/ how honest the numbers are

How honest the numbers are

Every measurement wobbles. We write down how much.

Why a single week means nothing

Ask the same AI the same thing three times on one day and the result flips in almost 9 per cent of cases. Between the four services the spread is five times larger: on 43 per cent of our 991 measurement days they disagree about who deserves a mention. Check one service and you check the wrong one.

So every figure here carries a range. When the range is nearly as wide as the figure itself, the cockpit tells you: that means there are still too few measurements to say anything. We do not read single weeks, we read trends across weeks.

What that cost us in August

Until 19 August, three repetitions of the same question counted here as three independent measurements. That was too generous and made our ranges artificially narrow. Since then they count as one.

The result: our ranges got wider and our citation rate got lower. Same measurement, more honest number. We made our own figures look worse because the old arithmetic was wrong, and we wrote to our customers about it the same day.

We also drop recommendations

ECHO used to recommend publishing an llms.txt file. Then the evidence arrived: Google officially ignores the file, 97 per cent of these files are never fetched, and across more than 300,000 domains there is no measurable effect.

We dropped the recommendation, from the tool and from our advice. An agency that actively sells less when the evidence is missing has an argument no feature list beats.