How to Score Proxy Providers on One Comparison Sheet

Checklists answer "is this acceptable". Comparisons answer "which of these is best for us". Those are different questions and they need different instruments.

Put three proxy vendors side by side on a feature checklist and you will get three columns that are each about 80 percent ticked, differing on items that are not equally important. One has better country coverage and worse session control. One is cheaper on paper and sells a different unit. One has excellent documentation and a 90-day minimum term. The checklist tells you all of that and refuses to rank it, so the decision gets made by whoever argues most confidently in the meeting.

A weighted sheet does two things a checklist cannot. It forces you to declare what matters before you look at the answers, which is the only real defence against motivated reasoning. And it makes the disagreement explicit: when two people score the same vendor 2 and 5 on the same criterion, you have found a criterion that is badly defined rather than a vendor that is ambiguous.

Some requirements are not criteria. They are gates, and a weighted score must never be allowed to rescue a vendor that fails one.

Run these as pass or fail before anything is scored. A failure removes the vendor from the sheet entirely rather than costing them points:

• Authentication methods you require. If your runners have dynamic egress addresses, a whitelist-only vendor is out. No score compensates for not being able to connect.
• Geographies you must have. Not "nice to have more". The specific countries or cities a contract or a client requires.
• Protocols you must have. SOCKS5, if anything non-HTTP passes through.
• Payment method you can actually use. Procurement that needs an invoice and a purchase order cannot buy from a card-only vendor, however good.
• Acceptable use policy compatibility. If your workload is prohibited by their published AUP, the relationship ends at the first incident. Read it first; proxy acceptable use policies explained covers what these documents typically restrict.
• Legal and compliance requirements your organisation imposes: data processing terms, jurisdiction, sourcing disclosures.

Gates are usually where a shortlist of eight becomes a shortlist of three, and that filtering is free. Do it before you spend an afternoon scoring.

Score each criterion 0 to 5. The anchors matter more than the scale: a criterion without a written anchor is an opinion with a number attached.

Fit

1. Measured success rate on your targets. Anchor: 0 if worse than your incumbent by more than the confidence interval of your sample, 3 if statistically indistinguishable from your incumbent, 5 if better by more than the interval. Measured, not quoted. Vendor dashboards do not count.

2. Exit type and coverage match. Anchor: 0 if it lacks a geography you need, 3 if it covers your required list, 5 if it covers the list plus city or ASN granularity you would use. Pool size is not part of this score, for reasons proxy pool size claims sets out.

3. Session control. Anchor: 0 if rotation is not configurable, 3 if sticky sessions exist with a documented duration, 5 if you can choose per-request rotation, timed rotation and held sessions, and the behaviour matches the documentation when you test it.

Cost

4. Normalised cost per 1,000 successful requests. Not the sticker price. See the next section for the conversion, and score relatively: 5 × (cheapest vendor's normalised cost / this vendor's normalised cost).

5. Cost predictability. Anchor: 0 if the bill moves with an axis you cannot forecast, 3 if it moves with one you can, 5 if it is flat monthly regardless of workload. This is the criterion that separates a quote from a budget.

Operations

6. Integration surface. Anchor: 0 if credentials are manual and singular, 3 if there is an API and sub-user support, 5 if you also get programmatic sub-users and credential rotation so each job holds its own revocable identity.

7. Observability. Anchor: 0 if usage is a single monthly number, 3 if there is per-day usage by sub-user, 5 if you can see per-session detail and export it. You cannot manage a cost you can only see in arrears.

8. Support responsiveness. Anchor: measured by your own tickets during evaluation, not by the promise on the website. 0 if a substantive question took more than a working day, 3 if answered same day, 5 if answered same day by someone who understood the question without escalation. Open one real ticket per vendor during the trial specifically to score this.

Risk

9. Commercial flexibility. Anchor: 0 for a long minimum term with no downgrade path, 3 for monthly billing with prorated changes, 5 for monthly billing, mid-cycle downgrades and a refund window that survives real usage. Monthly versus annual proxy plans covers the trade you are pricing here.

10. Transparency and compliance. Anchor: 0 if key terms are unpublished, 3 if pricing, AUP and an SLA are all public, 5 if they are public, specific and survive the reading in our guide to reading a proxy SLA. Unpublished figures score low here even when the vendor answers them in a sales call, because a private answer is not a commitment.

Ten criteria is deliberate. Twenty produces a sheet nobody fills in honestly, and five loses the distinctions you called the meeting about.

This is the step that makes the sheet work, and it is the step most comparisons skip.

Proxy vendors sell incompatible units: per address per month, per gigabyte, per concurrent thread, per API credit. You cannot score those against each other, and putting the entry prices side by side compares numbers that are not the same kind of thing. Convert everything to one unit instead:

cost per 1,000 successful requests
    =  (monthly plan cost × 1,000) / (monthly requests attempted × success rate)

Two things follow from that formula, and both are worth stating outright.

Success rate belongs in the denominator of price, not in a separate row. A vendor 12 percent cheaper with a success rate 10 points lower is not cheaper. Scoring price and success rate as independent criteria double-counts the good news and hides the interaction, which is why criterion 1 is about relative success rate and criterion 4 already contains the absolute one.

The conversion needs your volume, not the vendor's example volume. Per-gigabyte pricing is a volume curve, per-address pricing steps with tier size, and per-credit pricing has package breakpoints. Run the formula at the volume you will actually buy, and run it again at 3x that volume so you can see which vendor's curve is friendlier if you grow.

