What Cognitive Load Costs at Checkout
The number everyone quotes for removing a form field has no study behind it, and one practitioner who removed fields watched conversion fall. The friction research that does hold up is about cookie banners and about effort, not about clicks.
Answer to the question
What does an extra form field really cost?
Nobody has published a credible figure. The famous per-field conversion numbers trace to vendor blog posts and one confounded cross-section of 404 landing pages. The solid evidence is elsewhere: a 2023 meta-analysis of 184 field-experiment papers covering more than 2.2 million people found that interventions changing effort beat interventions changing motivation, and consent-banner trials show structural design moving choices by double-digit percentage points.
Key takeaways
- The best-documented vendor test in the genre put five-field forms at 13.4 per cent conversion against 10 per cent for nine-field forms, with no sample size, randomisation detail or statistical test published.
- The widely cited curve from 18.2 per cent conversion at one field to 4.2 per cent at nine or more comes from 404 landing pages observed cross-sectionally, so the offers needing nine fields are not the offers needing one.
- A 2023 meta-analysis of 184 field-experiment papers covering more than 2.2 million participants found interventions that change effort outperform interventions that change intrinsic motivation.
- Removing the opt-out button from a consent notice's first page raised acceptance by 22 to 23 percentage points, in a field experiment with just 40 participants.
The question every checkout team asks is how much one more form field costs.
Ask it and somebody will hand you a dollar figure attached to Expedia and the year 2008. There is no study behind that figure. The better question is what the extra step does to the person taking it, and here is where we land before any of the evidence arrives: cognitive load at checkout is real, the working mechanism is effort rather than a full working memory, and the only trustworthy numbers come from cookie-banner field trials and from pooled behavioural work. Every per-field price list in circulation traces back to a vendor's blog post.
Two of our pieces already hold the other half of this subject, cancellation and the economics of exit friction and the question of who the friction actually lands on. This one is about entry friction and its mechanism, which is a different argument with a different literature behind it.
Where does the twelve-million-dollar field go?
The best-known number in checkout design says that removing a single field from one form recovered twelve million dollars a year. We went looking for the study. The retellings disagree about which field was removed, a company-name box in one version and a cut from six fields to five in another, and there is no paper from the company, no conference proceedings and no peer-reviewed record of any kind that we could find. The trail ends at marketing blogs citing each other.
There is no peer-reviewed number for the cost of one field or one step, and we looked.
The vendor trail, weakest item first:
1. The Expedia story. A dollar figure with no study attached, repeated since 2008, the tellings disagreeing about which field went.
2. The Marketo and MarketingExperiments registration test, which is the best-documented item in the whole genre. Five-field forms converted at 13.4 per cent against 10 per cent for nine-field forms. It is a vendor A/B test published without a sample size, a randomisation note or a statistical test, it was measuring cost per lead rather than conversion as such, and its own conclusion was that buying the data beat asking for it.
3. The NP Digital dataset of 404 landing pages, where conversion falls from 18.2 per cent at one field to 4.2 per cent at nine fields or more. Observational, cross-sectional, vendor-published and confounded, because the offers that need nine fields are not the offers that need one.
The useful counter-example comes from inside the industry. Michael Aagaard, writing at ConversionXL, removed form fields and watched conversions fall, because the fields he cut were the ones visitors engaged with most. He kept nine fields, rewrote the labels and reported a 19.2 per cent increase. Which fields, not how many. We also searched for a peer-reviewed field experiment that manipulated the number of checkout steps or form fields and measured completion, and we could not find one. That is not a claim that none exists.
Does the capacity argument actually apply?
The authority behind the per-field claim is cognitive load, and the chain runs from George Miller to Nelson Cowan to John Sweller. Miller's 1956 paper is a synthesis of other people's experiments rather than an experiment of his own, and he treated the number seven as a rhetorical device. Cowan's 2001 reconsideration put the real limit at about four chunks, and he also set out the conditions under which it shows up at all: information overload has to restrict chunks to single items, and recoding into bigger chunks has to be blocked. A shopping interface does neither.
Sweller's 1988 founding paper built cognitive load theory to design instruction for novices acquiring schemas, which are the organised patterns that let an expert see a problem whole. His own twenty-year review with Jeroen van Merriënboer and Fred Paas states the architecture plainly.
> Once information is in long-term memory, the capacity and duration limits of working memory disappear.
That sentence is why a familiar checkout costs a shopper almost nothing and an unfamiliar one costs something. An abandoned basket looks like fickleness. Read functionally it is a person declining to spend effort on a problem the interface invented, with no schema on offer in return.
