A contract manufacturer's test report arrives with one line that decides the award: "ICT test coverage: 95%." That single number is quoted in every RFQ response, repeated in every quality deck, and almost never defined. It is the most load-bearing number in a test quotation and the least standardised.
Two suppliers, the same board, the same month, can both report 95% and mean different things. One counted nets, one counted pins, and one quietly excluded the power planes before doing the arithmetic. On the board used as the worked example below, those three choices produce 88.4%, 94.7% and 97.1% — a 8.7-point spread on an identical design with an identical fixture. None of them is dishonest. All of them are unusable as a comparison until the denominator is written down.
This page is the audit method: what coverage is a percentage of, how to recompute a claim from the netlist in about half an hour, and the clause that stops the number drifting between quotation and first shipment.
What "ICT Test Coverage" Is Actually Measuring
In-circuit test verifies that the assembled board matches the design: components present, values within tolerance, nets connected where the netlist says they connect, and no shorts between nets that should be isolated. Coverage is the fraction of that opportunity set the fixture and program actually exercise.
Four different quantities get called "coverage", and they differ by an order of magnitude in what they prove:
| Unit Counted | What It Proves | Typical Reported Value | Weakness |
|---|---|---|---|
| Nets (logical connections) | Every intended connection was verified for opens and shorts | 92-97% | Treats a 200-pin ground net and a 2-pin signal net as equal |
| Pins / nodes (physical access points) | Every solder joint was probed or sensed | 85-93% | Falls sharply on fine-pitch and bottom-terminated parts |
| Components (devices verified) | Every placed part was measured or presence-checked | 88-96% | Misses connector and mechanical defects entirely |
| Testable opportunities (access-filtered) | Every net that could be probed was probed | 96-99.5% | Self-referential: the supplier defines the denominator |
The fourth row is where most inflated claims live. If a net has no probe access, is it an untested net (coverage goes down) or a non-testable net (coverage goes up)? Suppliers who answer "non-testable" report 99% routinely, and they are not lying — they are using a denominator they defined. The fix is never to argue about the number. It is to fix the denominator in writing before the quotation is compared.
Two things follow from this. First, coverage on its own is not a quality metric; a board can have 97% coverage and still ship with every connector unverified. Second, coverage is a design property as much as a test property: the probe access that determines the achievable number is fixed in layout, long before the fixture is quoted. That is why the design rules in our design for testability guide and the fixture economics in the ICT fixture NRE model have to be read together with the coverage arithmetic.
How Is ICT Test Coverage Percentage Calculated?
Short answer, for the search snippet: ICT test coverage is the number of nets (or pins) the fixture actually contacts and verifies, divided by the total number of nets (or pins) on the board, expressed as a percentage. The percentage is meaningless without the denominator: 95% of all nets, 95% of pins, and 95% of nets that have probe access are three different claims, and on a typical mixed-signal board they differ by 6 to 11 points.
The arithmetic itself is trivial:
coverage % = (verified units / total units) x 100
verified units = nets or pins with a probe, boundary-scan cell,
or functional observation that produces a pass/fail
total units = nets or pins on the netlist, minus any exclusion
the buyer has agreed to in writing
Everything that matters is in the last line. A supplier who applies no exclusions gets a low number and looks worse than they are; a supplier who applies generous exclusions gets a high number that does not describe the board. Neither is a solvable argument. The only workable approach is to demand the exclusion schedule alongside the number, then check both against the netlist.
Three Denominators, One Board: 88.4% vs 94.7% vs 97.1%
The board: a six-layer industrial control PCBA, 1,842 nets, 4,210 pins, one 484-ball BGA, two 0.5 mm QFNs, a 96-pin DIN connector, four press-fit headers, and 214 decoupling capacitors on shared power nets. The fixture: 1,410 probes, plus boundary scan on the BGA.
