One controlled study on web forms is worth more than the rest of the literature combined, and it reports first-attempt success rising from 42 percent to 78 percent. Sixty-five participants, three real registration forms from live company websites, each tested in its original version and in a version rewritten to comply with twenty published guidelines.
That is the good news. The complication is that it tested the twenty guidelines together, so it cannot tell you which of them mattered. And when you go looking for evidence behind the individual guidelines, most of them do not have any.
This article separates the two, because a form that loses meetings is worth fixing on the basis of the study that exists rather than the fourteen that are quoted as if they did.
The study that carries the weight
Worth reading in detail, because its design is unusually clean and its limitations are unusually clear.
The method. An eye-tracking laboratory study, between subjects, 65 participants with a mean age of 27.5, all self-rated experienced internet users. Thirty-two were assigned to the original-forms condition and 33 to the improved condition, with groups matched on age, education and computer knowledge.
The material. Registration forms from three real company websites, each in its original state and in a version rewritten to satisfy twenty published guidelines.
The results, form by form, on first attempt. On the first form, 10 of 32 succeeded originally against 24 of 33 improved. On the second, 9 of 32 against 24 of 33. On the third, 22 of 33 against 28 of 32. All three differences were statistically significant.
Aggregated, that is 42 percent against 78 percent.
Completion time is the honest footnote. It improved significantly on two of the three forms and not on the third, where the difference did not reach significance. Reporting the aggregate without that detail overstates the finding.
And the limitation that matters most for practice. The study applied twenty guidelines together. It establishes that following published form guidance produces a large improvement. It does not establish that any particular guideline is responsible, and it cannot be cited for one.
This is the uncomfortable inventory, and it is worth doing because it tells you where to spend attention.
Label placement, the most repeated advice, has the weakest source. The article everyone cites reports saccade times of 500 milliseconds for left-aligned labels, 170 to 240 for right-aligned, and 50 for labels above the field. It states no participant count and identifies no equipment.
And a peer-reviewed study reached the opposite conclusion. An eye-tracking study published in 2008 found right-aligned labels produced the shortest completion times, at least for multi-column forms. Its sample size is not publicly available either, so the honest summary is that the question is open rather than settled.
Field count has a benchmark but not a causal result. A checkout study reports “the average checkout flow in 2024 is 5.1 steps long and contains 11.3 form fields”, against a stated view that “most sites need only 8 form fields in total”, with the trend falling from 12.7 in 2019. That is descriptive. It does not say that removing fields raises conversion.
Inline validation has a real study, and it is small. Twenty-two participants across six form variations, with the best-performing variant showing a 22 percent higher success rate, 22 percent fewer errors and 42 percent faster completion. No significance tests are reported, and it is industry research.
And several familiar guidelines have nothing behind them. Published guidance on placeholders, on marking required fields, on single-column layout and on error message specificity carries no cited study or sample size in its source. That does not make the advice wrong. It makes it advice.
One guideline does have a hard number, from an unusual source. A dataset of 2,130 conference participants found “99.9% of city names were 19 characters or shorter, making 19 characters a reasonable width for a city field.”
Between the study and the folklore sits a set of rules with actual force, and they cover most of what goes wrong on a booking flow.
Every field that asks about the person must declare its purpose. At AA, the input purpose must be programmatically determinable, which means correct autocomplete tokens: name, organization, organization-title, email, tel.
Nothing may be asked twice in one process. At Level A: “Information previously entered by or provided to the user that is required to be entered again in the same process is either: auto-populated, or available for the user to select.”
Errors must be text. At Level A, the item in error must be identified “and the error is described to the user in text.” A red outline is not a description.
And a correction must be offered where one is known. At AA, if suggestions for correction “are known, then the suggestions are provided to the user.”
The placeholder question is settled by the HTML specification rather than by usability opinion. “The placeholder attribute should not be used as an alternative to a label.”
Targets have a minimum. 24 by 24 CSS pixels at AA, which catches small radio buttons, tight date pickers and the close control on a booking modal.
One that specifically affects scheduling widgets. Any function using a dragging movement must be achievable “by a single pointer without dragging”, which reaches time-range selectors and drag-to-pick calendars.
Asking for qualification information before the person has decided to talk. Company size, budget and timeline on the first screen convert a booking into an application.
Requiring an account. The most common self-inflicted step, and the one furthest from the visitor’s goal.
Hiding the calendar behind a form. The visitor came to see whether a time exists that works. Making them earn the sight of it inverts the order.
Not saying what happens next. Duration, who will be on the call, and what to bring. Absent, the visitor supplies the worst assumption.
Losing the entry when a session expires. Covered by an AAA criterion for a reason: “when an authenticated session expires, the user can continue the activity without loss of data after re-authenticating.”
And the one nobody notices until it costs a deal. No confirmation that survives the browser. An email with the time, the link and a way to reschedule, sent immediately, is the whole of it.
Reversing two steps costs nothing to build and removes the most common reason people leave. Source : Method (2026)
What to do with this
Take your booking flow and count the steps between arriving and seeing an available time. If the answer is more than one, that is the change to make first, and it does not need a test.
Then apply the requirements that are not opinions: correct autocomplete tokens, nothing asked twice, errors written as text, targets at 24 pixels, and no drag-only interactions in the scheduler.
Move every qualification question after the booking is confirmed. You will still get the answers, from people who have already committed to the conversation rather than from people deciding whether to.
And treat the rest as convention rather than evidence. Label alignment, placeholder style and single-column layout are all worth a decision and none is worth a fortnight, because the published sources behind them do not carry the weight the argument assumes.
In one controlled study, yes and substantially. First-attempt success went from 42 percent to 78 percent across three real forms. The caveat is that it applied 20 guidelines together and cannot tell you which one did the work.
How many fields should a form have?
No study answers that causally. A checkout benchmark reports an average of 11.3 fields against a stated view that 8 suffice, but that is a descriptive count, not a conversion result.
Should labels go above the fields?
The most-cited source for that advice publishes no participant count and no equipment, and a peer-reviewed eye-tracking study found the opposite for multi-column forms. Treat it as convention, not evidence.
Is inline validation worth it?
The best-known study reports a 22 percent higher success rate for the best variant, from 22 participants across six variations with no significance tests reported. Promising, small, and industry-run.