A page-one query gave me 2,246 impressions and 19 of them came from a phone
Google Search Console told me one of my blog pages had 2,246 impressions over 28 days for the query "workout app development", sitting on page one. Semrush has no US volume figure for that query at all. When I split the same query by device in Search Console, 19 of those 2,246 impressions were mobile.
So I checked a second one. "ai manufacturing cost reduction case study": roughly 825 impressions a month in Search Console against Semrush's 320 a month, and 1 mobile impression.
Desktop-only impressions at that scale, on queries with no measurable human search volume, are crawlers and scrapers rather than readers. A page can look like it is winning in Search Console while almost nobody has seen it.
That cost me a real experiment. I had an on-page change running on one of those pages and I was scoring it on impressions and average position. Since 25 September I have not scored it on impressions at all, because its numbers were almost entirely desktop, and there was nothing in them to read. I judge page changes on mobile impressions and clicks now.
The two signals also move independently, which I did not expect. Over the same stretch my 28-day average position went from 19.3 to 23.5, so worse, while CTR went from 0.24% to 0.32%, so better. Ranking and getting clicked are not the same problem and the same report shows both.
Do you split Search Console by device before calling a page change a win?
Replies
This is a fantastic catch. Those desktop impression spikes on zero-volume queries are almost always automated rank-tracking tools and SERP scrapers polling Google. Because they run on headless desktop Chrome by default, they completely distort desktop impression data.
Treating GSC impressions as a vanity metric is easy; filtering by device to isolate mobile is one of the quickest ways to separate human intent from bot pings.
It also highlights why average position can be misleading: ranking 19 vs. 23 doesn't matter if nobody was clicking anyway. Scoring experiments on mobile clicks and qualified conversions avoids optimizing for robots talking to robots.
Great catch. The App Store has its own version of this. When I track where my app ranks for a search term, the rank the store shows on a phone and the rank its search API returns can differ by tens of places on the same term at the same time, one keyword sat at 63 in the store and 25 in the API. So I stopped trusting any single instrument. A move counts only when two sources agree on the direction, and the number I actually score changes on is installs from that term, which no bot produces. Your mobile split is the same idea: find the signal that only a human can generate.