Apollo and Clay disagree by segment, not overall

Neither vendor is a superset of the other. Apollo contributed 527 people Clay never returned — slightly more than the 520 both found. Clay's advantage is breadth, not containment, and a team using only one of them is missing a population the other sees.

This measures agreement with LinkedIn, not correctness. LinkedIn is self-reported, frequently stale, and in the title field is often marketing copy rather than a job title. Where a vendor and the reference profile disagree, the defensible statement is that they disagree — not that the vendor is wrong.

By set

Clay's perfect name score is partly a consequence of returning less. It carries 74 non-comparable names against Apollo's 31, and abbreviates roughly 15% of surnames (Jin C.), which the rubric scores as consistent rather than wrong. It also has 163 non-comparable locations against Apollo's 40, concentrated in Singapore where 94 of 321 rows carry no city at all. That is why n/c is printed beside every rate and never folded into it.

The difference between these vendors is volume, not per-row quality. Clay returned 1.8× more rows at a valid-row rate 1.9 points higher — inside the range that set-level variation moves. That aggregate reverses by set, so it is not a ranking; see valid rows by set.

Scope

Five ICP definitions, fixed before any data was pulled, chosen to stress vendor coverage along different axes. Full specification in the methodology.

Set 5 is only approximately harmonised, and is marked wherever it appears. Apollo's SaaS filter is an opaque vendor keyword tag with no Clay equivalent, and three defensible Clay encodings of “SaaS” span a 180× range in result count. The industry-enum encoding was chosen. The cost of that choice is that set 5's coverage comparison confounds vendor coverage with vertical definition, and its company-agreement figure is the least trustworthy of the five.

What the shapes show

An earlier Clay pull was excluded from every figure. It used a phrase-matching encoding not comparable to Apollo's. It is retained but unreported. The cost of excluding it is that this study cannot answer “what does a practitioner get by typing this in naively” — a legitimate question, and a different one.

How verdicts were reached

Rules decide first and escalate only where they decline. The model therefore settles few comparisons and disproportionately many contested ones.

Location compliance never reached the model. It is a closed structured vocabulary and is fully rule-decidable. Company and title agreement escalate hardest, which is where judge quality matters most — read instrument validation before trusting those two rates.

These counts are the third measurement of this quantity, not the first. The scraping actor reports its own quota refusals in the same shape as a per-profile failure. Reading a refusal as a per-profile failure wrote 2,428 fake dead links, which would have been published as a data-quality finding about both vendors when it was a fact about our billing plan. Correcting that too aggressively then discarded 1,428 profiles already paid for. The discriminator is a field the per-profile error carries and the actor-level refusal does not; 1,348 of the discarded profiles were recovered at no extra cost and the remaining 80 proved genuinely dead. The first two measurements were wrong in opposite directions.

Instrument validation

Two parts of this study are themselves measuring instruments. Both were measured, and both were found imperfect.

Two caveats, both material. The hand labels were produced by a language model validating another language model, so shared blind spots inflate agreement — this is screening, not independent ground truth, and no figure on this page rests on the judge alone without the underlying values being retrievable. And no acceptance floor was agreed before measuring, so 94.5% is what was observed, not a threshold that was cleared. A pre-registered floor would be a stronger claim and is not available here.

Disagreements are published rather than summarised. The model treated OPW Retail Fueling and OPW Fueling Components as different employers where the labeller read one employer with two division names, and the same for OSG USA, INC. and OSG Tap & Die, Inc. It matched Bolivar (Ohio) to Ciudad Bolívar (Venezuela) on string similarity. It accepted a headline fragment sitting in a vendor's name field as not comparable, where the labeller scored it a disagreement on the grounds that the value is present and false.

What this does not show

Named gaps, not hedges.

Email or phone accuracy.

Not measured — and the axis most commonly cited in this category. It would need a separate deliverability study.

Pricing, latency, enrichment depth, intent data.

Out of scope. Nothing here speaks to what either vendor costs or how fast it answers.

Coverage outside these five ICPs.

Five deliberately chosen segments in four countries. Results generalise to similar segments, not to either vendor overall.

Either vendor's total database size.

Nothing in this design supports a claim about how many qualifying people exist.

Which vendor is correct where LinkedIn disagrees.

LinkedIn is self-reported and often stale. Where a vendor and the reference profile differ, the defensible statement is that they differ.

Languages beyond English and German.

Set 2 is the only non-English market. A wider benchmark would likely widen the gap on title handling.

No right of reply was conducted. Neither Apollo nor Clay was shown these results before publication, given an opportunity to correct factual errors, or asked to comment. That is a real gap, and it is disclosed here rather than described as a process. Scores were not, and will never be, altered in exchange for access, data, or commercial consideration.

Data and reproduction

Published under CC BY 4.0. Attribution required; commercial use, redistribution and tabular reproduction permitted. Every figure on this page is computed from the published file at build time rather than transcribed, so the page and the file you can download cannot disagree.

Recomputing these figures

Derived values must be computed the way the page computes them, or the numbers will not match:

• Agreement rate = agree / (agree + disagree). Never include not_comparable in the denominator.
• Valid rate = valid / returned.
• Share of union = by_source[vendor] / union_size. This is not recall.
• Compliance rate = compliant / returned.

Re-running ingest, resolution and reporting is free and deterministic. Re-running the reference scrape costs about $9.71 and would produce different results, because people change jobs.

Corrections. Errors found after publication are recorded with the original value preserved and visible. Figures are never silently amended.

Version

1.0

Run window

2026-09-10 to 2026-09-12

Measured

2026-09-12

Re-verification due

2026-12-12

Published

2026-09-12

Licence

CC BY 4.0

Attribution

unresolved