aidos.tech Transparency on Google reviews
Methodology

How aidos surveys, computes and publishes

Every figure on this site must be traceable to a publicly visible statement. This page explains where the data comes from, what we compute from it — and what we deliberately do not do.

1. The data source

The basis is the Google Maps transparency notice: since 2026, Google states on public business profiles how many reviews were removed in the past 365 days “due to complaints about defamation”. aidos records exclusively what is visible to everyone there:

Not recorded: the content or authors of individual reviews — neither visible nor removed ones.

2. The survey

  1. Candidates: business profiles are systematically compiled from open sources (incl. OpenStreetMap) — city by city, across industries.
  2. Checking: every profile is checked for the transparency notice. Profiles without a notice are counted too — that is the only way to say credibly how widespread removals really are (a measured denominator instead of anecdote).
  3. Monthly repetition: the survey re-runs monthly. Every entry carries a survey date; entries that are not re-confirmed disappear from the public view.

On reach: the notice is rare — in the cities surveyed so far, only around 3–5 % of checked profiles carry it. Since July 2026 the survey logs every checked URL with its outcome; the resulting measured rate per city is shown in the data-basis overview.

3. What we compute

Ranges, never false precision

Google publishes ranges, not exact figures. We compute with these ranges and report them as such. For “over 250” the true number is unknown — there we compute with 250 as a lower bound and mark everything beyond it as open-ended, rather than inventing a ceiling.

The “rating without removals” estimate

For sufficiently documented profiles we estimate where the rating would stand if the removed reviews still counted. The computation is a single line: the removed reviews are put back into the displayed average.

ρ′ = (S + R · a) / (N + R)
ρ = displayed rating · N = displayed review count · S = ρ · N
R = number of removed reviews (from Google) · a = their assumed star value

Everything except R and a is public on the Google profile. Those two are unknown — but unknown to very different degrees, and that is the heart of the method.

a — the star value of the removed reviews

The removed reviews are gone, so their star value is fundamentally unobservable. Removal follows a defamation complaint, which makes them negative by their nature; we set the support to 1–2★.

The mean within that support is measured, not asserted. The only empirical anchor is the negative reviews that survived at the very same businesses — across all surveyed profiles with a star distribution:

StarsShare of surviving reviews
1★2.1 %
2★1.1 %
3★4.4 %
4★18.0 %
5★74.4 %

1★ outnumber 2★ by 1.99 : 1 (median per business: exactly 2 : 1), which gives a mean of about 1.34★; the constant we compute with is 1.335★. The measurement is repeated after every survey — so far it stays between 1.33 and 1.34★, which is why the constant does not drift from month to month. 3★ are deliberately excluded: they make up 59 % of the whole 1–3★ pool, would pull the mean to 2.3★ and roughly halve the reported effect — and a middling 3★ review is essentially never removed as defamatory.

R — where the uncertainty actually sits

We measured how much each unknown really moves the result. Sweeping a across its entire 1★–2★ support changes the reported effect by only about 0.1★ at the median. Google’s range for R moves it more than that for 93 % of profiles. The argument about star values is therefore the smaller question; the spread shown on the site comes overwhelmingly from Google writing “151 to 200” instead of a number.

So we pair the ends of the corridor to make it a genuine envelope rather than decoration:

upper edge = fewest removals (Rmin) × mildest assumption (2★)minimum effect
lower edge = most removals (Rmax) × harshest assumption (1★)maximum effect

The upper edge is what we report, as an “at least”. It assumes the lowest count Google names and the mildest rating at the same time, so it cannot be dismissed as having been computed too harshly.

“Over 250” — a corridor with no floor

For the most conspicuous profiles Google publishes only a lower bound. For them no upper edge of R exists, and therefore no lower edge of the corridor. We do not substitute a stand-in value; we draw the corridor open-ended and say “at most X★”. A stand-in would have the opposite effect: it would make precisely the most heavily affected profiles look the safest.

Why the estimate is systematically too low

Google’s figure is a rolling 365-day total. Anything older drops out of it — while still weighing on the displayed overall average, which is all-time. The reported effect is therefore a floor for the total distortion, not a measure of it.

Continuous monthly measurement can raise that floor — but not by addition. Two readings a month apart overlap by eleven twelfths: “151 to 200” in July and again in August largely describe the same removals, not 400. Three bounds are defensible, and we use the largest:

1 · the largest window ever read — every window is a subset of the total
2 · first window + every observed increase — from Wt − Wt−1 = r(t) − r(t−12) and r ≥ 0 it follows that r(t) ≥ Wt − Wt−1
3 · sum of readings ≥ 12 months apart — their windows do not overlap and count in full

Within the bands we use the lower edge for increases and the upper edge for deductions; a capped “over 250” contributes nothing and can therefore never inflate the figure. Across a simulation of 5,675 trajectories this bound never exceeded the true count and understated it by 42 % on average — it is a floor, not an estimate. It appears on a profile page only once it says more than Google’s current window.

All of it remains explicitly an estimate: if the removals were justified — a fake campaign, say — the displayed rating is the more accurate one. The computation says how much the display depends on the removals, not whether they were warranted.

The aidos score

The score (0–100) is a percentile of the minimum effect: it ranks a profile by how far the removals demonstrably lift its displayed rating. A score of 90 means “more affected than 90 % of surveyed profiles” — a neutral statistical classification, not a value judgement and no proof of misconduct.

Until August 2026 the score counted the number of removed reviews. That mostly measured business size: a restaurant with 8,000 visible reviews and “over 250” removals ranked near the top although its rating rises by less than 0.1★, while a business with 113 reviews and 101–150 removals — whose rating is flattered by at least 1.3★ — did not register at all. So we sort by the effect. The count Google publishes still appears in every row; it is the raw finding.

The monthly comparison — why we count businesses, not reviews

Because Google publishes ranges only, two surveys cannot yield a count of removed reviews. Subtracting one month's range midpoints from another's measures the width of the ranges, not what happened: a single business moving from “151 to 200” into “201 to 250” arithmetically produces “+50” when the actual increase may be 1. Conversely, any change inside a range stays invisible. Such a difference is neither an upper nor a lower bound but a computational artefact — we do not publish it.

What we report instead is the smallest reliably measurable event: how many businesses crossed a range boundary, and in which direction. That is a hard fact from Google's own display. “Unchanged” explicitly does not mean “no removals” — with a rolling 365-day total it can equally mean that new removals and expiring older ones cancel each other out.

4. Whom we name — and whom we don't

Names are published only for legal entities and chains. Only corporations (GmbH, AG …), larger chains and high-revenue businesses appear publicly by name. Sole traders and micro businesses enter the statistics only anonymised in aggregates (industry, city, count) and are stored pseudonymised internally. Details in the privacy policy (DE).

5. Limits of the data

6. Corrections

Mistakes happen — and get corrected. Affected businesses can contest their data via report / correct data (DE) or contact@aidos.tech. If a review is not completed within 5 days, we precautionarily remove the entry from public view. Substantive corrections are noted on the affected page.