There is real published research on CAPTCHA friction. It does not support a single headline percentage — and the percentages you will find quoted online mostly trace back to nothing.
Search for the conversion impact of CAPTCHAs and you will find confident numbers. Follow them back and most terminate in a vendor blog post citing another vendor blog post, with no sample size, no methodology and no link. We are a vendor too, which is exactly why this page cites only work you can go and read.
SpamKill previously published a conversion-improvement figure. We could not substantiate it to our own satisfaction, so we withdrew it, and you will not find it on this page or anywhere else on this site. A page arguing that unsourced conversion numbers should be distrusted cannot lean on one.
The most substantial recent work is Dazed & Confused: A Large-Scale Real-World User Study of reCAPTCHAv2 by Searles, Prapty and Tsudik at UC Irvine. It observed more than 3,600 distinct users over 13 months on a live account-creation and password-recovery service, with participants unaware they were being studied — so the behaviour is natural rather than lab-conditioned.
The authors conclude that reCAPTCHAv2 and similar technology should be deprecated on the grounds of, in their phrasing, immense cost and no security. UC Irvine's own write-up of the work puts the aggregate human cost at 819 million hours.
The same study measured perception, not just time: image challenges were rated annoying and scored "OK" on the System Usability Scale, while checkbox challenges were perceived as easy and scored "good". That gap matters operationally, because which one a visitor sees is not their choice — it is an escalation decision made about them.
The Stanford evaluation by Bursztein and colleagues, published at IEEE Symposium on Security and Privacy in 2010, covered more than 318,000 CAPTCHAs across 21 schemes — 13 image, 8 audio. It found audio CAPTCHAs particularly problematic, and non-native English speakers slower in general and less accurate on English-centric schemes. Difficulty is not evenly spread across your audience, so neither is the abandonment.
The W3C's Inaccessibility of CAPTCHA note states that every type of CAPTCHA will be unsolvable by users with certain disabilities. For those visitors the conversion effect is not a percentage — it is total. See is reCAPTCHA accessible?
No published work gives you a universal figure for how much a CAPTCHA reduces conversion on your form, and it is worth being clear why not:
That last point is the important one. The cost of a CAPTCHA is systematically harder to see than the benefit: blocked spam is counted, lost enquiries are not.
SpamKill has no challenge to abandon: detection happens after capture, scored on behaviour rather than on a test the visitor has to pass, at 99.9% accuracy across more than 100M+ submissions screened for 1,500+ businesses. When we are unsure about a real person they verify in one step and continue, and every held submission is recoverable — so a false positive is a delay you can see and fix, not a customer who silently vanished. We are not going to quote you a conversion uplift figure. Measure it on your own form; that is the only number that means anything. From $29/month with a 30-day free trial, no credit card.
Nobody can honestly give you a single number, and you should distrust any page that does. The effect depends on your traffic mix, device split, challenge type and how often the challenge escalates. What the published research establishes is direction and mechanism — CAPTCHAs cost real time, users find image challenges annoying, and some users cannot complete them at all — not a universal percentage.
Because they are repeated rather than measured. Trace a striking conversion figure back and it usually terminates in a vendor blog post with no methodology, no sample size and no link. That is not evidence, and citing it makes your own case weaker rather than stronger — including when the vendor is us, which is why our own retired conversion claim is not on this page.
It reduces it for the median visitor and concentrates it on a minority. Score-based systems only challenge visitors who score poorly — disproportionately people on assistive technology, older devices, unusual browsers and privacy tooling. Your average looks better while a specific group is quietly having a worse time, and they rarely tell you.
Run it as an experiment rather than trusting a benchmark. Measure form starts, form completions and spam volume for a fixed period, change one thing, and measure the same window again. Watch completions per start rather than raw submission counts, or a fall in spam will look like a fall in conversion.
Yes — on genuinely high-risk actions where a wrong decision is expensive, and where you have no better signal. The mistake is applying that reasoning to a contact form, where the cost of a fake submission is low and the cost of losing a real enquiry is high.
Paste your form HTML into the Form Transformer and see the protected version — no challenge for a visitor to abandon.
Open the Form Transformer