Same value, same odds. Not the same streets.

Start with the county's own geography, which we published in the second article in this series. Line Fulton's cities up by over-assessment rate and the county stops looking like one place. In East Point, 31.2% of the homes we can confidently value are over-assessed. In Roswell, 7.3%. Same tax authority, same office, same year, more than four times the odds.

Now put a second thing on that same chart. The cities where over-assessment is worst are the county's Black neighborhoods, and the mildest are its white ones. That isn't a subtle signal buried in the data — it's the shape of the chart, and anyone who knows Atlanta suspected it before opening it. It is not a clean line, either: Milton is a largely white city carrying an above-county rate, and it sits in the scatter where you can see it (Figure 1).

It holds up at a finer grain, which is what makes it worth reporting — cities are big, and a correlation that only appears when you average across whole municipalities usually isn't one. Group Fulton's census tracts⁶ by their share of Black residents instead, and the rate is flat across the two majority-white groups, steps up sharply above the 50% mark, and then holds — a raw gap of 7.3 percentage points between majority-Black and majority-white tracts (Figure 2). Measured a second way — on city boundaries, where our address matching is complete, rather than on tracts, where it isn't — the gap comes out wider still. The number we're publishing is the conservative one.⁷

Rates alone would make this sound tidier than it is, so here is the size of the thing. Roughly a quarter of Fulton's homes sit in five south-Fulton cities — East Point, South Fulton, Union City, Fairburn and College Park. Those five hold roughly a third of all the over-assessment in the county.

That is a real finding. It is also not the finding most people will assume it is.

Here is the test that matters, and it's a simple one.

Take every home we can value and ask whether being in a majority-Black tract predicts being over-assessed. It does: +7.3 points, exactly as above. Now add a single control — the home's own market value — and ask again. The 7.3-point gap doesn't shrink. It disappears, landing at −1.18 points with a p-value of 0.44, which is a statistician's way of saying there is nothing there to see. Specified continuously against tract share Black rather than a majority cutoff, same answer. We ran it on two separate versions of the county roll a week apart, one of them three thousand over-assessed homes larger than the other, and the result barely moved (Figure 3).⁸

What that means in plain terms: two homes of the same value, one in East Point and one in Roswell, have statistically indistinguishable odds of being over-assessed. There is no measured surcharge for the house being in a Black neighborhood. If there were one, this test is where it would show up, and it doesn't.

That result cuts both ways. It forbids us from saying Fulton County's assessments penalize Black homeowners for being Black — we looked, with the best instrument we have, and we don't find it. It equally forbids us from calling the outcome race-neutral, because the 7.3-point gap from the previous section is still sitting there, still true, still landing on actual households who pay actual bills. Both sentences hold at once, and an account that drops either one is a worse account.

That result also does more than rule something out. If a gap vanishes the moment you account for what the homes are worth, then what the homes are worth is what the gap was made of all along. Which points at the real question.

This is the part I think matters most, because “the overcharge lands in the middle of the market” is an outcome, not an explanation. Without a mechanism behind it, there's no reason to expect it to hold next year, or in another county, or under a different way of measuring.

Recall what the county is actually doing. It has to put a value on every home in Fulton — hundreds of thousands of them, every year, without visiting them. So it fits a model: recent sales, plus a handful of recorded facts about each property, plus a market-area factor, producing a number for every parcel whether or not the evidence for that parcel is any good. That model is graded on how well its values track sales across a group of homes, not home by home.

Two properties of that process, taken together, determine where its errors land.

Fitted models compress. Any model built from imperfect predictors produces estimates less spread out than reality. It has to: the part of a home's price it can't explain gets absorbed into the average rather than assigned to the home. So within any market area, the model's range of values is narrower than the actual range of prices. Homes whose true worth sits above their area's center get pulled down, and homes whose true worth sits below it get pushed up. Those second ones are over-assessments. Not mistakes — the arithmetic of being right on average.

Compression bites hardest where the model is confident and the houses aren't alike. This is the crux. Compression needs two conditions to produce a lot of over-assessment: enough sales for the model to commit to a specific number, and enough real variation among the homes it's committing about. At the top of the market you get the second without the first — luxury homes are close to unique and trade rarely, so the comparable sales that would bracket one cleanly, one larger and one smaller, grow scarce, and the model is reaching. So are we: an over-assessment finding up there rests on thinner evidence than the same finding does in the heart of the market, where similar sales are dense. The high rate at the top is real, and it is also softer and higher-variance than the middle's. At the bottom, below roughly $200,000, the county's values sit so far under what sales support that even a compressed estimate lands low: only 3.5% of those homes come in over-assessed, and they are just 2.3% of all the over-assessment in the county. The bottom of the roll is not where this problem lives — which is worth saying carefully, because it is not the same claim as “assessment is not regressive in Fulton County,” and it does not contradict the national work finding that it is.¹¹

The middle is where both conditions hold at once. In the $250,000–$500,000 stretch there are plenty of arm's-length sales — enough that the county's model produces a confident number and enough that we can check it against genuinely comparable homes. And the homes themselves are heterogeneous in ways the county's records capture poorly: a renovated 1962 ranch three doors down from an untouched one, a 1998 subdivision backing onto a 2021 infill build, a lot on the quiet side of a street and one on the arterial. Same square footage, same bed-and-bath count, same neighborhood code, materially different prices. The model can't see the difference, so it assigns the average, and it assigns it with confidence. That band is where a fitted surface has the most to be wrong about and the least excuse for hedging — and it shows up exactly where the mechanism predicts, a hump of over-assessment in the middle of the market, with the $300,000–$400,000 band alone accounting for more than a fifth of every over-assessed home in Fulton County (Figure 4).

