You build a clean calibration curve in dilute acid, prove its linearity, run your sample, and read a number off the line — and it is off. Not because the curve was bad, but because your sample is not dilute acid. It is a digested lube oil, a soil extract, a serum, a brine, and that matrix is changing the analyte's signal in a way your pristine standards never saw. An external curve assumes the sample and the standards respond identically; a matrix effect is exactly the case where they do not. The first moves keep you on a clean curve: matrix-matched standards that respond like the sample, or an internal standard that rides the matrix out — in plasma work, the usual first line. Only when the matrix cannot be reproduced in a standard and no internal standard tracks it do you stop bringing the sample to the calibration and bring the calibration to the sample. That is standard addition.
The idea in one line
Take the sample itself, split it into equal aliquots, and spike each with a different known amount of analyte — zero, then a little, then more — bringing every aliquot to the same final volume, so the sample is diluted identically in all of them. Measure each. Because the matrix is present, and equally diluted, in every measurement, its effect sits inside every point. Plot response against the concentration you added and you get a line: R = m·Cadd + b. The analyte already in the sample shows up as the intercept b — the response with nothing added — and the slope m is the sensitivity in this matrix. Divide one by the other and you have what was there: C₀ = b / m.
That one line of algebra rests on a short list: the response is net (any blank/baseline already subtracted); it is linear across the spiked range; there is a single multiplicative sensitivity m that the matrix scales but the spike does not change; the added analyte is the same chemical species as the native one; and any additive background has already been removed. Break any of these and the intercept moves. Most of what follows is just those assumptions, examined.
Why it works
The trick is that the analyte you spike in and the analyte already there are the same molecule in the same matrix. Whatever the matrix does to the sensitivity — suppress it, enhance it — it does to both, through the same m. So the slope of your response-versus-added line already carries the matrix effect inside it; you never had to measure that effect or reproduce it in a blank, because you never left the matrix. The external curve's fatal assumption — that standards and sample share a sensitivity — is made true by construction, because the "standards" here are the sample, dosed.
That is also why the answer lives at a negative added amount. With nothing added the instrument still reads the native content; to drive the response to zero you would have to remove analyte — add a negative amount — and the size of that hypothetical removal is exactly what was there. The x-intercept is the geometry of "how much would I have to take out to reach nothing."
One spike, or several
The sound version uses several additions — the un-spiked sample plus three or more increments — so you fit a real line, confirm it is straight across the range, and get an intercept with a known uncertainty. A single-point addition is faster and common in routine work: it gives C₀ = Cadd·R₀ / (Rspiked − R₀), but it assumes linearity rather than testing it. Choosing the additions is not about bracketing the answer — the answer always sits at −C₀, off to the left of every point you measure; you never straddle it. It comes down to three practical rules: keep every addition inside the linear range; make the largest addition roughly one to three times the native level, so the extrapolation back to −C₀ stays short and the added signal is large enough to measure well; and keep spike volumes small (or the total volume constant), so you are not diluting the matrix as you climb your own line — dilute the matrix and you have changed m partway up.
Standard addition cancels a multiplicative matrix effect: a matrix that changes the sensitivity (the slope), scaling the response up or down. It does not cancel an additive interference — a background or overlap that contributes signal regardless of how much analyte is present. An additive term lifts the whole line vertically, moves the intercept, and biases the result, and no amount of spiking removes it. In ICP that additive case has concrete names — a spectral line overlapping yours in optical emission, a polyatomic ion landing at your mass in mass spec — and standard addition is blind to all of them. Additive interferences are a separate job (a proper blank, background correction, better selectivity), done before or alongside, never by the addition itself.
The price you pay
Standard addition is not free, and a good analyst reaches for it because they must.
It is an extrapolation. You read the answer off the line beyond your measured points, and the uncertainty of an intercept b/m is not the tidy scatter of a point sitting among standards — it widens as the slope gets shallow and as the crossing falls farther from the centre of your added levels. So a standard-addition result is usually less precise than a reading off a good external curve — though far more accurate than a biased one. Precision is what you trade, not accuracy; more additions and honest replication buy some of it back.2
It costs sample and time. Every result now needs several prepared aliquots of that one sample, not one injection against a shared curve — a real burden across a long run, which is why standard addition is a targeted tool, not a default.
