007, does. [5] SNIF-ACT implements the information foraging equations inside ACT-R, a cognitive architecture, and simulates web navigation. The model selects links according to information scent estimates and departs pages according to a translated MVT-derived threshold. Against data from navigation studies, the model produces paths resembling those of actual users.
This is useful evidence — a formalism that could not produce plausible navigation behavior would be a genuine problem. But matching behavior with a model is not the same as confirming that users implement the underlying optimization. A user applying a simple, ruder heuristic — “keep reading until the page feels unhelpful, then leave” — might produce superficially similar navigation patterns without performing anything like the MVT calculation. The surface pattern does not distinguish between the formal optimization and a fast-and-frugal approximation of it.
The MVT makes at least one non-obvious, testable prediction that could cut between these possibilities. The patch departure rule says that optimal departure time depends on travel cost. If navigating to a new site is expensive — slow page loads, many clicks required, unfamiliar domain — the average rate of gain from the habitat drops, and the optimal threshold for leaving any given page drops with it: a user following the MVT-derived rule should stay longer in the current patch when traveling between patches is harder, holding page content constant. A user applying a context-free “I’m done with this page” heuristic would not show this effect. Whether empirical navigation data confirms this prediction remains open.
The broader scope of the theory raises a related question. Information foraging theory has extended beyond web navigation to library search and database retrieval. In some of these applications, the information scent concept is operative while the MVT-derived departure equations are not — “foraging” has become the framework and the formalism has receded. This is the normal lifecycle of a formal import: the equations legitimate the framework; the framework generates vocabulary; the vocabulary travels further than the equations.
The honest answer is that it accomplished two things: it provided a mathematically grounded account of information-seeking behavior, and it legitimated interface design research as a domain where formal predictions are possible. The first thing is real but not directly testable in the way the ecological original is. The second thing is real and has produced research programs.
The hawk leaves the patch when the catch runs thin. The website visitor closes the tab when the page stops yielding. The Marginal Value Theorem makes both into the same formal problem, and the formal translation holds — the mathematical structure that governs one governs the other, given the variable substitutions.
That is what the import accomplished: a principled optimization account of information-seeking, formally grounded, capable of generating testable predictions at the level of patterns. Less than the ecology original offers, because direct measurability did not transfer. More than a loose metaphor would produce, because no predictions come from those. The formal apparatus was imported. The measurability was not. Both facts are true, and the theory holds within the limits they define.