Average position 3.2 tells you how many organic results Google put above you. It tells you nothing about whether a user has to scroll past an AI Overview, four ads, a local pack and a People Also Ask block to see you. Those are different facts, and only one of them predicts clicks.
This is why CTR-by-position curves keep breaking. The curve assumes position is a proxy for visibility. On a modern SERP, the mapping between ordinal rank and vertical pixel offset is unstable across queries, devices, and weeks — so a stable position number can sit on top of a collapsing click rate.
What Search Console's position number actually counts
Google documents position as the topmost position your site occupied in the results for that query. If your page appears twice, you get the higher one. Positions are counted within the result set, and different result types are grouped: a link inside a carousel or an image block is positioned relative to its container, not to the blue links around it.
Two consequences people routinely miss:
- Position is ordinal, not spatial. It counts results above you, with no weighting for how tall they are. An AI Overview that fills a phone screen and a one-line sitelink-free blue link both consume rank slots in a way that does not reflect their pixel footprint.
- Impressions are not eyeballs. Google documents that a link on the loaded results page counts as an impression even if the user never scrolls to it; the exception is elements that need an additional action, such as scrolling a carousel sideways or expanding an accordion. So "impressions flat, clicks down" is exactly what you would expect if your result moved further down the page without changing rank.
That second point is the measurement trap. Search Console's two visibility metrics — impressions and position — are both insensitive to vertical displacement. The metric that is sensitive to it, CTR, is the one you are trying to explain.
Why the same position moves hundreds of pixels
The height of everything above you is a function of the query, not of your page. Within a single keyword set you will see:
- AI Overviews that render anywhere from a few lines to most of a mobile viewport, sometimes expandable, sometimes not.
- Ad blocks that vary from zero to four units depending on auction density that day.
- Local packs, video carousels, product grids, Top Stories — each hundreds of pixels tall.
- Sitelinks on the top result, which can double the height of a single organic entry.
- People Also Ask, which can appear above or below you and expands on click.
None of these changes your ordinal position in a way Search Console exposes. They all change whether your listing is above the fold.
How to measure pixel depth without guessing
You need the vertical offset of your result from the top of the document, at a fixed viewport, for a fixed query set, sampled repeatedly. The method is a headless browser, a stable viewport, and the bounding box of the element containing your URL.
// Playwright: offset of the first result linking to your domain
const box = await page.evaluate((host) => {
const a = [...document.querySelectorAll('a[href]')]
.find(el => el.href.includes(host) && el.offsetParent);
if (!a) return null;
// climb to the result container, not the inline link
const r = a.closest('[data-sokoban-container], [data-hveid]') || a;
const rect = r.getBoundingClientRect();
return { top: rect.top + window.scrollY, height: rect.height };
}, 'example.com');
Four caveats that decide whether the numbers mean anything:
- Fix the viewport and device emulation. 393×852 mobile and 1440×900 desktop produce different layouts and different above-the-fold thresholds. Report them separately; never average them.
- Fix location and language. SERP feature composition is strongly geo-dependent. A rotating proxy pool will add variance you will later misread as a trend.
- Scripted queries against Google conflict with its terms of service, and detection changes the results you get. Most teams who do this at scale use a commercial SERP API. If you run it yourself, keep volumes small and treat the data as a sample, not a census.
- Sample repeatedly. SERP features flicker. One run per query per week is noise; several runs per day on a small query panel is signal.
What to log so the data stays comparable
Store one row per query, per device, per run. The useful fields are not just your own offset:
| Field | Why it matters |
|---|---|
query, device, locale, timestamp | Required to compare anything to anything |
ordinal_position | Lets you join to Search Console |
pixel_top, element_height | Your actual location and size |
above_fold (boolean, viewport-relative) | The thing CTR responds to |
blocks_above (list: ads, AIO, local, PAA, video) | Explains why depth changed |
aio_present, aio_cites_you | AI Overview displacement is the biggest current mover |
total_page_height | Normalises depth across layout changes |
With that schema you can compute the thing Search Console cannot: a CTR curve indexed by pixel depth and by the set of blocks above you, rather than by rank alone.
Reading the data
Three patterns show up quickly once you have a few weeks of samples.
Stable position, falling CTR, rising pixel depth. A new feature block moved in above you. This is a SERP composition change, not a quality problem with your page, and rewriting the title will not fix it. Confirm by checking whether blocks_above gained an entry on the same date.
Position improved, CTR unchanged. You moved from 4 to 3 but both sit below the fold on mobile. Ordinal gains below the fold are worth much less than gains that cross it; this is the clearest argument for targeting queries where the fold is reachable.
High variance in pixel depth across runs of the same query. The SERP is unstable — often ad load or an intermittent AI Overview. Average position will look calm while actual visibility swings. Treat forecasts for these queries as wide intervals, not point estimates.
One honest limitation: pixel depth from a headless browser is a model of what users see, not an observation of it. Personalisation, zoom level, browser chrome, and ad blockers all shift the real fold. The measurement is still far better than ordinal rank, because it moves in the same direction and roughly the same magnitude as the thing you care about.
What to do next
- Pick 50–200 commercially important queries and run a twice-daily pixel-depth sample on both mobile and desktop emulation. Start narrow; breadth adds cost, not insight.
- Join the panel to Search Console by query and date, and build your CTR curve against
pixel_topandabove_foldinstead of position. - Before you open a CTR investigation, check
blocks_abovefor the affected queries first. If a block appeared, your diagnosis is displacement, and the response is query selection or feature eligibility, not another title rewrite. - Report visibility to stakeholders as "share of tracked queries where we appear above the fold on mobile", with average position as a secondary metric. It is the number that actually tracks clicks.