It suits comparing alternatives rather than absolute time prediction

What it is

GOMS is most reliable for comparing operation sequences for the same task across interface alternatives: alternative A has six fewer steps, one fewer mode switch, and one fewer decision point, which is more robust than reporting "8.4 seconds faster." Absolute time depends on operator constants, devices, users, and task assumptions. This card and the previous one, "it predicts only error-free skilled performance," are the same limitation showing up in two different uses: that one says the model's number cannot be taken as real elapsed time; this one goes further and shows that, even so, that inaccurate number remains useful for comparison — not because the number becomes accurate, but because the act of comparing cancels out the part that is not.

Why it happens

Absolute time is unreliable because it depends on a whole set of operator constants — keystroke time, pointing time, mental preparation time — each of which is itself affected by input device, user proficiency, system latency, and individual differences; get any one constant wrong and the resulting seconds shift as a whole. But if the comparison is between two different alternatives for the same task, done by the same population on the same device, these constant errors act on both alternatives simultaneously, in roughly the same direction and magnitude, and largely cancel out when subtracted. What is left mostly reflects the genuine structural difference between the two alternatives in operator count, decision points, and mode switches. This is why "alternative A has six fewer steps than alternative B" is a more robust conclusion than "alternative A is 8.4 seconds faster": the former is a structural fact that holds regardless of how the constants were calibrated, while the latter rests entirely on those error-prone constants — a 20% miscalibration can make "8.4 seconds" unrecognizable, but the fact of "six fewer steps" does not change.

Where it stops holding

If the two alternatives change not just the operation sequence but the user's actual decision strategy or error rate, comparing step counts under the "ideal sequence" alone can mislead — alternative B might have fewer steps because it merges two originally separate confirmation actions into one, but in real use that makes users more likely to overlook something that needed confirming, and the resulting rise in error rate can offset or exceed the benefit of fewer steps, a consequence entirely invisible to a "step count comparison." When the theoretical time difference between two alternatives is itself small, it can be completely swamped by genuine random noise, and relying purely on the model's relative difference to draw a conclusion is unreliable without real measurement to confirm it. Tasks with substantial waiting, network latency, or differences in cognitive complexity are also poorly suited to comparison by step count alone, since the time variation those factors cause often exceeds the difference in the operation sequence itself. Fewer steps cannot automatically be equated with "better" either — discoverability and the cost of error are two things the model simply cannot express, and both need to be brought in separately.

Applying it

• Write a GOMS or KLM sequence for each candidate alternative and compare the differences in operator count, decision points, number of mode switches, and waiting segments.
• Report relative differences, key assumptions, and a concrete step list as the primary finding, rather than throwing out a single number in seconds that reads like a precise measurement.
• For the alternative ultimately selected, time skilled users in real use to confirm the model-predicted relative advantage actually holds up, rather than stopping at the paper analysis.
• How to check: list discoverability and error-recovery cost separately, since the model cannot express them, and weigh them alongside the step-count comparison in the final judgment — an alternative with fewer steps but worse discoverability should not automatically win just because it leads on step count.

Related

• Same group: B4.04.1 Decompose tasks into goals, operators, methods, and selection rules · B4.04.2 It predicts only error-free skilled performance
• Nearby: A10 Reaction Time and Movement Time · B2.26 Directness and Indirection Trade-offs
• Search terms: GOMS comparison · KLM · relative cost · noise cancellation