Data vs Information vs Knowledge: Key Differences

A crawler asks for one product page and gets about 2.9 MB back. In the HTTP Archive's July 2025 crawl, the median desktop page weighed 2,862 KB, and the HTML document inside that was 22 KB of it. The thing you actually wanted, a stock number, a price, the currency, the minute you saw it and the address you saw it at, is about two hundred bytes. Everything between 2.9 MB and 200 bytes is refinement work.

Three words describe three states that material passes through on the way, and blurring them is how the work goes missing from the plan. That is the project with a dashboard nobody acts on, and the project with a terabyte of stored HTML that never became a decision. What follows is the distinction, the pyramid that formalizes it, the parts of it that do not survive contact with real data, and what the ladder costs when you turn web data into something you can act on. Every figure, definition and quotation below was read from its primary source on 10 August 2026.

Data: the raw material

Data is a recorded fact with no context attached. A symbol, a number, a measurement, a run of characters. On its own it tells you nothing, because you do not yet know what it represents.

• The string $1,299 is data.
• The characters a p p l e are data.
• The pair 48.85, 2.29 is data, and it shows how little that is. Two decimal places of latitude is 1.11 km on the ground. Two decimal places of longitude at that latitude is about 0.73 km. The pair names a box of roughly 0.8 km². It says nothing about which datum, which hemisphere, or whether latitude comes first.

Is $1,299 a price, a salary, a credit limit? Is apple a fruit, a company, or half of somebody's password? Nothing in the token answers that. In scraping terms, the response body your HTTP client just received is data in its purest form: complete, present and mute.

Raw is a decision, not a property. The 2013 MIT Press collection edited by Lisa Gitelman is called "Raw Data" Is an Oxymoron for exactly this reason, and the point lands hardest on people who write extractors. Somebody chose to request that URL and not another. Somebody kept the price node and dropped the shipping estimate two lines below it. Somebody set a timeout, which decides what a slow page counts as. By the time bytes reach disk, several judgments are baked in, and none are recorded in the bytes.

Data is what is left after you have made several decisions and stopped writing them down.

Information: data with context and structure

Information is data with enough context to answer what, who, where and when. The context is not a mood, and you can check item by item whether it is present:

• Schema. Which field is this, and what type is it? A bare column of numbers is not a price column until something says so.
• Units. Google's merchant listing documentation requires priceCurrency alongside price, in three-letter ISO 4217 format, and requires a price greater than zero. The largest consumer of structured data on the web refuses a number without its unit. That tells you how often the number turns up without one.
• Identity. Which object does this describe? Not the model, the variant: capacity, color, seller, region.
• Time. When was this true? A price with no timestamp is a rumour with a decimal point.
• Provenance. Which crawl, which address, which renderer, which parser version. The W3C's Data on the Web Best Practices, a Recommendation since 31 January 2017, puts this at best practice 5: "Provide complete information about the origins of the data and any changes you have made."

Two corrections are owed here, both against the earlier version of this article.

A label that names the wrong thing is not context. The earlier version used "iPhone 16 Pro, list price $1,299" as its example of data turning into information. Apple's own September 2024 launch announcement put iPhone 16 Pro at $999 to start, in four capacities from 128 GB to 1 TB. So $1,299 is not that model's list price. It might be a larger capacity, a different model or a reseller's markup, and the label does not say which. That is precisely the failure the example was written to illustrate. A record that looks labeled and is under-labeled is worse than an unlabeled one, because nothing downstream will question it.

Information theory does not measure meaning, and citing it here was sloppy. Shannon's 1948 paper in the Bell System Technical Journal says so in its second paragraph: "These semantic aspects of communication are irrelevant to the engineering problem." Entropy measures how much uncertainty a message removes about which message was sent, out of a known set. It has nothing to say about whether $1,299 is a price. The everyday sense of "gaining information" and the mathematical one share a word and not much else.

