Measurement

We Moved Two Keyword Clusters 19 Positions in 96 Days. They Earned One Click.

Eleven supporting articles. Twenty archived scans. Nineteen positions of rank. One click. The depth lever did exactly what it promises and paid nothing, and the reason turns out to be the same reason our title rewrites failed.

By Thicket AI Team · August 16, 2026 · 9 min read

Between May and August 2026 we built eleven supporting articles around two calculator pages. Both clusters improved by roughly nineteen and thirteen positions in Google. Over the same period they produced 972 impressions and one click. Topical depth moved rank exactly as advertised; it stopped at position 28, and position 28 earns nothing.

What we did

We run a daily scan that groups Search Console queries by the page they land on. Any page with three or more distinct queries gets flagged as a cluster with an obvious recommendation attached: write supporting articles for the query variants. It is standard advice and we followed it.

Around a bonus-tax calculator we published eight articles — the flat versus aggregate withholding methods, the state-by-state breakdown, spot bonuses against annual ones, a step-by-step walkthrough. Around a daily-steps calculator we published three more. The earliest went live on 12 May 2026, which makes this read 96 days old at minimum. We hold ourselves to not judging content before thirty days, because we once ran a validation ten days after publication and measured nothing at all. This one has had its full run.

Rank: the lever worked

Impression-weighted average position, traced across every archived scan:

Scan date	bonus cluster	steps cluster
`2026-05-07`	46.6	42.0	first archived scan
`2026-05-23`	52.1	59.7	5 bonus articles now live
`2026-06-13`	53.9	55.1	both clusters drift worse
`2026-07-10`	48.2	42.8	recovering
`2026-07-26`	36.6	29.2	the step change
`2026-08-07`	34.7	28.8	holds
`2026-08-16`	27.9	29.4	best positions recorded

From 46.6 to 27.9, and from 42.0 to 29.4. Note that both clusters got worse through late May and June before recovering — if we had judged this at day 30 we would have concluded the opposite of the truth. The step change lands between 10 and 26 July, and we cannot attribute it cleanly: our own concentration decision and whatever Google was doing that fortnight are not separable from this data.

Clicks: the lever paid nothing

Cluster	Queries	Impressions (28d)	Clicks (28d)
bonus (8 supporting articles)	33	470	1
steps (3 supporting articles)	46	502	0

Zero clicks on 7 May. One click on 16 August.

One caveat we have to state because it cuts against us elsewhere: Search Console reports about 28% of this property’s finalized days as near-zero, so any window total is a floor rather than a point estimate. That inflates nothing here. Correcting one click for a 28% null rate gives about 1.4 clicks. The finding never rested on the impression count.

The control, and what is wrong with it

Clusters that received no supporting articles over comparable windows:

Cluster	Treated?	Δ position
bonus	8 articles	−18.7
steps	3 articles	−12.6
bulk QR	none	−6.3
bulk image compress	none	+2.5
amortization	none	+8.2

The direction agrees. The control is still weak and we are not going to dress it up. The treated clusters started at positions 42 to 47 and the untreated ones at 82 to 90, and rank movement is not linear in difficulty — climbing from 46 to 28 is not the same job as climbing from 82 to 76. Only one control spans the full window. Read the rank comparison as suggestive and no more.

The click result does not need a control. One click on 972 impressions is not a comparison; it is a measurement.

The part that changed our mind

Earlier this year we ran four title and meta-description rewrites to fix click-through rate. All four earned zero clicks. The diagnosis we wrote at the time was that a page sitting at position 28 has no clickable impressions that a better title can win.

The depth work has now arrived at 27.9 and 29.4.

Both of our on-page levers failed at the same coordinate, from opposite directions. One tried to win clicks at position 28 and could not. The other tried to escape position 28 and could not get past it. That is not two lever failures. It is one wall, measured twice.

What is left is the thing neither lever touches: the off-site signals that decide whether a young domain is allowed into the top twenty at all. We are not presenting that as demonstrated — it is the hypothesis standing after two eliminations, and it agrees with a separate result of ours where third-party attention, not self-published volume, tracked whether a company got named at all.