Worked shape, using your own inputs: a plan at $140 a month, 2.4 million requests attempted, 94 percent success. That is 140 × 1000 / (2,400,000 × 0.94), which is about $0.062 per 1,000 successful requests. Do the same arithmetic for every vendor on the sheet, including the incumbent, then score criterion 4 relative to the cheapest.

Weights are the decision. Scores are just inputs. Here are five profiles, each summing to 100, to start from and then adjust for your situation.

Read the columns for the reasoning. High-volume scraping puts 60 of its 100 points on success rate, cost and predictability, because at volume the bill and the retry rate are the whole story. Account management moves 20 points onto session control, because a rotating exit breaks the product. Geo and ad verification puts 25 on coverage, since a country you cannot reach is a job you cannot take. Enterprise procurement spreads weight almost evenly and loads the risk group, which is exactly what a procurement function exists to do, and our enterprise proxy procurement checklist covers the paperwork side.

Adjust these. What you must not do is adjust them after you have seen the totals.

Six process rules. They cost nothing and they are the difference between a defensible decision and a rationalised one.

1. Lock the weights in writing, with a date, before any vendor is scored. Email them to yourself if that is what it takes. This is the single most important rule on the page.
2. Score criteria in rows, not vendors in columns. Score every vendor on criterion 1 before touching criterion 2. Scoring a whole vendor at once lets a strong first impression bleed into unrelated rows.
3. Two scorers, independently, then compare. Record the spread, not just the average. Any criterion where two scorers differ by 3 or more points is a criterion whose anchor is broken; fix the anchor and rescore that row.
4. Missing data is a score on criterion 10, not a zero everywhere else. A vendor who does not publish their pool sourcing scores low on transparency. Do not also score their capability as zero on the basis of the absence.
5. Write one sentence of evidence per cell. Not a paragraph. One line saying where the number came from: a page you read, a ticket you opened, a measurement you ran. Cells without evidence get revisited.
6. Score the incumbent too. A comparison that omits the thing you already have cannot tell you whether switching is worth the work.

Where the evidence comes from splits cleanly. Criteria 2, 4, 9 and 10 come from published pages and terms, so read them and date what you read. Criteria 1, 3 and 7 come from your own testing, and our complete proxy testing guide covers the methods. Criteria 6 and 8 come from actually using the account during the evaluation window.

Three shortlisted vendors, deliberately anonymised. Every score below is invented to demonstrate the arithmetic. None of it is a measurement of any real provider.

Vendor A sells flat monthly capacity by concurrent thread. Vendor B sells rotating residential by the gigabyte. Vendor C sells cheap addresses per month. Their normalised costs per 1,000 successful requests came out at $0.06, $1.40 and $0.09 respectively, so criterion 4 scores 5.0, 0.2 and 3.3 under the relative rule.

Apply the high-volume scraping weights and divide by the 500-point maximum:

Same vendors. Same scores. Different weights, and the ranking at the top changes. A wins scraping by nearly 15 points and is a statistical tie with B on account work. The decision lives in the weights, not in the scores, which is why locking them first is not bureaucracy.

Two other things this example shows. Vendor B's near-zero on normalised cost does not eliminate it, because on the account-management profile that criterion carries only 10 points. And Vendor C, the cheapest sticker price on the shortlist, finishes last on both profiles, because a low success rate multiplies through the price formula and then costs points again on criterion 1.

Three rules for interpreting a total, in order of how often they are needed.

A gap under 5 points on a 100-point scale is a tie. Your inputs are not precise enough to support a finer distinction, and pretending otherwise is false rigour. When two vendors tie, stop scoring and decide on something you deliberately excluded: which contract you would rather exit, which support team you would rather call at 3am, which one your team already knows.

A large gap driven by one criterion is a finding, not a verdict. If Vendor A wins by 15 points and 12 of them come from criterion 4, you have learned that the decision is about price. Confirm the price inputs are right before you act, because one wrong volume assumption moves that criterion by more than the gap.

A vendor that wins on both a cost-heavy and a risk-heavy profile is a genuinely strong choice. Run at least two weight profiles, including one that reflects how your needs might look in a year. The vendor that survives both is the one to sign.

Finally, sanity-check the winner against the warning signs in how to spot a fake proxy provider before you buy. A scoring sheet rewards the vendor who answers the most questions well, and answering questions well is cheap.

The first and the fourth are the expensive ones. Weight drift produces a decision that was made before the sheet existed, and unit confusion produces a decision based on numbers that were never comparable.

The sheet has a second life, and it is worth more than the first.

Keep it. At renewal, rescore the incumbent on the same ten criteria with the same weights, using a year of actual evidence rather than trial impressions. The rows that move are the argument for whatever you want next: a better rate, a shorter term, an SLA credit, or a switch. Walking into a renewal with a documented drop on criteria 1 and 8 is a materially stronger position than walking in with a sense that things have got worse, and it is exactly the kind of evidence that negotiating volume discounts turns into money.

Do it 60 days out, not 7. Seven days out you have no alternative and nothing to bargain with, and the sheet will tell you so.

For reference while scoring criterion 4, SparkProxy's own published plans are Starter $75 a month for 100 threads and 5 whitelist slots, Core $140 for 250 threads and 10 slots, Boost $240 for 500 threads and 15 slots, and Plus $440 for 1,000 threads and 25 slots, all flat monthly with unlimited bandwidth and 30 days validity. Flat monthly pricing scores well on criterion 5 by construction, and that is the honest reason a vendor selling it recommends the criterion. Weigh it accordingly.