The long tail advantage is 2.27x, and narrower inside a position band

Chase the long tail. It is the most durable advice in search, and we believed a strong version of it: our own notes recorded a head term converting at 0.06 percent against a specific variant of the same subject at 7.50 percent, a gap of roughly 125 times.

This is the check of that claim. The advantage is real and it is far smaller than that, and getting to the honest number took two attempts, because our first correction made an error of its own.

What it looks like before you control anything

One cluster, one window, 3,328 queries. Band them by how much volume each query draws and the pattern is exactly the one everybody describes.

Read end to end that is a 5.3x advantage, and published as it stands it is a strong argument for chasing the tail.

THAT 5.3x IS NOT THE NUMBER THIS FINDING TESTS, and an earlier version of this page treated it as though it were. It compares the single query above 100,000 impressions against everything under 10. The position-controlled table further down uses wider cohorts, 1,000+ against under 100, so setting 5.3x beside it measured two changes at once and credited both to position. On the cohorts actually used below, the uncontrolled gap is 2.27x: 0.1103 percent for queries above 1,000 impressions against 0.2499 percent for those under 100. That is the baseline the control has to beat.

The column that ruins it

Read the position column. It moves with volume throughout: 8.96 at the head, 5.90 in the deep tail. On the two cohorts this finding uses, the head averages 8.75 and the tail 7.67, so the rarer queries also sit about one position higher up the page. Two things are moving and the table credits one of them.

Holding position still

Cross-band the same queries by position and by volume together. Topic, template, site and window are constant throughout, because this is a single cluster.

The same cohorts that give 2.27x across the cluster give 1.57x at positions 5 to 8 and 1.23x at 8 to 11. Both within-band figures are smaller than the overall one, and both are smaller than the headline anyone would have taken from the first table.

No further gain appeared past the first step

Look at the step from mid to tail rather than from head to tail. At positions 5 to 8 it is 1.00x and at 8 to 11 it is 0.86x, so in these cells the deep tail did not convert better than the merely specific.

READ THAT AS AN ABSENCE OF EVIDENCE, NOT EVIDENCE OF ABSENCE. Those two cells carry 31 and 33 clicks. At that size the ratios are point estimates with intervals wide enough to contain a real advantage and a real penalty alike, so they do not establish equivalence and certainly do not establish that the deep tail is worse. An earlier version of this page said going further "buys nothing", which claimed more than 64 clicks can carry.

What can be said is narrower and still useful: the advantage that is clearly visible is earned in the first step away from the head term, and we did not observe it continuing. Since the cost of chasing queries rises steeply as they get rarer, the burden of proof sits with the deep tail, and this data does not discharge it.

What is left standing

Something real, and much smaller than advertised. Within one position band, moving from the head term to specific variants is worth roughly 40 to 60 percent more clicks per impression. On a cluster this size that is worth having and it is not a strategy.

The other thing in the data, and it is an observation rather than a mechanism: the rarer queries also rank better. Head and tail sit at 8.75 and 7.67 across the cluster, and the difference persists inside the bands. Whether ranking is what produces the click-rate gap, or both follow from something else about specific queries, is not something this comparison can separate. Doing so needs an adjustment model fitted to the position distributions, and none was fitted here.

The claim we withdrew

Our own record held a head term at 0.06 percent against a specific variant at 7.50 percent, about 125 times. On this window the best query above 100 impressions reaches 2.83 percent against a head term at 0.105 percent, which is 27 times, and it rests on three clicks.

The 125x does not reproduce. It came from a different window, it was never controlled for position, and a rate computed from a handful of events is the thing this site exists to refuse. It should not be requoted, including by us.

The limit that cuts against this finding

This export contains 636,317 impressions against 833,763 counted at page level for the same cluster and window, so roughly 24 percent of it is not returned.

What that 24 percent consists of, we do not know. Search Console anonymises rare queries, and it also documents additional loss when results are grouped by query rather than reported in aggregate. An earlier version of this page asserted the gap was "by definition the rarest queries", which assumed a composition we never measured and which happened to be the one most convenient to the argument.

The honest position: the deep tail measured here is the tail the instrument returned, and a quarter of the cluster sits outside it for reasons we cannot decompose. That is a limit on the conclusion in both directions.

How to check your own

Pull queries filtered to a single cluster, so topic and template stay fixed. Band them by volume and look at the position column before anything else. If position moves with volume, and it usually will, the uncontrolled comparison is measuring both and crediting one.

Then cross-band by position and volume together and read the step from mid to tail rather than head to tail. That step is where the advice either holds or stops, and on this cluster it stopped.