Page Depth and Position: Our Own Data

We publish our own AI-visibility numbers, so it seemed inconsistent not to publish the search side too. This is one site — ours — with 101 pages, three months of Search Console data, and rendered word counts taken from the built HTML rather than from source markdown. It is not a study. It is a worked example with the inconvenient rows left in.

The short version: word count is a threshold you have to clear on a contested query and does nothing for you on an uncontested one, and we have pages that prove each half.

What we actually measured

The window is the three months ending 11 August 2026: 2,204 impressions across 213 distinct queries, with 73 of our 101 pages earning at least one impression. Two clicks. That number is not a typo, and it is the reason this analysis exists rather than a footnote to it.

For every page with impressions we joined three fields: impressions, average position, and the number of words a parser actually reads from the rendered page. That last one matters more than it sounds. Counting words in your markdown counts things nobody sees; counting words in the rendered HTML counts what an extractor or an answer engine can lift. Our posts came out consistently longer in source than in output, because frontmatter, code fences and component props do not survive to the page.

The table

Page	Impressions	Avg. position	Rendered words
/blog/track-ai-visibility/	582	89.1	668
/geo-agency/dubai/	425	79.3	1,953
/geo-agency/singapore/	349	88.4	1,743
/geo-agency/abu-dhabi/	186	84.3	1,688
/blog/geo-vs-aeo-explained/	120	80.9	679
/blog/show-up-in-chatgpt/	105	80.2	656
/blog/geo-agency-dubai-buyers-guide/	78	61.3	1,311
/blog/optimize-content-for-llms/	48	94.2	586
/geo-agency/munich/	45	18.3	1,703
/geo-agency/frankfurt/	26	6.9	1,676
/geo-agency/luxembourg/	26	15.0	1,585
/research/	25	8.1	1,117

Twenty-three of our twenty-seven blog posts at the time were under 800 rendered words, with a median of 675. The single post above 800 that carried real search volume was the Dubai buyer’s guide at 1,311 words, and it was also the best-ranked of them at position 61 against position 80 to 94 for the rest.

The pattern that holds

Every page we own that reaches the first two pages of results on a term anyone else is competing for is at least 1,100 rendered words. There are no exceptions in our data. The three city pages at positions 6.9, 15.0 and 18.3 are all between 1,585 and 1,703 words; the research index at position 8.1 is 1,117.

That is a floor, and it is worth understanding what a floor is. It does not say that adding 400 words to a page moves it up. It says that we have never once cleared a competitive query with a page that thin, which is a much weaker and much more useful claim.

The pattern that does not hold

Four of our posts sit at roughly 660 words and rank between position 9 and 33: why-reddit-is-cited at 33, rank-in-chatgpt at 11, ai-vendor-shortlist-sales at 9.5, how-llms-choose-citations at 15. If length drove position those rows would not exist.

What they have in common is that each carries between 1 and 19 impressions. Nobody is competing for those queries. A 660-word page is entirely sufficient to win a question nobody else is answering, and no amount of depth would earn it more traffic, because the demand is not there to win.

Then there is the row that does most of the work in this analysis. Our thirty-one city pages are all between 1,585 and 1,953 rendered words — a tight band, same template, same depth — and they span position 6.9 to 88.4. Frankfurt at 6.9 and Singapore at 88.4 differ by more than eighty positions with essentially the same word count. Whatever separates them, it is not length. It is how many other people want the same query.

Why “write longer” is the wrong conclusion

If you take one thing from the table, take this: the variable that predicts our position is competition, and word count is only visible because it correlates with whether we bothered to treat a query as contested.

The causal story that fits our data is unglamorous. On a query several people are answering, the page that covers the question completely beats the page that covers part of it, and covering a question completely takes more words than covering part of it. The words are a symptom of the coverage. Pad a thin page to 1,200 words without adding coverage and you have a longer thin page, which is worse than the short one because it takes longer to discover that it does not answer anything.

There is also a ceiling nobody talks about. Our longest city page, Dubai at 1,953 words, sits at position 79. Depth did not rescue it, because the Dubai query is the most contested in our entire set.

The measurement trap we walked into

Worth recording because it nearly sent us the wrong way. After rewriting four posts to roughly double length, our internal content-coverage metric reported that all four had got worse at matching the search demand they were written for.

They had not. The metric scored whole-page cosine similarity against a short query, and that is length-biased: a longer document spreads its normalised weight across more terms, so similarity to any short query falls mechanically as the page grows. The metric was quietly rewarding thin pages — the exact opposite of what the position data says. Scoring each query against a page’s best-matching passage rather than the page average fixed it, and moved our measured content gaps from 14 to 4 out of the same demand set.

If you are using an embedding-based coverage score to decide what to write, check whether it is doing this before you trust a number that tells you your best work got worse.

Limitations, stated plainly

This is one site, in one category, over one three-month window, with two clicks in it. Every honest caveat applies and several are load-bearing.

Position and word count are confounded by intent: we wrote longer pages for the queries we cared about, so of course the longer pages are the ones on contested terms. Average position in Search Console is itself an average over wildly different queries, and a page can be at position 6 for one phrase and 90 for another. Our sample is small enough that a single unusual page moves a median. And the whole window is a period of rapid page growth, which drags average position down mechanically as new pages index at depth — ours fell from about 50 to about 83 while impressions rose more than fivefold.

We have since rewritten the four thinnest high-impression posts, roughly doubling each. We have not yet re-measured their positions, and we are not going to guess. When there is a post-change reading we will publish it here whether it supports this analysis or embarrasses it, the same way we publish our AI visibility numbers including the run where they fell.

Questions this raises

Does any of this transfer to AI answers rather than search? Not directly, and we would not claim it does. Our own AI share of voice moved from 12.8% to 11.1% across the same period in which these search impressions rose more than fivefold. Publishing depth on your own domain is not what gets you into someone else’s shortlist.

Is there a number to aim at? For a contested query on our data, 1,100 rendered words is where our pages start winning. Treat it as a diagnostic rather than a target: if you cannot reach it without padding, you have not finished the research.

Should I rewrite my thin pages? Only the ones on queries somebody is competing for. Rewriting a page that ranks 11th on a query with four impressions a month is work you will never be paid for.

The bottom line

Depth is a floor that competitive queries impose on you, not a lever you pull. Find out which of your pages are on contested queries, check whether those pages cover their question completely, and leave the rest alone. Everything else here is our data rather than a rule, and one site is not evidence — which is why we would rather show you the table than tell you the conclusion.

The same discipline applied to the other surface is in how to track your brand’s visibility in ChatGPT, and the claims we do and do not stand behind in this field are graded on our evidence register.