Inside one position band, click-through varied nearly nine times over

Almost every traffic forecast in search rests on a published click-through-by-position curve. Reach position three and expect some percentage; slip to position eight and expect a smaller one. The curves differ between vendors but they agree on the shape, and the shape is what gets used: it turns a ranking target into a traffic promise.

On this property the shape does not hold. It does not hold weakly, it inverts.

The whole site, bucketed by position

The two emphasised rows are the ones worth arguing from. The 1 to 3 band is 99 impressions across 61 pages and proves nothing in either direction; it is printed because leaving it out would be a choice about which data the reader gets to see. URLs are canonicalised before counting, and every share uses the sum of these page rows as its denominator rather than the property day chart.

Pages averaging position 10 to 20 convert at 0.580 percent. Pages averaging 5 to 10, five ranks better, convert at 0.439 percent. Being further down the page is associated with more clicks per impression, not fewer.

Holding position still

The obvious objection is that the bands contain different pages, so something other than position is moving. That is the finding, and it can be shown inside a single band.

The 5 to 10 band carries 95.0 percent of every impression this site earned. Split it by whether the page answers a question that resolves to a single value.

Same position band, same site, same window, same template. Click-through differs by 8.7 times.

POSITION IS NOT FULLY HELD CONSTANT HERE, and an earlier version of this page wrongly said it was. A five-rank-wide band is not a fixed position: inside it the single-value group averages 7.65 and the comparison group 6.69, a gap of 0.96 of a position. So position is not eliminated as a contributor, only narrowed. What can be said is that the residual difference runs AGAINST the single-value group, which ranks nearly a position worse. How much of the 8.7x that residual could account for is NOT established here: a difference in group means does not bound the effect of the underlying position distributions, and separating the two properly needs finer matching than a band, or a model fitted to those distributions.

Why the bands invert

Because the single-value clusters all live in the 5 to 10 band. They are 80.7 percent of it, they convert at 0.177 percent, and they drag the band average down to 0.439.

The 10 to 20 band contains none of them. Not a small share: zero percent. It is made of ordinary pages answering questions that need more than one value, which is why a worse position produces a better rate.

The inversion is not a paradox and it is not an anomaly of this site. It is what happens whenever a position band is not homogeneous in query shape. A curve that maps position to click-through has no term for query shape by construction, so it cannot distinguish the two populations inside this band - that is a property of the form, not a survey of what has been published.

What this breaks

Any forecast of the form "we currently rank eight, moving to four is worth this much traffic". That arithmetic multiplies current impressions by a curve value, and on this site the curve value is wrong by up to an order of magnitude depending on which page you picked, in a direction the curve cannot indicate.

It also breaks the reverse inference, which is more common and does more damage: seeing a low click-through rate and concluding the page is underperforming its position. Two thirds of this site is doing exactly what its queries permit.

What we are not claiming

Not that position does not matter. Within one query family it plainly does, and nothing here measures that, because these buckets mix families.

Not that published curves are wrong in general. They describe an average across sites whose query mix is not ours, and the failure reported here is a failure of TRANSFER, not of arithmetic.

And page-level average position is a blend. A page reported at 7 may sit at 3 for some queries and 15 for others, so these buckets are softer than they look. That is a real limit, and it is also the number the interface hands you and the number forecasts get built from, which is the reason for testing that one rather than a cleaner one nobody uses.

How to check your own

Export your pages with clicks, impressions and average position. Bucket by position, then sum clicks and impressions inside each bucket and divide. Do not average the per-page rates: that gives a page with nine impressions the same vote as a page with nine hundred thousand, and it is the most common way this analysis goes wrong.

Then split your largest bucket by whether the query can be answered in the result itself. If the two halves differ by more than about two, your site-wide position curve is a blend of two populations and you should stop forecasting from it.