The transfer is undefended on both sides. Nicholas Carr applies an instructional-design theory to leisure media use in The Shallows without arguing for the move, and our own application of cognitive load to checkouts is an extension rather than a validated finding. Ton de Jong's 2010 critique marks the honest limit. He grants that the theory has explained a large body of experimental results, then argues that its division of load into intrinsic, extraneous and germane is hard to measure apart and hard to falsify. That is a conceptual argument, not a failed replication.
What does the screen itself cost a reader?
Reading is the one kind of effort at a checkout with large-scale evidence behind it. A 2018 meta-analysis led by Pablo Delgado at the University of Valencia pooled 54 studies and 171,055 readers published between 2000 and 2017, and found a small but consistent advantage for paper over screens on comprehension. Three things moved it. The advantage grew when reading was timed rather than self-paced, it held for informational and mixed texts but not for narrative ones, and it grew across the publication years.
Virginia Clinton's 2019 review, restricted to randomised experiments, covered 33 studies from 29 reports and 2,799 readers. Her picture is the same and narrower: a small paper advantage for expository text, nothing for narrative, no reliable difference in reading time, and readers on paper somewhat better at judging how well they had understood. That last result is the one a checkout designer should care about. A shopper reading terms on a screen is a worse judge of whether they followed them.
Carr's own hypertext evidence does not survive the same handling. He quotes Diana DeStefano and Jo-Anne LeFevre's 2007 review as covering 38 experiments and reporting that the demands of decision making and visual processing in hypertext impaired reading performance. We could not open that review, so we can only report Carr reporting it. The item in his chapter best shaped for the one-more-click argument, a 1999 study of link density, has no retrievable published record at all, so none of its findings appear here.
What do cookie banners prove?
This is where consumer cognitive load has been tested properly, because the law forced a billion interfaces into existence at once. Christine Utz and colleagues ran three field experiments on a live German website with more than 80,000 users in 2019. Notices placed lower on the screen, and lower-left, drew more interaction. A binary choice produced more acceptances than mechanisms asking for permission per category or per company. Their own summary is the sentence to keep: seemingly small implementation decisions can substantially affect whether and how people interact with a notice. Consent fatigue is the framing of that paper rather than a finding of it, and we found no study measuring how decisions degrade across repeated consent requests.
Midas Nouwens and colleagues scraped 680 sites among the top 10,000 in the United Kingdom in 2020 and found that 11.8 per cent met the authors' minimal reading of European law. Their field experiment then put 40 people through the eight commonest designs. Banner or barrier made no difference. Removing the opt-out button from the first page raised consent by 22 to 23 percentage points, and putting granular controls on the first page lowered it by 8 to 20 points. Forty participants is a small experiment, so those swings are a direction rather than an estimate.
Two later studies sharpen it. Hana Habib, with Lorrie Cranor among her co-authors, inspected 191 consent interfaces and then ran 1,109 people through 12 variants inside a shopping task on a prototype shop, which is the closest published setting to the one this article is about. Benjamin Maximilian Berens and colleagues measured actual behaviour in 2023 and found that the visual treatment of the accept and reject options mattered, that the labelling of the reject option mattered, and that biasing text had no significant effect. The lever is the button, not the copy.
Biasing text had no significant effect on what people chose. The visual treatment of the buttons did.
Which lever is stronger, effort or motivation?
The pooled answer arrived in 2023. Yu Luo and colleagues meta-analysed field experiments with real behavioural outcomes, 184 papers covering more than 2.2 million participants between 2008 and 2021, sorted by the cognitive process each intervention targeted. Interventions that changed effort outperformed interventions that changed intrinsic motivation.
Interventions that added friction and interventions that removed it came out similar in size, so the effort channel runs both ways and it is the strong one. The authors ask for caution because of possible publication bias, and that caution travels with the result. Richard Thaler's one-page 2018 editorial supplies the name for friction pointed the wrong way, sludge, and nothing more than the name.
The option-count literature is the place to be careful, and two meta-analyses have to be printed together. Benjamin Scheibehenne and colleagues pooled 63 conditions from 50 experiments and 5,036 people in 2010 and found the mean effect of assortment size on choice overload virtually zero, with no sufficient conditions for it identifiable, which is the jam study and what the pooled evidence did to it. Alexander Chernev and colleagues pooled 99 observations and 7,202 people in 2015 and recovered a reliable effect once moderators were modelled, with decision task difficulty, choice-set complexity, preference uncertainty and an effort-minimising goal each making overload more likely.
Both are serious. We read them as agreeing more than they look: the number of options is close to inert, and the difficulty of the task is not. Difficulty, rather than count, is what the cognitive load argument can honestly claim at a checkout.
What can this evidence not settle?