The bookkeeping is unforgiving. Two clean patterns keep it honest: put equal sample aliquots into a set of flasks, add increasing volumes of a spike stock, and make every flask up to the same final volume — the sample is diluted identically everywhere, m stays constant, and the intercept is C₀ in that final solution (multiply by Vfinal/Valiquot for the original); or add tiny volumes of a concentrated spike straight in — you still compute the added concentration (Cstock·Vspike/Valiquot), but the volume is small enough to skip the matrix-dilution correction. What you must not do is add large, varying volumes of a dilute spike — that dilutes the matrix by a different amount at each point and quietly breaks the constant-m premise. The spike must also match the analyte's species and be given time to equilibrate: an aqueous metal spike dropped into an incompletely digested oil, or a redox-active element like chromium or mercury added in the wrong form, will not behave like the native analyte, and the method fails without complaint.
A quiet diagnostic: the matrix factor
Run a standard-addition line and a clean external curve for the same analyte and compare their slopes. If the matrix were inert they would match; when they do not, the ratio mmatrix / mclean is the matrix factor — the sensitivity remaining in the matrix relative to a clean standard, below 1 for suppression and above 1 for enhancement. Two honesties about it. It is a multiplicative matrix factor, not a spike-recovery (which also folds in preparation losses — a different experiment). And it is not an independent confirmation of your result: for a linear system the external reading and the addition result differ by exactly this same ratio — one fact stated twice, not a cross-check, and that holds even when an additive offset is present (the offset biases both the same way, which is exactly why standard addition is blind to it). Where the slope ratio and the two readings genuinely diverge, the cause is elsewhere — the response is curved, or the un-spiked external reading and the un-spiked addition aliquot were not the same thing.
Standard addition, or internal standard?
Both fight the matrix, and for a plasma-spectrometry audience the order matters. In ICP the internal standard is the routine, primary control for multiplicative matrix effects — nebulisation efficiency, plasma loading, ionisation, drift — because a well-chosen reference element tracks those and the ratio divides them out on every sample, riding along with an ordinary calibration. Standard addition is the heavier, occasional method for what the internal standard cannot track: an analyte-specific sensitivity change, or a technique with no good internal standard at all. They are not interchangeable and often work together — when they do, you plot the analyte-to-internal-standard ratio against added concentration, so the addition line inherits the internal standard's correction.
Pedagogical, not a method: real oil wear-metal work more often dilutes in solvent (ASTM D5185) than acid-digests, and a clean aqueous curve confounds acid, carbon, and dissolved solids with the true matrix. Treat the numbers as illustration.
| Added (ppm in solution) | Net response (a.u.) |
|---|---|
| 0 | 5.33 |
| 5 | 8.58 |
| 10 | 11.83 |
| 15 | 15.08 |
The points fall on a line of slope 0.65 per ppm and intercept 5.33, so C₀ = 5.33 / 0.65 = 8.2 ppm in the measured solution (then × Vfinal/Valiquot for the oil). The same digest read against a clean aqueous curve gave 6.3 ppm — the matrix factor 6.3/8.2 = 0.77 says the aqueous curve saw only ~77% of the sensitivity, reading ~23% low. That 0.77 equals the slope ratio; it is the matrix quantified, not a second, independent check. And note what it still cannot touch — a spectral overlap or a polyatomic at your mass would be additive, and would need its own correction.
The judgement
Standard addition answers one specific, honest admission: I cannot build a standard that responds like my sample. When that is true — a complex, variable, or one-of-a-kind matrix — it is often the only route to a defensible number, and if you also ran an external curve it hands you the matrix factor as a by-product. When it is not true, matrix-matched standards or an internal standard riding an ordinary curve are faster, more precise, and less sample-hungry. Knowing which situation you are in — and being honest that a matrix effect is present, rather than trusting the clean curve because it is convenient — is the whole of the skill. The method cannot decide that for you; it only rewards you for deciding correctly.
A foundations piece — the reasoning behind standard addition, not a validation protocol. For spike levels, acceptance criteria, and where the method of standard additions is mandated, work from the relevant EPA or ASTM method and your laboratory's quality system.