I used queries that we developed for the hCaptcha account creation A/B test analysis for a snapshot analysis, with data from the first three weeks of January 2026. We find that the data allows us to easily answer the first two questions, and that we're unfortunately unable to answer the last question, which might be the most important one to answer. I'll be filing a subtask of T394744 to update the instrumentation to fix that problem.

What percentage of users who start filling out the form do not complete it?

This is called the abandonment rate, and it's the converse of the completion rate. For wikis that do not have hCaptcha enabled (all but eight wikis), we get the following table:

The same table for the eight wikis that have hCaptcha enabled for account creation (Chinese, English, Indonesian, Japanese, French, Persian, Portuguese, and Turkish):

How long does it typically take users to submit the form?

For both categories of wikis (hCaptcha and non-hCaptcha as defined above) and across both platforms, the median time to completion is less than a minute.

A place where we find some differences is that the eight wikis that have hCaptcha see a higher proportion of account creations that are completed relatively quickly. For the "less than 30 seconds" category, hCaptcha wikis have a substantially higher proportion (desktop web: 26.9%, mobile web: 24.6%), than the other wikis (desktop web: 19.3%, mobile web: 16.1%). Additional analysis is needed to understand more about these behavioural differences, e.g. if they're driven by automated creations or if there are other reasons behind them).

Additional observations:
We've previously pulled data twice and see a lot of variation in completion rates. This indicates that we might be interested in monitoring completion rates more closely, i.e setting up an Airflow job to aggregate it on a daily basis. That would then allow us to possibly identify challenges such as significant increases in creation attempts at specific wikis, sudden changes in completion rate, etc.

At what step in the account creation funnel do users who start filling out the form bounce?

I reviewed bounce rates for users who start filling out the form, using the element_id field to understand which fields in the form are contributing most to user drop-off. Please see summary of results below and let me know if you have any questions.

Data
Data reflects events logged in March 2026 across all wikis (excluding wikis in the hCaptcha trial) with splits by desktop and mobile.

The tables below shows the proportion of account creation attempts where that specific field was the final interaction (Field). The Last Action column shows the final recorded user interaction with the field. This can be either:

• action = type: The user was in the middle of entering data into a field when they dropped off
• action = blur: The user finished interacting with a field and moved their focus away (likely hit a validation error)

See results below and let me know if you have any questions:

Results:

Some Key Insights:

• On mobile, most users drop off at the user_name field after interacting with the field. This is most likely due to a validation error.
• On desktop, the majority of drop-offs were associated with the captcha field (element_id = captcha AND action = type), accounting for 71% of drop-offs. After this, most users dropped off at user_name field.
• Password and confirm_password drop-offs account for roughly 26% of all bounces on mobile, while they only account for 6.75% on desktop. This likely highlights users' difficulty typing characters using the smaller mobile screen and editor.

Desktop Bounce Rates While Filling out Form

Mobile Bounce Rates While Filling out Form

cc @KStoller-WMF

Thank you @MNeisler!

On desktop, the majority of drop-offs were associated with the captcha field (element_id = captcha AND action = type), accounting for 71% of drop-offs.

It's surprising to see this so high on desktop and so much lower on mobile.

@EMill-WMF or @kostajh
Is that likely because bots, scripts, and scraping tools disproportionately run from desktop-like environments? Or curious if you have any hypotheses or additional knowledge here?

In T402533#11857478, @KStoller-WMF wrote:

On desktop, the majority of drop-offs were associated with the captcha field (element_id = captcha AND action = type), accounting for 71% of drop-offs.

It's surprising to see this so high on desktop and so much lower on mobile.

@EMill-WMF or @kostajh
Is that likely because bots, scripts, and scraping tools disproportionately run from desktop-like environments? Or curious if you have any hypotheses or additional knowledge here?

Well, it's not that bots, scripts, and scraping tools are running _from_ desktop-like environments, but that they would be registering _as_ desktop-like environments. I'm not personally familiar with exactly how we run these categorizations, but it wouldn't surprise me if desktop is the catch-all, while mobile environments have to be affirmatively detected as likely-mobile.

There is also the specific use case of bots that are headless browsers. Those could likely be configured by the orchestrating entity to advertise themselves as either a desktop or mobile client, but I would expect the default (and preference for blending in) would be to show up as a desktop client.

In any case, it does not surprise me to see a higher desktop rate of drop-off from the CAPTCHA -- not due to it being harder for humans to solve on desktop for some reason, but for it to be doing its job (as best as FancyCaptcha can) at deterring simple scripts.