No peer-reviewed number exists for the cost of one field or one step. Cognitive load theory has never been validated on consumer interfaces, and nobody has published a shopper equivalent of the instructional experiments. The consent percentages come from one small field experiment in one legal regime. The screen-reading results are about comprehension, not purchase, and they vanish for narrative text. Consent fatigue is asserted and unmeasured. Both cookie-banner nulls, on banner style and on biasing copy, are real nulls, and they say that some of the things practitioners spend most on do nothing detectable.
Two products are being sold under one name. One is a price list per field, which rests on folklore, and the other is effort reduction, which rests on more than 2.2 million people in field experiments plus the consent literature and research on dark patterns and where the blame sits. Conversion optimisation charges the same for both, and only one of them has anything behind it.
Common questions
What does an extra form field really cost?
Nobody knows, and the published numbers are not studies. The twelve-million-dollar Expedia story has no paper behind it, the five-versus-nine-field test showing 13.4 per cent against 10 per cent is a vendor A/B test with no sample size reported, and the 404-page dataset running from 18.2 to 4.2 per cent is a confounded cross-section. We looked for a peer-reviewed field experiment and found none.
Is cognitive load a real effect at checkout?
The theory is real and mainstream, and its application to shopping is an extension nobody has tested. Cognitive load theory was built for teaching novices, and Sweller's own review notes that working-memory limits disappear once material sits in long-term memory, which is why a familiar checkout costs a shopper almost nothing. The defensible claim is about task difficulty, not about a capacity ceiling.
What is the best evidence that reducing friction works?
A 2023 meta-analysis led by Yu Luo covering 184 papers and more than 2.2 million people in field experiments found that interventions changing effort beat interventions changing intrinsic motivation, with friction added and friction removed coming out similar in size. The authors warn about possible publication bias. It is the strongest pooled evidence that effort, rather than persuasion, is the lever.
Do cookie banner designs change what people choose?
Yes, and the effective parts are structural. Christine Utz and colleagues found notice position and the shape of the choice mattered across more than 80,000 users. Midas Nouwens and colleagues found that removing the first-page opt-out raised consent by 22 to 23 percentage points among 40 participants. Benjamin Maximilian Berens and colleagues found the button labelling mattered and the persuasive text did not.
Sources
Miller, G. A. (1956). 'The magical number seven, plus or minus two: Some limits on our capacity for processing information.' Psychological Review, 63(2), 81-97. doi:10.1037/h0043158. [Published record read at Crossref; the 1956 article carries no abstract and reports no experiment of its own, so no sample size exists to cite.]
Cowan, N. (2001). 'The magical number 4 in short-term memory: A reconsideration of mental storage capacity.' Behavioral and Brain Sciences, 24(1), 87-114. doi:10.1017/s0140525x01003922. [Abstract read in full via OpenAlex, cross-checked against the open-access Cambridge PDF link.]
Sweller, J. (1988). 'Cognitive load during problem solving: Effects on learning.' Cognitive Science, 12(2), 257-285. doi:10.1016/0364-0213(88)90023-7. [Abstract read in full; no sample size stated.]
Sweller, J., van Merrienboer, J. J. G., & Paas, F. (2019). 'Cognitive architecture and instructional design: 20 years later.' Educational Psychology Review, 31(2), 261-292. doi:10.1007/s10648-019-09465-5. [Abstract read in full; open-access PDF at Springer. A review, not a study.]
de Jong, T. (2010). 'Cognitive load theory, educational research, and instructional design: some food for thought.' Instructional Science, 38(2), 105-134. doi:10.1007/s11251-009-9110-0. [Abstract read in full. A conceptual and methodological critique, not a failed replication.]
DeStefano, D., & LeFevre, J.-A. (2007). 'Cognitive load in hypertext reading: A review.' Computers in Human Behavior, 23(3), 1616-1641. doi:10.1016/j.chb.2005.08.012. [Abstract NOT retrievable: OpenAlex returns null, no deposited abstract at Crossref, closed at Elsevier, not indexed in Europe PMC. Published record, authors and pagination verified. The 38-study count and all quoted phrases reach us only through Nicholas Carr's account at p. 129 and are reported as such.]
Delgado, P., Vargas, C., Ackerman, R., & Salmeron, L. (2018). 'Don't throw away your printed books: A meta-analysis on the effects of reading media on reading comprehension.' Educational Research Review, 25, 23-38. doi:10.1016/j.edurev.2018.09.003. [Abstract read in full; open-access PDF at ScienceDirect. 54 studies, 171,055 participants, 2000 to 2017; Hedges' g = -0.21; paper advantage larger under time constraint, absent for narrative-only texts, growing across publication years.]
Clinton, V. (2019). 'Reading from paper compared to screens: A systematic review and meta-analysis.' Journal of Research in Reading, 42(2), 288-325. doi:10.1111/1467-9817.12269. [Abstract read in full. 29 reports, 33 studies; reading performance k = 33, n = 2,799, g = -0.25; reading time k = 14, n = 1,233, g = 0.08, no reliable difference; calibration k = 11, n = 698, g = 0.20 favouring paper; expository g = -0.32, narrative g = -0.04.]
