The comprehension tax
Every reading product quotes words per minute, a figure that is trivially easy to inflate and means nothing on its own. Here is how to benchmark a speed reader honestly, including this one, and what you will probably find.
What this piece argues
- Words per minute alone is meaningless; report rate and recall together, with the test.
- Most speed-reading measurements fail on the passage, the test, or the missing control.
- Test after a delay, on three separate days, against yourself reading normally.
- We shipped the instrument that can contradict our own marketing, and sometimes it will.
Every product in this category quotes a number, and it is always the same number. Words per minute. It is the easiest quantity in reading to measure and the easiest to inflate, because nothing in it refers to whether anything was understood. You could raise your own figure to four thousand words per minute this afternoon. Hold the page at arm’s length and turn it.
The unit that would mean something
The claim that survives scrutiny is not a rate. It is a pair — a rate and a recall score, reported together, alongside the test that produced them. Neither half means anything alone. Comprehension at ninety per cent is unimpressive if the rate was ninety words per minute; a rate of seven hundred is uninteresting if recall fell to chance. The useful quantity is something closer to comprehension per minute, and the reason almost nobody reports it is not a conspiracy. It is that measuring it properly is tedious, and measuring it improperly produces a better number.
This is a methodology piece, then, and it applies to us before it applies to anybody else. By the end of it you will have a protocol capable of embarrassing our own product, which is the point. Our position is that most of the industry’s numbers are artefacts of measurement rather than findings, and that position is worthless unless we hand you the means to hold ours to the same standard.
Where the number goes wrong
Speed-reading measurements fail in a small number of well-understood ways, and they fail in the direction of the flattering result almost every time. That asymmetry is worth noticing on its own. An error that made the number look worse would have been found and fixed within a week.
Start with the passage. If you already know the material you are not reading it, you are confirming it, at whatever rate confirmation runs. Predictable prose behaves the same way: a familiar genre with familiar structure lets you infer half of each sentence before it arrives, which is a genuine skill but not the one under test. Then the questions. Multiple-choice items written from a passage are notoriously answerable from general knowledge; give them to somebody who never saw the text and they will typically score well above chance, which means part of every reported comprehension score was never about the reading at all.
Timing matters as much as content. Test immediately and you are measuring what is still sitting in working memory, which decays. Test after a day and you are measuring what was learned, which is what anyone actually wants from a document. Almost every speed-reading demonstration tests immediately. Then there is self-reported comprehension, the weakest instrument in the set: readers routinely overestimate how much they took in, the illusion is strongest when the material was fluent and easy to process, and speed reading is close to a perfect machine for producing fluency without understanding.
Two more, and these are the ones well-intentioned people still miss. Practice effects: run four trials and the fourth will beat the first regardless of technique, because you have warmed up, learned the test format and worked out what is being asked of you. And the missing control — the same reader, reading normally, on matched material, tested the same way. Without it there is no comparison being made at all, only a number floating free. A great deal of what gets reported as a speed-reading gain is the practice effect wearing a jacket, which is roughly the story of what happened to Evelyn Wood.
Even the rate is soft. A word can be a whitespace-delimited token or a standardised five characters, and the two disagree by a fair margin on technical prose. Time can be counted from first word to last, or across the session including pauses. And any tool that modulates its own pacing shows a nominal figure running above the true one — ours stretches dwell up to 2.8× at paragraph breaks, 2.2× at sentence ends and 1.5× on long words, so on dense text the words actually delivered sit under the number on the dial. We report that gap rather than hide it, and you should assume it exists wherever a tool does anything clever with timing.
| Error | What it does to the number | Fix |
|---|---|---|
| Familiar passage | Confirmation read at confirmation speed; rate rises, recall holds | Unseen material from your own reading pile |
| Easy or predictable prose | Half of each sentence is inferred before it arrives | Match difficulty to what you actually read, not to a demo text |
| Questions answerable without the text | Scores well above chance on general knowledge alone | Give the test to somebody who never read the passage |
| Immediate testing | Measures working memory, which decays | Wait several hours at minimum, a day if you can |
| Self-reported comprehension | Tracks fluency rather than understanding, often inversely with speed | Score against fixed targets, not against how it felt |
| Practice across trials | The last run is the best whatever the technique | Counterbalance the order, or discard the first run |
| No control condition | Leaves nothing for the number to be better than | Read matched material normally, same test, same delay |
A rate without a recall score is not a measurement. It is a boast.
The only unit that means anything
- The passage. What it was, how long, how difficult, and whether the reader had seen it or anything closely resembling it before.