Parsing is what produces this layer. Feed raw markup to an extractor, get back a record, and the record answers questions the bytes could not. Cleaning and normalizing that output is the step that makes the answers trustworthy rather than merely present. Several standards exist to carry the context along with the values instead of in somebody's head: schema.org, whose vocabulary reached version 30.0 on 19 March 2026; the W3C's Model for Tabular Data and Metadata on the Web, a Recommendation since 17 December 2015, which exists because a CSV file alone is not information; and Data Package, now at version 2, which bundles files with the schema that explains them. JSON Schema is the odd one out. Its current published version is still 2020-12, unchanged since December 2020, with the next release unshipped.

Some of that context is already sitting in the pages you scrape. In the Web Almanac's October 2022 crawl, 37% of pages carried JSON-LD and roughly a quarter carried microdata. Reading a page's own structured data beats parsing its rendered text, and it is the first thing to check before writing a selector.

Knowledge: information applied and joined

Knowledge is information understood, connected to other information and put to work. It answers how and why, and it supports a decision.

One labeled price is information. Ten thousand labeled prices, tracked across competitors and across weeks, become the finding that a rival discounts this category every Tuesday and undercuts you by 8% on average when they do. One location is information. A map of where your customers cluster is something you can site a shop on.

The step that does the work is joining, and joining is where the effort hides. The same laptop sold by forty retailers is one product only because somebody wrote a rule saying which strings mean the same object, then handled the cases where the rule is wrong. That job has a name, entity resolution, and it is the difference between forty price series and one. Get it wrong in the generous direction and you compare a 512 GB model against a 256 GB one. Get it wrong in the strict direction and your competitor appears to have stopped selling the product.

Machine-readable knowledge has standards of its own, and they are moving. RDF, the model behind knowledge graphs, is being revised: RDF 1.2 Concepts reached Candidate Recommendation Snapshot on 7 April 2026. A graph earns its cost when your questions stop being about rows and start being about relationships between things you named.

The old illustration still holds. A sentence in a language you do not read carries information, since it clearly means something to somebody. Until it is translated it yields no knowledge at all. Data mining is the systematic version of that translation: the climb from a pile of records to a statement you can act on.

The pyramid, and who actually built it

These layers are usually drawn as the DIKW hierarchy, with each tier built from the one below.

That table has five rows, and almost every version you will meet online has four. The four-tier drawing is usually credited to Russell Ackoff, and Ackoff's own version had five. His paper "From Data to Wisdom", presented in 1988 and published in the Journal of Applied Systems Analysis in 1989, puts understanding between knowledge and wisdom, and the memorable line depends on it: "An ounce of information is worth a pound of data. An ounce of knowledge is worth a pound of information. An ounce of understanding is worth a pound of knowledge."

The credit is also older than Ackoff. T.S. Eliot's 1934 pageant play The Rock asks where the wisdom lost in knowledge went, and where the knowledge lost in information went. Nicholas Henry drew a working distinction between data, information and knowledge in 1974, in the same paper that used the phrase knowledge management. Milan Zeleny mapped the tiers onto know-nothing, know-what, know-how and know-why in 1987. Robert Lucky's Silicon Dreams, also 1989, is where the layers get drawn as a pyramid. Ackoff gave the sequence its canonical statement rather than inventing it, and the tier the retellings drop is the one that answers why.

Where the ladder stops working

Take the pyramid as a checklist of remaining work and it earns its keep. Take it as a theory of how understanding forms and it falls apart, which the information-science literature has been saying for twenty years.

The logical objection. Martin Frické's 2009 paper in the Journal of Information Science, volume 35, argues the hierarchy contains a central logical flaw and rests on operationalism and inductivism, positions the philosophy of science had already left behind. Data does not arrive prior to theory and then get refined into it. You have to know roughly what you are looking for before you can decide what counts as a record, which makes the bottom rung depend on the top one.

The discontinuity objection. David Weinberger put the practical form of it in Harvard Business Review in 2010. Data and information are things a computer holds. Knowledge and wisdom are things a person holds. The pyramid draws a smooth ascent across a break, and a smooth ascent is exactly what software cannot deliver you.

The definitional objection. Jennifer Rowley's 2007 review of textbook treatments found plenty of pictures and little discussion, with wisdom barely defined. Chaim Zins, surveying information-science scholars for JASIST the same year, came back with several incompatible models and no agreement on whether E = mc² is information or knowledge.

