If HbA1c is a 90-day average, what does a 0.5-point fall actually mean in glucose?

My HbA1c went from 7.4% to 6.9% over four months on semaglutide. Everyone tells me half a point is good. Nobody has told me what half a point is, in units I can picture.

I understand the outline: haemoglobin gets glycated in proportion to ambient glucose, red cells live about four months, so the measurement integrates glucose over that window. What I do not understand:

• Is the average genuinely flat across 90 days, or is recent glucose weighted more heavily? If it is weighted, by how much? This matters because I want to know whether my 6.9% reflects the last four months or mostly the last four weeks.
• Is there a formula that converts an A1c into an average glucose I can compare against a meter reading? I have seen one quoted but not derived, and I do not know how much scatter there is around it.
• How long after a dose change should I wait before an A1c reflects it? I escalated at week 8 and the draw was at week 16, so I do not know what I am looking at.

I would rather do the arithmetic myself than keep asking, so a worked conversion would be more useful to me than a reassurance.

Adding the reference change value for HbA1c, because it bounds what a 0.5-point difference can support and it is a smaller number than people expect.

The criterion for a real difference between two measurements combines analytical and within-person biological variation:

RCV(95%) = 2.77 × sqrt(CVa² + CVi²)

HbA1c is unusually well behaved on both terms. Analytical CV on a certified platform is about 2%; within-subject biological CV is about 1.7%, which is remarkably low for any analyte.

• CVa² + CVi² = 4 + 2.89 = 6.89.
• sqrt(6.89) = 2.62.
• RCV = 2.77 × 2.62 = 7.3%.

Applied to a baseline of 7.4%: 7.4 × 0.073 = 0.54 points. So a fall of 0.5 points is just inside the noise band for a single pair of measurements, and a fall of 0.6 would clear it. That does not mean your result is meaningless — it means the correct statement is "probably real, and consistent with the direction the drug produces", not "definitively a 0.5-point improvement".

Two things make your case stronger than the bare arithmetic suggests. First, you have a mechanism and a prior: this drug class lowers A1c substantially, so you are not testing an arbitrary hypothesis. Second, if you have concurrent weight loss in the same window, you have a correlated second measurement pointing the same way. Two weakly significant concordant measurements are worth more than one, though not by as much as intuition suggests.

The general lesson worth extracting: HbA1c's RCV of ~7% is the best of any routinely available metabolic analyte. Compare triglycerides at roughly 59% or ALT at roughly 71%. When people complain that A1c is a crude measure, the comparison they should be making is not against an idealised continuous glucose record but against the other things on the same requisition form, all of which are far noisier. A1c is crude in what it measures and unusually precise in how it measures it.

One practical consequence: do not switch labs mid-series if you can help it. HbA1c is well standardised through the NGSP and IFCC networks, so between-lab differences are smaller than for most analytes, but they are not zero, and a 0.2-point method offset eats a third of your detectable change.

A footnote on what the 0.5 points is worth, since "good" was the word you were offered and it is not a unit.

Two separate translations exist and they are often conflated.

Diagnostic translation. Thresholds are 48 mmol/mol (6.5%) for diabetes and 42 mmol/mol (6.0%) for the upper end of normal in most systems, with the intervening band variously named. Your move from 57 to 52 mmol/mol keeps you above the diagnostic threshold, so nothing about your classification has changed. That is worth knowing because people read a fall as a reclassification and are then surprised that their care has not changed.

Outcome translation. This is the one people actually want, and it is where the honest answer gets uncomfortable. Observational data relate higher A1c to more microvascular and macrovascular events, and it is tempting to read a 0.5-point fall off that curve as a proportional risk reduction. But the intervention trials that lowered A1c aggressively by other means did not deliver the macrovascular benefit the epidemiology predicted, and one of them found harm. So A1c is a good marker of glycaemic state and an unreliable surrogate for the benefit of lowering it — the benefit depends heavily on what you lowered it with.

That last point is why the outcome trials in this drug class matter more than the A1c numbers do. In SUSTAIN-6 semaglutide reduced major adverse cardiovascular events with a hazard ratio of 0.74 in people with type 2 diabetes at high cardiovascular risk [1], and in FLOW it slowed kidney disease progression at a dose that produces only modest weight loss [2]. Those are outcome measurements. Your A1c fall is consistent with being on a therapy that has them; it is not itself the benefit.

Practical version: track A1c because it is cheap, precise and tells you whether the glycaemic effect is present. Do not convert it into a risk percentage in your head, and do not let a plateaued A1c persuade you the therapy has stopped doing anything, because the outcome effects in this class are not proportional to the glycaemic effect. Discuss the actual targets with a clinician who knows your cardiovascular and renal status.