- The test. Format, number of items, and some evidence that the items are not answerable without having read the passage.
- The delay. How long between reading and testing. Immediate and next-day are different measurements and deserve different names.
- The control. The same readers, reading normally, on matched material, tested in the same format at the same delay.
- The spread. How many readers, and how much they varied. A mean with no spread attached is half a result.
A protocol you can actually run
None of this is secret and none of it is difficult. It is simply expensive in the only currency a product page cares about: time, and the risk of a disappointing answer. A properly controlled comparison takes days, needs more than one reader, and can come back saying the technique did nothing. A demonstration in which somebody flicks through a paperback and answers three questions about the plot takes ninety seconds and never comes back saying anything at all.
What follows takes about forty minutes a day for three days. It is not a controlled trial and it will not produce a publishable result. It will produce something more useful to you than a published result could: an estimate of what speed costs you, on the material you actually read, measured by an instrument you set up yourself.
Choose matched passages
Four to six extracts of roughly a thousand words each, from the same source or sources of comparable difficulty, none of which you have read. Use your own material — the filings, papers or chapters you would have read anyway. A children’s story chosen to flatter the result is not a test.
Write one test format and reuse it
The same number of questions each time, in the same style, targeting specific content rather than gist. If your tool ships a recall test, use it — but check that its items are not answerable by somebody who never saw the passage.
Run the control first
One passage read normally, on paper or screen, at whatever pace feels natural. Record the rate honestly: words divided by minutes, not the figure a device shows you. This is the only number everything else gets compared against.
Change one thing at a time
One passage per condition — the technique at your normal rate, then the technique faster. If you change pace and mode together you will not know which of them moved the score.
Wait before testing
Several hours at minimum, a day if you can manage it, and the same delay for every condition. Otherwise the delay becomes the variable you are unknowingly measuring.
Repeat on three separate days
Counterbalance the order so that no condition is always first. A single session is noise: attention, sleep, coffee and the passage itself move the score more than most techniques do.
One addition will make the result more honest still, and it takes ten seconds. Before each test, write down the score you expect. The gap between your prediction and your result is often more informative than the result itself, because it tells you how far you can trust your own sense of having understood something — and that sense is what every speed-reading claim is quietly resting on.
The instrument that can contradict us
The app ends every passage with a recall check. It mixes words that genuinely appeared with plausible distractors drawn from the same lexicon, so recognising the general shape of the topic is not enough to pass it. The score is shown next to the rate rather than on some separate screen, and the whole session — rate, settings, resolver breakdown, comprehension history — exports as a plain-text report. The words-read counter records the furthest point you reached, so rewinding cannot inflate it.
Building that had an obvious consequence, and we would rather state it than wait to be asked. We have shipped the instrument that can falsify our own marketing. Some proportion of the people who run it properly will find that their comprehension drops as they speed up, that the paired images did nothing for them, or that the voice made both numbers worse. We do not know what that proportion is. Nobody has tested this particular combination in a controlled trial, ourselves included, which is why the science page labels it a bet rather than a finding.
Which leaves the goal somewhere less comfortable than this industry likes to put it. The aim was never faster at no cost. There is a trade-off curve, it is well documented, and past roughly 400 to 500 words per minute on unfamiliar prose comprehension falls away until what is left is skimming — a real skill, a useful one, and not reading. The honest question is narrower: for this document, this afternoon, what is the fastest rate at which the cost is one you would accept? A contract and a magazine feature give different answers, and a reader who knows both is doing something no words-per-minute figure has ever described.
A note on what this is. Signal is written in-house by the team that builds Reader Inc., so treat it as an argument rather than a review. Nothing here is medical, psychological or educational advice, and the app is not a treatment, therapy or diagnosis for any condition. Where we describe research we describe it in general terms; where we are reasoning past the evidence we say so. The app is free, runs entirely on your own device, and ships with a comprehension test switched on — which means you can check every claim we make against your own reading rather than taking our word for it.
Read next
A field guide to speed readers
Six different machines share one shelf in the app store, and they do not do the same job. What each family removes, what it puts back, and where the evidence runs out — including for ours.
What we will not claim
The accessibility market has a long habit of shipping confidence ahead of evidence. We sell a reading app, so the habit is available to us too. This is the list of sentences we have decided not to write.
Let the word come to you
Word-by-word presentation removes eye movement completely, and eye movement is a real cost. It is also a much smaller cost than the speed-reading industry needs it to be. Here is the arithmetic, with nothing rounded up.