Thirteen thousand keyword probes later: which phrases actually return volume data
Across 13,103 keyword tool probes run between 2026-10-13 and 2026-10-25, 72.6% returned nothing: zero rows, no volume, no keyword difficulty. The single strongest pattern we found: engineered long-tail phrases built as <category> for <condition> fail 95-97% of the time, while plain spoken category words almost always return data.
What percentage of keyword probes return zero results?
72.6% of the 13,103 probes returned nothing — no volume, no keyword difficulty, zero rows. Each probe was one phrase sent through Semrush Keyword Magic Tool; each empty result is a CSV with a header and no data rows. In total the non-empty probes returned 80,674 matched keywords, which means the empty results were not rare edge cases. They were the majority of the dataset.
The probes ran in a dual-gate setup. Gate A required volume of at least 100 and keyword difficulty of 20 or less; gate B required volume of at least 500 with no difficulty filter. A phrase counts as empty only when both gates return zero rows, so the 72.6% rate cannot be blamed on one strict filter.
| Word shape | Probes | Empty rate | Note |
|---|---|---|---|
Engineered long-tail — <category> for <condition> / by <parameter> | ~700 | 95-97% | Consistent across three consecutive batches; the hardest finding in this dataset |
Spoken category words — <category> calculator / cost / size / how many … do i need | ~1,900 | 22-71% | Same tool, same database; the word shape changes everything |
| Control batch, all filters removed (any entry counts) | 140 | 91% | Only 9% of phrases have any entries at all |
| Total | 13,103 | 72.6% | 80,674 matched keywords returned |
A rate this high is not in any tool’s public documentation, and we could not find another public probe-level dataset — most teams run a handful of phrases and stop. Our 49 batches over 13 days exist because we fund the probes as a cost of picking directions, not as research.
Which word shapes actually have volume?
Split the 13,103 probes by word shape and the 72.6% average hides two opposite worlds. Plain spoken category words — <category> calculator, <category> cost, <category> size, how many <category> do i need — returned data 29-78% of the time depending on the category. Engineered long-tail phrases built as <category> for <condition> or by <parameter> returned data only 3-5% of the time.
Same tool, same database, same two weeks of October 2026. Roughly 1,900 of the probes were spoken category words; roughly 700 were engineered long-tail. Both groups went through identical gates, so the gap is not a filtering artifact — it is in the phrases themselves.
Why do “for / by / with” conditional phrases fail?
Because almost nobody types them. These phrases live in content briefs and AI-generated topic lists, not in search boxes. Three consecutive batches in mid-October 2026 produced empty rates of 95-97% for this shape — the most consistent result in the whole dataset. When a seed pattern fails at that level across independent batches, the pattern is the problem, not the tool.
Does a low-KD filter cause the empty results?
No, and we ran the control to prove it. On 2026-10-22, in batch 41, we re-ran 140 phrases with every filter removed — no minimum volume, no difficulty ceiling, unfiltered phrase match. 91% of them still returned zero entries; only 9% had any keyword data at all. A strict gate hides weak keywords, but it cannot hide keywords that do not exist.
This control rules out the cheapest explanation for the 72.6% headline. If the empty rate were caused by our volume and difficulty gates, the unfiltered control would have filled most of those rows. Instead, on 2026-10-22, 128 of 140 phrases came back empty with every gate removed.
How should you write keyword seeds, then?
Start from the category word a person would say out loud, then attach one modifier with task intent: calculator, cost, size, or how many. We covered the category-versus-long-tail gap from the selection side in our keyword matrix analysis. And if you build pages from seeds like these, the next bottleneck is getting them indexed — we measured Search Console’s indexing quota the same way, one logged attempt at a time.
How was this measured?
Every number on this page comes from the same archive, and the archive is kept raw — no sampled screenshots, no reconstructed logs. If a claim in the table above cannot be traced to a dated entry in that archive, it does not appear on this page. The pipeline, exactly as run:
- Tool: Semrush Keyword Magic Tool, Phrase match mode, US database, USD.
- Gate A: volume at least 100 and keyword difficulty at most 20. Gate B: volume at least 500, no difficulty limit. Gate C: no filters at all — control batch only.
- Empty result: the phrase returns zero data rows, which in the archive is a CSV containing only the header line.
- Sample: 13,103 probes in 49 batches, run between 2026-10-13 and 2026-10-25.
- Counting: one probe is one phrase checked in both gates; a probe is empty only if both gates return zero rows.
- Verification: after this page was drafted on 2026-10-26, we re-scanned the raw archive independently — every probe CSV across all 56 round directories. That re-scan found 13,988 probe files with the same 72.6% empty rate, and the 2026-10-22 control batch re-counted at 140 probes, 91.4% empty. The small difference in totals is re-probe deduplication in the published aggregate; the rate is identical either way.
What does this data not tell you?
Three honest limits, stated because a data page should know its own edges. First, this is one tool, one regional database, and one match mode — another tool or another region may return different numbers. Second, an empty result means the tool has no data for the phrase; it does not prove nobody searches for it, and the 2026-10-22 control exists precisely because empty and nonexistent are different claims.
Third, our probe set was chosen to pick site directions, so it over-represents the word shapes we test most. Treat the category-level rates as the robust finding, not any single batch.
Frequently asked questions
What counts as an empty keyword probe?
An empty probe is a phrase that returns zero data rows after server-side filtering — in our archive, a CSV file that contains only the column header. Each probe is one phrase checked under two gates: volume at least 100 with keyword difficulty at most 20, and volume at least 500 with no difficulty limit. If both return zero rows, the probe counts as empty. The 72.6% rate is counted this way across all 13,103 probes; the unfiltered control batch is reported separately from that headline number.
Is the 72.6% empty rate specific to one keyword tool?
Yes. This page measures Semrush Keyword Magic Tool with phrase match on the US database, and the number should not be transplanted to other tools. What transfers is the method: run your own seeds through your own tool, write down the empty ones, and aggregate by word shape. We would expect any volume database to show a similar failure pattern for engineered <category> for <condition> phrases — not because every tool agrees, but because no tool can have data for phrases nobody searches.
Does an empty result mean nobody searches for the phrase?
No. An empty result means the tool’s database has no entries for that phrase under the filters used. On 2026-10-22 we ran the control: 140 phrases with every filter removed, and 91% still came back empty. That tells us the phrases lack data in this database — not that human demand is zero. Search demand exists outside any single keyword database, especially for new or spoken-language topics that seed tools have not caught up with.
Which keyword seeds are most likely to return volume data?
Plain spoken category words with a task modifier: <category> calculator, <category> cost, <category> size, or how many <category> do i need. In our sample of about 1,900 such probes the empty rate ranged from 22% to 71% depending on the category — high variance, but every category returned data at least some of the time. Engineered conditional phrases of the form <category> for <condition> returned data only 3-5% of the time across about 700 probes in the same archive.
How often should seed rules be re-measured?
Whenever the tool’s database updates or your niche drifts — in practice, once per selection round rather than on a calendar. Our rates come from 49 batches between 2026-10-13 and 2026-10-25, and we treat them as per-round facts, not permanent constants. A cheap habit that keeps this honest: keep every probe CSV, re-count empties at the end of each round, and retire any seed pattern whose empty rate climbs for three consecutive batches.
More measurement notes like this one live on the blog; the free tools we run are listed on /tools.