The law runs the ladder backwards. Article 4(1) of the GDPR defines personal data as "any information relating to an identified or identifiable natural person". Information is the genus and data is the species, the reverse of the pyramid. This is not a quibble about words. Reason that you hold only data, not information about anybody, and the regulation does not follow you there. A row relating to an identifiable person is personal data the moment it lands on disk, before any parsing or joining.

The ladder is a good map of what work remains. It is a poor account of what anybody knows.

What changes between one page and ten thousand a night

A one-off script and a nightly crawl are different problems, and the difference shows up first as arithmetic.

Ten thousand pages at the median 2,862 KB is roughly 28 GB a night, moved and paid for, to keep perhaps 2 MB of records. Fetch the HTML document alone, at the median 22 KB, and the same run moves about 220 MB. Deciding which layer you need is a two-orders-of-magnitude decision about bandwidth, and it is usually made by accident, by whoever picked a headless browser because it was easier to start with.

Schema drift is silent. A retailer moves the price into a different node and your parser keeps returning records. They are the wrong records. Nothing raises an exception, because a string was found where a string was expected. The failure surfaces weeks later as a trend nobody can explain. The defense is to assert what a valid record looks like and fail the run when the assertion breaks.

Time turns information into a series. One price is a fact about a page. A price with a timestamp, repeated, is a fact about a market. Store observation time and collection time separately, because they diverge on retries and the difference is what tells you whether Tuesday's discount was real or your scheduler was late.

Provenance stops being paperwork. At one page you remember where the number came from. At ten thousand a night across a proxy pool, you cannot re-derive an analysis without knowing which crawl, which parser build and which render produced each row. Lineage tooling exists to carry that, OpenLineage among others, now at version 1.52.0.

Cleaning does not scale by adding people. Interactive tools are the right instrument at a thousand rows and the wrong one at ten million. OpenRefine, the successor to Google Refine and still the best answer for hand-cleaning a messy extract, last shipped 3.10.1 on 4 March 2025. Past a certain size, cleaning has to become code that runs on every batch.

One missing unit, found late, costs a full re-crawl: 28 GB and a day of staleness for one field.

The same ladder with a storage bill: bronze, silver, gold

Data engineering rebuilt the pyramid in the last decade and gave it different names. In the medallion architecture, as Databricks documents it, the bronze layer holds "raw, unvalidated data" in the source's original format, the silver layer holds "validated, cleaned, and enriched versions of the data" after deduplication and normalization, and the gold layer holds "aggregated data tailored for analytics and reporting".

Bronze is data, silver is information, gold is the bottom of knowledge. The renaming adds two things the original never had. Each layer is materialized, so it carries a cost you can read off an invoice. Each layer is kept, so when the silver logic turns out to be wrong you re-derive it from bronze instead of re-crawling the web. DIKW draws its arrows upward and never says to keep the bottom rung. Anyone who deleted raw responses to save disk, then found a parser bug, has learned that at full price.

Name the layer before you ask for it

"Get me the data" and "get me the answer" are requests of a different order. Quoting them as the same thing is where projects go wrong before anyone writes code.

Collection is not the finish line. A crawl that returns pages has produced data. Without parsing, validation and joining it never becomes information, and it never becomes a decision. Budget the whole ladder or expect to fund the rest of it later, in a hurry.

Value climbs with the rungs. Raw pages are cheap and abundant. An answer that changes a price list is neither. The figure everybody quotes, that data teams spend eighty percent of their time cleaning, traces back to self-reported vendor surveys rather than to a measurement of anyone's calendar. Treat it as a mood rather than a number, and measure your own.

Buy at the rung you need. If what you want is analysis-ready records rather than stored markup, that is what a data as a service arrangement delivers, and what a managed web scraping service is for on targets that fight back. If what you want is the answer to a business question, say that, because the distance between the two is most of the work.

Four questions cover the common case. Which rung am I holding? Which rung does the decision need? What has to be attached to climb the next one, and who attaches it? What breaks silently if nobody does?

Data, information, knowledge. Three words, three separate jobs. Say which one you mean.