Utz, C., Degeling, M., Fahl, S., Schaub, F., & Holz, T. (2019). '(Un)informed Consent.' Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, 973-990. doi:10.1145/3319535.3354212. [Abstract read in full; author order verified at Crossref. Three field experiments on a live German website, more than 80,000 unique users. Consent fatigue is asserted in the framing and is not a finding of the paper.]
Nouwens, M., Liccardi, I., Veale, M., Karger, D., & Kagal, L. (2020). 'Dark Patterns after the GDPR: Scraping Consent Pop-ups and Demonstrating their Influence.' Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems, 1-13. doi:10.1145/3313831.3376321. [Abstract read in full; author order verified at Crossref. Scrape n = 680 across the top 10,000 UK sites, 11.8 per cent meeting the authors' minimal legal reading; field experiment n = 40; notification style no effect; removing the first-page opt-out +22 to +23 percentage points; granular first-page controls -8 to -20 points.]
Habib, H., Li, M., Young, E., & Cranor, L. (2022). '"Okay, whatever": An Evaluation of Cookie Consent Interfaces.' CHI Conference on Human Factors in Computing Systems, 1-27. doi:10.1145/3491102.3501985. [Abstract read in full; author order verified at Crossref. 191 interfaces inspected; 1,109-participant between-subjects experiment, 12 variants, shopping task on a prototype e-commerce site.]
Berens, B. M., Bohlender, M., Dietmann, H., Krisam, C., Kulyk, O., & Volkamer, M. (2023). 'Cookie disclaimers: Dark patterns and lack of transparency.' Computers & Security, 136, 103507. doi:10.1016/j.cose.2023.103507. [Abstract read in full. Visual representation and reject-option labelling had significant effects on actual behaviour; biasing text had no significant impact. NOT VERIFIED: sample size, absent from the abstract and the deposited record.]
Thaler, R. H. (2018). 'Nudge, not sludge.' Science, 361(6401), 431. doi:10.1126/science.aau9241. [Full text read; a one-page editorial with no data, cited for the definition only.]
Luo, Y., Li, A., Soman, D., & Zhao, J. (2023). 'A meta-analytic cognitive framework of nudge and sludge.' Royal Society Open Science, 10(11), 230053. doi:10.1098/rsos.230053. [Abstract read in full. Field experiments with real behavioural measures, 184 papers, k = 184, N = 2,245,373, 2008 to 2021. Effort interventions outperformed intrinsic-motivation interventions; nudge and sludge effect sizes similar; authors caution over potential publication bias.]
Scheibehenne, B., Greifeneder, R., & Todd, P. M. (2010). 'Can there ever be too many options? A meta-analytic review of choice overload.' Journal of Consumer Research, 37(3), 409-425. doi:10.1086/651235. [Abstract read in full. 63 conditions from 50 experiments, N = 5,036; mean effect size virtually zero; no sufficient conditions identified.]
Chernev, A., Bockenholt, U., & Goodman, J. (2015). 'Choice overload: A conceptual review and meta-analysis.' Journal of Consumer Psychology, 25(2), 333-358. doi:10.1016/j.jcps.2014.08.002. [Abstract read in full. 99 observations, N = 7,202; choice-set complexity, decision task difficulty, preference uncertainty and decision goal all moderate; effect significant once moderators are modelled.]
Carr, N. (2010/2020). The Shallows: How the Internet Is Changing the Way We Think, Read and Remember. Atlantic Books, London. ISBN 978-1-83895-258-7. [Full vault read-through, pp. 1-237 plus endnotes; seed claim at Ch. 7, pp. 115-143, cognitive load theory pp. 123-126.]
Vendor material, named as such and not as research: the Expedia 'twelve million dollars from one field' story, dated 2008, with retellings disagreeing on which field was removed and no company-published study, conference paper or peer-reviewed record located; the Marketo / MarketingExperiments registration-form test, five-field 13.4 per cent against nine-field 10 per cent, published without sample size, randomisation detail or statistical test and measuring cost per lead; the NP Digital dataset of 404 landing pages, 18.2 per cent at one field falling to 4.2 per cent at nine or more, observational and cross-sectional; Michael Aagaard at ConversionXL, who removed fields and saw conversions fall, then kept nine fields, rewrote the labels and reported a 19.2 per cent increase. [Retrieved via WebSearch 2026-10-17. NOT FOUND, after four OpenAlex query formulations and an extended WebSearch: any peer-reviewed field experiment manipulating checkout steps or form-field count and measuring completion or purchase.]