What Search Console Will Not Tell You, and How to Know When It Is Misleading You
Search Console is the most honest tool in SEO and it will still hand you a number that means something different from what you think. Not because it lies, but because it answers exactly the question you asked, including the parts of the question you forgot you set.
Short answer: Search Console withholds low-volume queries for privacy, silently keeps filters you set weeks ago, reports the date range you requested rather than the range containing data, and averages positions in a way that describes no real search. Every one of these is defensible on Google's side. All four will still mislead you if you act on the numbers without checking.
These are not hypotheticals. Each one below came out of a real export from my own site, and I have quoted the warnings my own tooling produced verbatim.
1. The filter you forgot you set
From the audit output
“This export is FILTERED (Country=Nepal). Every figure describes that slice only.”
Filters in Search Console persist. Set a country filter to check something in June, come back in August, and it is still applied. Export, and the file carries it — with no indication inside the numbers themselves that it did.
This is the most dangerous of the four because a filtered export is indistinguishable from an unfiltered one once the figures are in a spreadsheet or a client report. A country filter turns worldwide performance into one market. A device filter turns everything into mobile-only. Nothing in the data announces it.
The check: look at the filter chips above the chart before you export, every time. If you report on this regularly, record which filters were active alongside the file, or the export becomes unreproducible three months later when you cannot remember.
2. The clicks attributed to nothing
From the audit output
“9 clicks and 77 impressions are not attributable to any listed query. Google withholds low-volume queries for privacy.”
Google does not show queries searched by very few people, because a rare enough search can identify an individual. Those clicks still appear in your totals. They simply belong to no row in your query table.
On my export the totals were 26 clicks and 245 impressions. The listed queries accounted for 17 clicks and 168 impressions. More than a third of the traffic belonged to queries I will never see.
The trap is that the query table looks complete. Sum it, calculate a click-through rate from it, and you have quietly computed a rate for a subset while labelling it the whole.
The check: always compare the sum of your query rows against the reported total. State the gap explicitly rather than absorbing it. This gap is proportionally largest on small and new sites — exactly the sites whose owners are most likely to read the table as the full picture.
3. The date range that does not describe the data
From the audit output
“The filter claims Jun 1 2025–Aug 15 2026, but the first day with impressions is 2026-06-25.”
The range you select is the range you asked about, not the range in which anything happened. My export claimed roughly fourteen months. The site had existed in search for about seven and a half weeks.
Any rate calculated across the claimed range would have been wrong by roughly a factor of nine. “26 clicks over fourteen months” describes a dead site. “26 clicks over seven and a half weeks” describes a new one. Same data, opposite conclusions, and only one of them is true.
The check: sort by date ascending and find the first row with non-zero impressions. That is your real start date. Use it for every rate you calculate, and state it in any report.
4. Brand and non-brand look identical in a total
From the audit output
“Classified as brand by fuzzy match (likely misspellings): ‘shasikant’.”
This is not a flaw in Search Console, it is a flaw in how people read it. The single most useful thing you can do with the data is split it in two.
| Segment | Clicks | Impressions | CTR |
|---|---|---|---|
| Brand | 17 | 80 | 21.25% |
| Non-brand | 0 | 88 | 0.00% |
The total, 26 clicks, reads as a small but functioning site. The split reads as a site that has not started working: every attributable click came from somebody who already knew my name, and 88 impressions to strangers produced nothing.
Note the fuzzy-match caveat. A misspelling of your name is brand traffic, but classifying it is a judgement call, so the term judged should be named rather than silently folded into a total. That is the same principle as everything else here: disclose the decision instead of hiding it inside a number.
A fifth thing: average position describes no real search
Average position averages every impression. A page at position 3 for a few searches and position 80 for many shows a middling figure that describes neither situation. Worse, it only counts searches where you appeared at all, so going from invisible to position 90 can make your average position look worse while you are actually improving.
In a low-volume market like Nepal this matters more than elsewhere, because a single position change on a rarely searched term visibly moves the average. Trends across a group of queries are signal. Individual position movements are usually noise, and rank-tracking reports built on them are theatre.
Use the API if you do this repeatedly
The interface caps exportable rows and applies whatever filters are currently set. Both introduce silent error. The API lets you specify the date range and dimensions explicitly on every request, which makes the pull reproducible and removes the inherited-filter problem completely.
What it does not do is give you the anonymised queries. That limit is about privacy, not about which access method you chose, so the unattributed gap is present in both. Use the API for consistency, not in hope of more data.
The underlying principle
Every caveat above is a case of a tool answering precisely the question asked, including parts of the question the person forgot they had set. The failure mode is not bad data, it is unstated assumptions travelling downstream into a decision.
This is why the tooling I built labels every finding as measured, computed, or unverifiable, and turns anything it could not check into an explicit open question rather than a silent pass. The full account is in the SEO automation write-up, the internal linking discovery has its own article, and the Search Console parsing code is public on GitHub.
Common questions
Why do my query numbers not add up to the total?
Google withholds low-volume queries for privacy. Those clicks count in the total but belong to no listed query. On my export it was over a third of the traffic.
Does Search Console remember filters when exporting?
Yes. Filters persist across sessions and the export carries them, with nothing in the numbers to reveal it. Check the filter chips every time.
Why does the date range not match the data?
It reports what you requested, not where data exists. Find the first row with non-zero impressions and use that as your start date.
How do I split brand from non-brand?
Filter queries containing your business name, allow for misspellings, and treat the rest as non-brand. It is the most useful split available.
Is average position reliable?
As an average, yes. As a description of reality, no. It only counts searches where you appeared, so improving can make it look worse.
API or CSV?
API if you do it repeatedly, for reproducibility. It will not return anonymised queries either.
Want a second read of your own Search Console data?
Send me access and I will tell you what the numbers actually say, including which filters were applied, how much traffic is unattributable, and whether search is bringing you strangers or only people who already knew your name.