What we changed

The scan that produced this data had been recommending write supporting articles for the bonus cluster every day for 101 days, while eight such articles already existed. It only ever looked at Search Console and never at our own repository, so it could not tell “nobody has tried this” from “this was tried eight times and did not convert.” We have since taught it to count the articles that exist and to refuse the recommendation above position 20.

That bug is the more general lesson here. An automated recommendation that never checks whether it was already taken will keep proposing the same action forever, and its persistence reads as confidence.

What we are not claiming

This is one property, roughly 100 days old at the start of the window, with no meaningful referring domains. Nothing here tests what depth does for a cluster already inside the top ten, where a few positions translate into real click-through differences. Nothing here tests an established domain.

The bounded claim is the useful one: at position 28, on a young domain, eleven articles bought nineteen positions and one click. If your pages are out there too, the next article is probably not your problem.

Falsifier

If either cluster crosses into the top ten on the strength of the existing articles and starts earning clicks by 15 October 2026, then the lever was slow rather than capped and retiring it at position 28 was premature. That date is in our calendar and the result will be published either way.

Frequently asked

Does building topical depth around a page actually improve its ranking?

In our data, yes, and substantially. We built eight supporting articles around one calculator page and three around another, then traced both clusters across twenty archived scans between 7 May and 16 August 2026. The first cluster moved from an impression-weighted average position of 46.6 to 27.9, and the second from 42.0 to 29.4. That is a gain of 18.7 and 12.6 positions respectively, sustained across independent measurements. Anyone claiming supporting articles do nothing for rank is contradicted by this series.

So why did the traffic not follow?

Because the gain ended at position 28, and position 28 is page three. Across both clusters the same 28-day window recorded 972 impressions and one click. The depth lever did the job it advertises — it moved rank — and the rank it could reach does not convert. A lever that reliably produces an output nobody pays for is still a failed lever, and that distinction is the whole point of the result.

Could the missing clicks be a reporting artifact rather than a real absence?

No, and this is worth separating carefully. Google Search Console reports roughly 28% of this property's finalized days as near-zero, which deflates any window total, so our 972 impressions should be read as a floor rather than a point estimate. But a suppressed day cannot conceal clicks large enough to change the conclusion: correcting for the observed null rate moves one click to roughly 1.4. The argument never depended on the impression count. It depended on the click count being indistinguishable from zero, and it is.

What was the control group, and how good is it?

Four clusters that received no supporting articles over comparable windows. They moved -6.3, +2.5 and +8.2 positions, against -18.7 and -12.6 for the treated pair. The direction agrees with the treatment hypothesis, and we are still calling this control weak. The treated clusters started at positions 42 to 47 and the untreated ones at 82 to 90, and rank movement is not linear in difficulty — climbing from 46 to 28 is a different job from climbing from 82 to 76. Only one control shares the full window. Treat the rank comparison as suggestive. The click finding needs no control at all, because one click on 972 impressions is a measurement rather than a comparison.

How does this relate to rewriting titles and meta descriptions?

The two levers failed at the same coordinate from opposite directions, which is the part we found genuinely clarifying. Earlier we ran four title and meta rewrites aimed at improving click-through rate; all four earned zero clicks, and the diagnosis at the time was that a page sitting at position 28 has no clickable impressions a title can win. The depth work has now arrived at positions 27.9 and 29.4 and earned nothing either. One lever tried to win clicks at position 28 and could not. The other tried to escape position 28 and could not get past it. That is not two independent failures. It is one wall, measured twice.

What actually moves a page across the last twenty positions?

On this property, nothing we can do on the page itself has. The remaining candidate is domain authority — the off-site signals a young site with no referring domains does not have. We are not presenting that as a demonstrated conclusion; it is the hypothesis left standing after the two on-page levers were eliminated, and it matches an independent read from a different measurement. What we can say with confidence is narrower and more useful: if your pages sit beyond roughly position 20, more supporting content is unlikely to be the thing that pays.

Does this mean nobody should write supporting articles?

No, and we would be overreaching to say so. Our result covers clusters ranking beyond position 20 on a domain with no meaningful authority. Nothing here tests what depth does for a cluster already inside the top ten, where a few positions of movement translate into real click-through differences, and nothing here tests an established domain. The specific claim is bounded: at position 28, on a young domain, eleven articles bought nineteen positions and one click.