Here is the same fixture, honestly reported three ways.
| Basis | Numerator | Denominator | Coverage | What It Conceals |
|---|---|---|---|---|
| All pins | 4,210 pins; 1,410 probed + 624 BGA pins via boundary scan = 2,034 | 4,210 | 48.3% | Pins on ground nets are counted as untested even though a ground short would be caught |
| All nets | 1,842 nets; 1,629 verified by probe or scan | 1,842 | 88.4% | Treats the 214-caps-on-power-nets as fully verified when only the net, not each cap, is checked |
| Nets less agreed exclusions | 1,629 verified | 1,720 (planes, connector mechanical pins and three RF nets excluded by agreement) | 94.7% | Exclusions are legitimate in this case, but only because they were agreed in writing |
| Testable nets only | 1,629 verified | 1,677 (the supplier's own access-filtered count) | 97.1% | Silently reclassifies every no-probe net as non-testable |
Three lessons from the table, in order of how much money they save:
1. Pin coverage on any board with planes and BGAs is a misleadingly low number. 48.3% is technically true and it would fail every supplier benchmark, yet the fixture is a reasonable one. This is why suppliers stopped quoting pin coverage years ago; it is not, however, why buyers should stop asking what the number is measured against.
2. The gap between 94.7% and 97.1% is entirely exclusions. Forty-three nets, of which most are the piggyback capacitors. Agreeing to exclude what genuinely cannot be probed is normal engineering. Agreeing to exclude them after seeing that they would pull the number down is how a 94.7% board becomes a "97% coverage" line in a deck.
3. Nothing in the table measures whether the failures were caught. Coverage says a test exists. It says nothing about the test's detection threshold, its guard band, or whether a marginal joint passes. That is a different conversation, covered by the cost-per-defect-caught framing in ICT vs FCT coverage and cost.
The Five Exclusions Suppliers Take Without Saying So
In audit work on quoted coverage claims, the same five exclusions recur. Each is defensible on its own; taken together without disclosure they can move a reported number by 4-9 points.
| Exclusion | Why It Is Taken | Nets Typically Removed | What It Actually Hides |
|---|---|---|---|
| Power and ground planes | Planes cannot be probed; a short is caught indirectly by other nets | 2-12 | Split-plane and plane-isolation defects, which are real yield losses on high-layer-count boards |
| No-probe nets (no access) | Nothing can be done at fixture stage; the design already froze without a test point | 20-120 | The DFM conversation that should have happened. This is the largest and least visible exclusion |
| Ambiguity groups | Two or more nets sharing one node; a failure cannot be localised | 4-30 | Nets that are "tested" in the sense that a failure is detected somewhere, but not attributed |
| Connector and mechanical pins | Contacts are verified functionally or by mating, not by ICT | 8-96 | Every connector solder joint on the board, which is where a large share of field failures originate |
| Nets covered by functional test only | FCT exercises them at system level, so ICT does not duplicate | variable | Whether the FCT actually asserts on them. "Covered by FCT" is frequently a plan, not a test |
The second row is the one to focus on, because it is both the biggest and the only one the buyer still controls. Nets without probe access are a layout decision made months earlier. If a board arrives at the fixture stage with 120 no-access nets, the coverage ceiling is already set, and no amount of negotiation will lift it — only a design revision will. The test-point density rules that prevent it are in the DFT guide; the fixture cost consequence of fixing it late is in the fixture cost overview.
The fifth row deserves a harder line than it usually gets. If a net is excluded from ICT on the grounds that functional test covers it, the purchase order should name the FCT assertion. Without that, "covered by FCT" means covered by a test that may diagnose the symptom but never touch the net.
How Coverage Is Built: Probe Access, Boundary Scan, and What Each Buys
Three mechanisms generate coverage, and they have very different cost, lead time and rework profiles.
Bed-of-nails probing is the workhorse. Each node needs a test pad on the bottom side, 0.9-1.3 mm diameter in practice, with 1.5-2.5 mm clear of adjacent components and no via-in-pad. Coverage from probing is limited only by access. The constraint is mechanical: probe count drives fixture cost, roughly as modelled in the per-node NRE breakdown, and above about 1,800 probes the fixture mass and deflection start to become a design problem in their own right.
