Search research · September 9, 2026
When Page One Stops Paying: A Measured Case
Ranking first on a non-brand query returned 4.25%, not the textbook 28%. What 91 days of Search Console and Bing AI citation data showed, and what it changes.
By Shihab Shahriar Antor · Updated 2026-09-09
This is a single-site observation, not a study. One property, 91 days to 6 September 2026, Google Search Console plus Bing Webmaster Tools. It is worth writing down because the shape of it changed what we planned to do next, and because the usual diagnosis for these numbers would have been wrong.
The number that did not make sense
The site had pages sitting on page one and returning almost nothing. The reflex diagnosis is a snippet problem: bad title, weak description, rewrite and move on. Before doing that we checked the site's own clickthrough curve by position, on non-brand queries only.
| Position band | Clicks | Impressions | Observed CTR |
|---|---|---|---|
| 1 to 3 | 13 | 306 | 4.25% |
| 3 to 5 | 24 | 959 | 2.50% |
| 5 to 7 | 86 | 4,730 | 1.82% |
| 7 to 9 | 147 | 14,249 | 1.03% |
| 9 to 11 | 129 | 17,458 | 0.74% |
| 11 to 15 | 29 | 4,490 | 0.65% |
One property, 91 days to 6 September 2026, Search type Web, brand queries excluded. Small counts in the top band: treat the 4.25% as an order of magnitude rather than a precise figure.
Position one to three returned 4.25%. The commonly quoted figure for that band is somewhere near a quarter of all clicks. Nothing about the pages explains a gap that size, and we checked: the titles were keyword-first and inside 60 characters, and the descriptions answered the query with real numbers in them. Rewriting them would have been motion.
The second cut explained more than the first
Splitting the same queries at identical positions by whether they contained a competitor's brand name produced a cleaner signal than anything about the pages themselves.
| Query type | Clicks | Impressions | CTR |
|---|---|---|---|
| Contains a competitor's brand | 9 | 3,751 | 0.24% |
| Generic, no competitor brand | 267 | 27,956 | 0.96% |
Both rows are positions 7 to 11. The difference is the query, not the ranking.
Someone searching a competitor's brand plus "pricing" wants that competitor's pricing page. Ranking eighth for it produces impressions almost by definition, and no title changes that. Those impressions were never convertible, and treating them as a CTR failure would have meant optimising toward an outcome that cannot happen.
That is the part worth generalising. A large impression pool at a good average position is not automatically an opportunity. It can be a query you rank for and should not expect to win.
Meanwhile the citations went up ninefold
Over the same window, Bing's AI Performance report recorded citations growing from 416 in the first thirty days to 3,861 in the last thirty. The same pages that were not being clicked were being read, summarised, and attributed inside generated answers.
Which makes the next question the important one
If an engine is answering on your behalf, the lever stops being the click and becomes the content of the answer. So we pulled the grounding-query report to see what the citations were actually about.
Every one of them described the free tier. Templates, comparisons, getting started. Not a single grounding query in the classified set touched the paid capabilities, even though those are the reason the product has a price. The pages carrying the paid story held well under one percent of the site's search visibility, so there was almost nothing for an engine to find.
That reframes a lot of work. Publishing more pages aimed at the same queries would have produced more citations that said the same thing. The useful move was to make the paid capabilities describable on pages an engine already reaches.
What we would check on another site
- 01
Compute clickthrough by position on your own property
Not against a published curve. Segment out brand queries first, because a handful of navigational terms will otherwise carry the whole top band and hide the shape.
- 02
Split page-one underperformers by query intent
Separate queries naming a competitor from generic ones. If the competitor-brand rows are far worse at the same position, that pool is structural and no snippet work will move it.
- 03
Read the live titles before rewriting anything
Fetch what is actually deployed. Ours were already correct, and a rewrite programme would have consumed weeks for nothing.
- 04
Compute clicks per thousand by URL prefix
Clusters differ by more than an order of magnitude. One section returning 3 per thousand and another returning 29 are doing different jobs, and only one of them is a traffic asset.
- 05
Pull the AI grounding queries, not only the citation count
The count tells you that you are being read. The grounding queries tell you what you are being read as, which is what should drive the next thing you write.
Questions
- Does a low clickthrough rate on page one mean my titles are bad?
- Not necessarily, and checking is cheap. Compare against your own site's clickthrough curve by position rather than a published benchmark, and separate queries that name a competitor's brand. On the property measured here, competitor-brand queries converted four times worse than generic ones at identical positions, which no title rewrite would have changed.
- Why would ranking first only return 4.25%?
- Because the result is no longer the first thing on the page. When an answer is generated above the results, the query is often satisfied before anyone reaches a link. It shows up as healthy impressions, a stable average position, and clicks that do not follow.
- Are AI citations worth anything if nobody clicks?
- They are worth something different. A citation puts your name and your framing inside the answer a person actually reads, which influences what they believe before any click happens. It is measurable in Bing Webmaster Tools. It is not a substitute for traffic, and it should not be reported as though it were.
- What should I do about a large cluster with very low click yield?
- Diagnose before investing. Low yield has at least three causes that need opposite responses: an instant-answer cluster that is earning citations and doing its job, a navigational endpoint nobody clicks through from search, or a retired host still sitting in the index. Only the last one is a defect.
- Does adding structured data increase AI citations?
- There is no good evidence that it does. Google retired FAQ rich results in May 2026, and a controlled study published days later found no significant citation lift from adding JSON-LD. Keep structured data for classic rich results and machine readability, and do not build a citation strategy on it.
One property is one property. The curve on your site may look nothing like this, which is the reason to compute it rather than assume it. What generalises is the method: measure your own clickthrough by position, split by query intent before diagnosing, and read what the AI grounding queries say about you rather than only counting them.
Every table here came out of a script that reads a Search Console performance export and prints yield by URL prefix, clickthrough by position band, and the page-one queries converting under one percent. It is about two hundred lines of Node with no dependencies, and the five steps above are the whole method if you would rather write your own.