What a shared proxy pool really shares

Two rows on the same seller’s page. Dedicated, $28 a month. Shared, $6. Every other cell in both rows matched, down to the bandwidth line, and the word was doing all of the work.

I paid the $6, pointed it at a retailer I had scraped happily for a year, and the first request came back as a challenge page. The first one. I had not had time to do anything wrong yet.

That is the product behaving exactly as designed. It just was not described in a way that told me what I had bought.

I sell mobile proxy lines, one line to one customer, on Singapore carrier sims that run me about $10 a month in data plus roughly $1.50 in modem depreciation. My incentive is obvious, so read on with a hand on your wallet. I also buy shared pools most weeks for my own scraping, and for the bulk of what I do dedicated would be me setting money on fire.

Bandwidth is the one thing you are barely sharing

On a pricing page, shared describes capacity. Several customers, one address, the cost divided. You rent a slice of a pipe you were never going to fill.

That framing would hold up if bandwidth were the scarce resource. It isn’t.

A price monitor pulling 200,000 pages a month at 180 KB each moves 36 GB, over a link that was never near the bottleneck. I have never been throttled off a proxy by a neighbour’s throughput. What runs out on a shared address is its welcome.

A separate piece here on reading a proxy pricing page handles dedicated versus shared as a billing model, which is the right lens with two quotes open side by side. This is the other question. Once the money has moved, what is physically in the bundle.

What is actually in the bundle

An ip address arrives at a website carrying a record. Not a formal one, and not one anybody publishes. A scattered record, written independently by every site that has ever seen it.

A marketplace keeps its own note of which addresses behave like scrapers. An anti bot vendor keeps a score and sells it onward. A login form counts failed attempts by source and remembers.

None of those systems know which customer of which reseller was holding the address at the time, and none has a way to find out. The address is the identity, and the whole unit of judgement.

So a share of an address is a share of that record. All of it, including the parts written while you were nowhere near it.

The rate limit does not know your name

This is the part I get argued with about, so here is the arithmetic.

Say a target allows 120 requests an hour per address. Four customers sit on the pool, all four have read the same sensible advice about pacing, and all four settle at 40 an hour. Comfortably inside the ceiling.

The target counts 160 and starts refusing.

Then each of you opens your own logs, sees 40 an hour against a documented limit of 120, and concludes the target is broken or the proxy is junk. Nobody has misread their own numbers. Each set describes a quarter of the traffic.

Your pacing is real and it protects nobody, because the quantity you are pacing is not the quantity being measured. A separate piece here works out how much concurrency is safe, and every figure in it assumes the address is yours alone. On a shared pool you divide those figures by a number the seller will not give you.

Rotation moves you around inside the same neighbourhood

Big targets gave up scoring addresses one at a time years ago. They score the block an address sits in and the network it is registered to, because tracking a few thousand ranges is cheaper than tracking millions of addresses.

Once a run of addresses out of one registration keeps producing the same shape of traffic, the cheap defence is to stop treating them as addresses and start treating them as a category.

That decision covers a range. Rotating inside the range achieves nothing at all, since rotation hands you a different address in the same category.

Which makes the standard support remedy close to worthless. You report a blocked address, the seller rotates you, and the replacement comes out of the stock the first one came from. I burned through eleven addresses on one pool against one target in an afternoon and hit the same wall on nine. The pool had not handed me a bad address. It had handed me a bad neighbourhood, and no fresh ip out of that neighbourhood was going to repair it.

That is the failure mode that costs people a fortnight of rewriting their own scraper for no reason.

The half that is not on anybody’s page

Everything above is about the people on the address at the same moment as you. That is the smaller half.

The sharing also runs backwards through time. A rotating pool hands you an address at nine in the morning. Somebody held it at three. A customer who cancelled last year held it for eight months before that. Every one of them left marks on every site they touched, and all of those marks are still attached to the address you were just issued.

There is no way to inspect any of it. No log, no handover note, nothing the seller could produce even if they felt generous. A seller knows which of their accounts had an address. They do not know what those accounts did with it, or what one particular retailer concluded about it in April.

A reputation lookup gets you nearer than nothing. It tells you whether one of the commercial scoring services has already made its mind up. It tells you nothing about what your target remembers, and your target holds the only opinion in this that costs you money.

You are buying a used address with no service history. An honest listing would print that on the page.

Nothing you do on it accrues

The other direction gets less attention and bothers me more.

On a line only I use, three months of sane behaviour against one target is an asset. The address stops reading as new. Its failed login count sits at zero. That record builds while I sleep, and it belongs to me.

On a shared address you are writing into a notebook five other people also write in. Nothing accumulates. Whatever standing your careful Monday earns gets spent by somebody else on Tuesday, and it resets when the pool rotates you out.

So behaving well on a shared pool never compounds. It buys you the current week on the current address and then it is gone.

Where shared is the right purchase

Broad crawling across many unrelated targets. Touch two thousand domains, none of them more than a few hundred times, and no single site accumulates enough of a view to form an opinion. The sharing costs you nothing you could measure.

Low volume. Four thousand pages a month and you are neither the reason an address burns nor sitting on one long enough to inherit much.

Anything where failure just means retrying. If a block puts the record back on the queue and it lands twenty minutes later, a dead address is a rounding error in your run time, and paying four times more to dodge it is a bad trade.

Where it is a bad buy

A single high value target. If the business depends on rows out of one site, you have concentrated all of that inherited history onto the one opinion you cannot afford to lose. Worse, the other customers on that pool are probably aimed at the same site, since that is what the pool was marketed for.

Anything logged in. An account puts a name on the address. From that moment it stops being anonymous traffic and becomes your account’s home, and sharing that home with five strangers leaves something you care about sitting on behaviour you cannot see, cannot control, and could not explain if asked.

Anything where a block is expensive to undo. A permanent ban on an account you spent three months warming is not a retry. A data partner deciding you look abusive is not a retry either. When recovery is measured in days, the gap between $6 and $28 stops being the number worth optimising.

What I got wrong, from the selling side

I have sold shared capacity myself. If a customer had asked me what the previous holder of a rotated address did to one specific retailer, I could not have answered. I did not hold the data at that resolution, and nobody selling shared addresses does.

I also used to tell people that dedicated fixes this. It narrows it. On my own lines the history starts the day the sim went into the modem, and whatever the carrier’s block absorbed from real subscribers before it reached me is invisible to me too. Mobile addresses get recycled through people who were not careful. A separate comparison here of the three address types goes further into that.

Test the pool against the thing you actually care about

There is a full piece here on testing a proxy before you pay. The line from it that matters here: do not evaluate a shared pool against an ip checker or some neutral page and call it tested, because a neutral page has no opinion about anybody and will load identically for a burned address and a clean one. Send two hundred requests of your real shape at your real target, and count the rows your parser produced rather than the responses that came back.

The questions I put to a seller about how a pool is shared, and the check I run before one goes anywhere near a real job, are here.

Get new guides and videos first — join the Telegram channel.