None of this involves race, and none of it could. The county's model has no race variable in it; neither does ours. Both are built from property characteristics, location, and sales. The process is blind in the literal sense.

But it is not blind to price. It has a specific, structural, predictable error concentrated in a specific band of the market. And which homes sit in that band is not random at all.

In Fulton County, the $250,000–$500,000 range is disproportionately where the housing stock of the majority-Black neighborhoods sits. Nearly half of it — 46.3% — against 19.2% of the stock in majority-white tracts (Figure 6). That is not a claim about anyone's finances or choices; it's a description of a housing stock, produced by a century of decisions about where houses got built, who was permitted to buy them, and which neighborhoods appreciated when. East Point, South Fulton, Union City, Fairburn, College Park — this is the heart of their market. It is also the exact band where a mass-appraisal model, doing its job correctly, errs high most often.

That is the whole mechanism. A race-blind process with a price-shaped error, operating on a housing stock where price and race are correlated, produces a racial pattern in its outcomes without containing anything about race. It isn't a penalty. It's a pattern. The statistical test in section 02 is the proof: the entire 7.3-point gap is accounted for by the value of the homes. Nothing is left over for race to explain, because the value distribution already explains it.

Two things follow.

The first is that “no penalty” and “no problem” are different findings. A process can be perfectly even-handed in its rules and still deliver its errors unevenly to people, and this one does. Nobody in the Fulton County Board of Assessors decided this. It is nonetheless happening, in measurable quantity, to identifiable households.

The second is a question we are deliberately not answering. Home value is not a neutral variable in American housing — the reason Fulton's Black neighborhoods hold the stock they hold, at the prices they hold it at, is itself a history, and controlling for home value controls that history away. Whether that makes home value the wrong thing to control for is a serious question. It is not our question, and we don't have the standing or the evidence to settle it. We're reporting what the assessment data shows: no same-value racial surcharge, and a real racial gap that runs entirely through the value of the houses. We assert neither a causal racial penalty nor an all-clear.

Our finding also isn't the last word against the national literature, and it doesn't read as a contradiction of it. The best-known work on this question finds a substantial assessment gap by race while holding taxing jurisdictions and tax rates fixed, and finds that just over half of it arises between neighborhoods rather than within them — driven in part by assessments being less responsive to neighborhood attributes than market prices are.¹⁰ That is section 03's mechanism, reached independently and by a different route. We work in one county, in one year, against our own comparable-sales estimate rather than sale-price ratios, and our numbers are not their numbers — but what each of us ends up pointing at is the same.

One more cut, and the one most easily overstated.

Sort Fulton's homes by the median income of their census tract and over-assessment is more common toward the lower end of the range than the upper: the bottom five deciles average 19.0% against 13.2% across the top five. But the shape is not the one the story wants. The worst-hit decile isn't the poorest — it sits a little above it, and the lowest-income tenth of the county comes in close to the county rate rather than at the top of the range (Figure 5). So “the poorer you are, the likelier you're overcharged” is not quite what this data says, and we're not going to claim it.

What the data does show is sharper, and it isn't about how often. It's that the same overcharge is not the same burden.

The typical over-assessed Fulton home pays about $804 a year above what recent sales support. In the lowest-income tenth of the county — where the representative household is making around $38,000 — the typical over-assessed home pays about $590 a year, a smaller sum but 1.55% of income, roughly twice what the same overcharge comes to in the middle of the range.⁹ A smaller dollar figure, a heavier bill. How often the error happens and how much it costs are two different questions, and they do not have the same answer.

That number needs its caveat alongside it, because the denominator is doing real work. We have no household income; nobody does, at the parcel level. What we have is the median income of the tract a home sits in, and the lowest-income tracts are exactly where that proxy is worst — full of renters whose income is counted while their landlord's tax bill isn't, retirees on modest incomes in homes they finished paying for decades ago, and students. Read 1.55% as a signal about the shape of the burden, not as a household budget line.

Three claims. Over-assessment in Fulton County is geographically concentrated, and where it concentrates correlates strongly with race. That concentration runs through what the homes are worth — it sits in the middle of the market, where a mass-appraisal model is at once most confident and most wrong, and Fulton's Black neighborhoods hold a disproportionate share of their homes there. And the dollars land hardest, as a share of what a household has, at the bottom of the income range.

Two refusals. We do not claim that Fulton County assesses Black homeowners' homes higher than comparable white-owned homes; we tested it directly and the effect is statistically indistinguishable from zero. And we do not claim that a process blind to race produces an outcome neutral on race. This one doesn't.

And one thing a pattern like this can't do. It describes a county from above, and nobody lives in a county from above. That over-assessment concentrates in a particular band of the market, or in a particular set of neighborhoods, says nothing certain about the house anyone is standing in — there are correctly valued homes in the worst-hit neighborhoods and over-assessed ones in the mildest. Distributions don't settle individual cases. An overcharge still has to be shown one house at a time, against the recent sales of the specific homes nearest it. The pattern says where to look. It never says what you'll find.