Boundary scan (JTAG) converts a large BGA from an untestable black box into thousands of virtual test points at zero fixture cost, provided the device supports IEEE 1149.1, the scan chain is daisy-chained through the board, and the nets between scannable devices are accessible in the interconnect model. On the worked example above, boundary scan on the 484-ball BGA contributed 624 pins of coverage — more than a third of everything the fixture verified, at no probe cost.
Functional observation catches what the other two cannot: power rails under load, communication protocols, sensor calibration, current draw signatures. It is also the weakest for coverage counting, because the observation is indirect. A rail that reads 3.31 V proves the regulator works and proves nothing about the twelve capacitors on it.
The strategic point: boundary scan is the cheapest coverage a design can buy, and it is a schematic-time decision. Designs that enable scan chains and route them properly routinely reach 95%+ coverage without exotic probe counts. Designs that skip it end up buying probes, or accepting a lower number.
Recomputing a Supplier's Coverage Claim in 30 Minutes
This is the audit a buyer can actually run, before the PO, using documents the supplier already has.
| Step | What To Ask For | What To Check | Red Flag |
|---|---|---|---|
| 1. Netlist count | Net count from the IPC-2581 or ODB++ fabrication data | Compare against the design's own net count | Count differs by more than a handful (renamed nets, merged planes) |
| 2. Fixture net list | The probe schedule: which net connects to which probe number | Count distinct nets with a probe | Supplier offers only the percentage, not the schedule |
| 3. Scan coverage report | Boundary-scan interconnect test report from the tool | Which nets the scan chain actually covers | "Boundary scan enabled" with no interconnect report |
| 4. Exclusion schedule | Written list of excluded nets and the reason for each | Every exclusion falls into the five categories, with a name against it | Exclusions appear only in the summary cell |
| 5. Recompute | — | (probed + scanned verified) / (total − agreed exclusions) | Result differs from the quoted figure by more than 1 point |
The audit takes about half an hour on a mid-size board, and it produces a number the buyer computed rather than one the supplier quoted. In practice the most common finding is not fraud but drift: a coverage figure quoted during RFQ, before the fixture was built, that was never updated against the delivered fixture's actual probe schedule. A design change between quotation and launch routinely moves coverage 3-6 points in either direction, and nobody re-reports it.
Where a claim fails the audit, the fix is almost always cheaper than it sounds: additional test pads on a spare bottom-side area, enabling a scan chain, or accepting a written exclusion with a named reason. All three are cheaper than discovering three years later that a connector joint class was never tested.
The Coverage Clause for Your PO
Four lines do more than a page of quality boilerplate. Each is deliberately narrow enough that a supplier can accept it without exposing themselves to an open-ended commitment.
| Clause | Wording Pattern | What It Prevents |
|---|---|---|
| Stated denominator | "Coverage is reported as verified nets divided by total netlist nets, excluding only nets listed in Schedule A." | A number that silently changes basis between suppliers or between revisions |
| Exclusion schedule | "Schedule A lists each excluded net with one of: plane, no access, ambiguity group, connector, FCT-only." | Bulk exclusions absorbed into a single summary figure |
| Ambiguity-group cap | "Nets in ambiguity groups shall not exceed 1.5% of total nets without written agreement." | The slow creep of ambiguity groups that quietly reduces attributable detection |
| Change notification | "Any ECO affecting test point access, netlist or BOM shall trigger re-reported coverage within 5 working days." | The RFQ-era number that was never refreshed after two design revisions |
The change-notification clause is the one most often omitted and the one that matters most on a program that runs for years. Coverage is not a property of the supplier; it is a property of a specific board revision tested on a specific fixture. Every revision breaks the equivalence, and only a re-report restores it. The same logic applies to the fixture itself, which has its own ownership and revision questions, covered in the fixture NRE and ownership analysis.
Finally, coverage is one gate in a chain. A board can clear ICT at 95% and still fail at AOI on solder volume, or at X-ray on voiding in the same joints. How the gates fit together, and which defect classes each one owns, is the subject of our AOI vs X-ray vs SPI comparison and the test strategy and coverage planning guide.