I've been automating browser tasks across a few dozen sites for a small SaaS. This is a field report on what blocks you, ranked by how much time it cost me.
Not a "how to evade detection" post — most of what I do is submitting my own product to directories and publishing my own articles. The interesting part is how inconsistent the barriers are.
The barrier that stops everything: Google OAuth
Far and away the most common hard stop.
Flow: click "Continue with Google" → an account chooser appears listing the logged-in account → click it → nothing happens.
The URL stays on accounts.google.com. There's no error, no spinner, no CAPTCHA. The page just doesn't advance. I've seen this on four different sites (a directory, two blog platforms, one product directory), with a real logged-in Google session in the profile, headed mode, no automation flags.
Workaround: there isn't a good one. Your options are:
- Manual login once, reuse the profile. Works, requires a human for 30 seconds per platform, which is fine if you're doing this once.
- Find a non-Google login. This is the actual answer below.
The distinction that matters: who owns the login flow?
| Flow owner | Reliability |
|---|---|
| GitHub OAuth | ✅ works |
| Email + password | ✅ works |
| Firebase (Google) on a third-party site | ⚠️ sometimes works |
| Google's own login page | ❌ reliably fails |
Firebase-based Google login worked for me on one site. It's Google identity underneath, but the UI is served from the site's own domain, and that appears to be the difference. Google's own accounts.google.com page is the wall.
So my current rule when evaluating a new platform: does it offer "Sign in with GitHub"? If yes, it'll take five minutes. If it's Google-only, budget a manual login.
Cloudflare's interstitial
Not a CAPTCHA — the "Performing security verification" page that appears before content. Hit on npmjs.com package pages and a couple of tool directories.
For read-only scraping this is a genuine wall. My workaround was to stop scraping and use the API or just not check. For posting, it never appeared (different path through the site).
CAPTCHA types, in order of solvability
1. Cloudflare Turnstile (checkbox/interactive)
Fine in a headed window. Fails in headless — it reports auto/failure. If a human can click it, it works.
2. Image CAPTCHA, Latin characters
Sometimes readable at 4-5x upscale with contrast enhancement. Accuracy is bad enough that I don't do it. Autofilling a wrong answer often locks the account, which is much worse than not submitting.
3. Image CAPTCHA, Japanese kana with noise lines
Yahoo Japan and one Japanese blog platform. I could not read these reliably at all. Upscaling to 900px wide and the characters are still guesswork. This ended two channels for me.
4. Cloudflare challenge that never resolves
Some sites just loop. Nothing to do.
The subtle one: forms that look fillable and aren't
A directory site had a submission form with 11 fields. I filled all of them, clicked Submit, and the page reloaded empty. No validation message, no error.
Turns out the submit was gated on a React state that only updates on real keyboard events, not dispatchEvent. Two fields accepted programmatic values; the rest silently reset.
The fix is page.keyboard.type() on a focused element rather than setting .value:
await page.click('input[name="tool_name"]');
await page.keyboard.type("My Tool", { delay: 40 });
Slower, ugly, works. If a form silently discards your input, this is usually why.
Rate limits that look like failures
dev.to: publishing more than one post every few minutes returns HTTP 429 with Rate limit reached, try again in 300 seconds. That's a 5-minute cooldown, not a ban. Worth knowing before you conclude your script broke.
Things that are genuinely easy
For contrast, because the internet makes this sound uniformly hostile:
- GitHub OAuth on three sites: clicked through in under a minute each
- Plain HTML forms with no client-side framework: always work
- Sites with a public API: trivially better than automating the UI
If a platform has an API, use it. The API exists precisely because the UI is a bad interface for programs.
The general rule I've landed on
Automation difficulty correlates with how much the platform cares about spam, and how big its anti-abuse team is.
A small directory with a plain form: 30 seconds.
A Y Combinator portfolio site with Firebase auth: maybe.
Anything Google operates: no.
Anything Japanese consumer-facing: image CAPTCHA, good luck.
Which means: the channels that are easy to automate are also the channels everyone else has automated, so they're saturated and often paid. The hard ones are hard for a reason, and usually that reason is that they're real communities.
That's an uncomfortable conclusion when your goal is distribution. It suggests the answer isn't better automation — it's fewer channels, done properly.
(I write about building Speech Kobo and the open data behind it on GitHub.)
Top comments (0)