The queries your site ranks best for may be the ones no person ever typed. We measured 5,910 queries across ninety days and found two shapes that rank well and click at zero.
They aren’t a rounding error. Together they account for 31% of the distinct queries in the report.
Here’s what they look like, what they do to your averages, and how to pull them out before you read anything else.
What the Split Looks Like
We sorted every query by its shape, then compared impressions, clicks and average position.
| Query shape | Queries | Impressions | Clicks | Avg position |
|---|---|---|---|---|
| Imperative, instruction-like | 848 | 1,768 | 0 | 10.3 |
| Long conversational questions | 977 | 12,259 | 0 | 32.9 |
| Everything else | 4,085 | 185,810 | All of them | 47.2 |
Read the last column. The instruction-shaped queries rank 37 places better than the ordinary ones and produce nothing at all.
The Two Shapes
They look nothing alike, and only one of them is obviously strange.
Instructions rather than searches
These open with a verb aimed at a machine: evaluate, compare, analyse, list, assess. They read as tasks handed to a system, not as something typed into a search box.
They also tend to carry a product or company name and a set of options, in the shape of a template that has been filled in repeatedly.
Questions nobody would type
The second shape is a full sentence question, often ten or more words. It’s phrased the way someone speaks to an assistant rather than the way anyone types into a box.
These are the larger group by volume, and the ones most likely to be mistaken for a real opportunity. A fluent question at position 30 looks worth chasing.
What They Do to Your Numbers
They distort the two figures people check first.
- Average position improves: the instruction-shaped group sat 37 places above the ordinary one.
- Click-through rate collapses: impressions arrive, clicks do not, and titles had nothing to do with it.
- Striking-distance lists fill with noise: a page at position 9 may have no human audience.
We spent a session preparing title rewrites for exactly that reason before checking the query shapes behind them. The rewrites would have changed nothing.
How to Separate Them
You can do this in a spreadsheet in about ten minutes.
- Export queries with impressions, clicks and position. The query dimension alone is enough; you don’t need the page cut for this.
- Flag the imperative openers. Evaluate, compare, analyse, list, assess, generate, act as. A short prefix match catches most of them.
- Flag questions of six words or more. Then read a sample rather than trusting the filter, because plenty of long questions are ordinary people.
Then recompute average position and CTR with those rows removed, and compare. The gap between the two versions is the size of your measurement problem.
What This Does Not Prove
Google doesn’t label where a query came from, so nothing here identifies these as bots, assistants or automated tools. We can describe the shape and the click behaviour, and that’s the limit.
Three explanations survive the data, and this report separates none of them.
They could be assistant-mediated searches where nobody sees a results page. They could be automated evaluation running against live search. They could be ordinary searchers whose phrasing has changed faster than our assumptions.
What would settle it
Matching these query patterns against server logs, which record the agent making the request. We set out how to read those in server logs for AI search, and it’s the only place provenance is recorded.
Why This Sits Alongside Indexation
It’s the second thing Search Console will let you believe. The first is that impressions prove a page is indexed, which we took apart in a real audit of 285 pages.
Impressions cannot tell you a page is indexed, and they cannot tell you a person searched. Both are the same mistake: treating a count as evidence of the thing you care about.
Frequently Asked Questions
Why do I have impressions but no clicks
It happens because position explains most of it and query shape explains much of the rest. Long conversational questions and instruction-shaped strings both generate impressions and almost never a click.
Are these queries bots
Nobody can tell you from Search Console, because Google doesn’t publish query provenance. The shape is observable and the cause is not, so treat any confident answer as a guess.
Should I optimise for conversational queries
Not on click potential. They’re worth serving if your content answers them, since assistants read pages, but expect the return in citations rather than in sessions.
Does this mean my CTR is fine
It means your reported CTR is measuring something other than what you assume. Recompute it with the two shapes removed, then judge.
How often should I check
Once a quarter, and always before you commission title rewrites. The check costs ten minutes and can save a sprint aimed at the wrong problem.
Recompute Before You Rewrite
Pull the query export, flag the two shapes, and recalculate. If your averages move much, the page you were about to rewrite was never the problem.