How it decides, and where it fails
Most tools in this category won't tell you how detection works. It's the only thing worth judging us on, so here it is, including the parts that currently don't work well.
Five rules, applied in order
A signal has to survive all five to reach a person. Most don't.
A volume floor
Percentage change is never computed below an absolute threshold. Two searches becoming eight is a 300% rise and complete noise. This single rule removes more false positives than the other four combined.
Last year, not just last month
A rise is measured against both the recent baseline and the same week a year ago. Without the annual comparison, a tool confidently reports Black Friday as an emerging trend, every year, to a client who has planned for it for a decade.
Two horizons, not one
A slow steady climb is invisible to a short comparison, because the baseline absorbs the growth as it happens. So movement is checked over four weeks and over three months. Short catches events, long catches trends.
Sustained and spike differ
Three consecutive rising periods is a trend. One is an event. They need different responses, so they're never scored on the same axis.
Your own activity is excluded
Movement caused by the client's live campaigns is filtered out before anything is surfaced. Reporting a client's own campaign back to them as a discovery is the fastest way to lose their trust.
A judgement about fit
Whatever survives is assessed against that client's brand profile — buyer, positioning, tone rules, live campaigns — and either surfaced with a rationale and a window, or rejected with a reason.
What we currently get wrong
This list is maintained honestly and will change. If something here matters to you, ask about it before you buy anything.
| Problem | What it means for you |
|---|---|
| Acceleration is unreliable | Whether a rise is speeding up is computed from a short window and is mostly noise at realistic volumes. A genuinely accelerating query can be labelled steady. It's the most useful thing we could tell you and currently the least trustworthy. Being rebuilt. |
| Flattening rises get mislabelled | A real trend that's losing momentum is sometimes called a spike. The item still surfaces correctly, but the character is wrong, which affects what the draft suggests. |
| Long-tail search data is incomplete | Search Console anonymises rare queries, so 30 to 50% of query-level volume is missing. Fine for direction, wrong for absolute numbers. Never put these figures in a client report as totals. |
| New clients start slow | Seasonal comparison needs more than a year of history. Below that the annual rule switches off and we tell you it has, rather than quietly reporting seasonal noise as new. |
| Small evaluation set | Fit judgements are measured against a labelled set that's still small. It catches large regressions reliably and small ones unreliably. |
How we check ourselves
Every change to the judgement layer runs against a labelled set of cases across multiple brands before it ships. We measure how often it surfaces things it should, how often it surfaces things it shouldn't, and separately how often it surfaces something with no plausible connection to the brand at all.
That last number is the one we treat as a red line. A merely weak suggestion wastes a minute. An absurd one costs your team's confidence in the whole tool, and it's much harder to win back.
Causation
We record that you acted and what followed. We don't tell you the action caused the result, because with this data nobody honestly can. If a vendor in this category shows you a clean attribution number for trend activity, ask how it was calculated.
Benchmarks from other agencies
Not until there's enough data for it to mean something, and not without explicit opt-in and aggregation that can't be traced to a client.
Judge it on one client
The method only matters if the output is useful. One free workspace, one month, your own comparison.
No card. One workspace stays free permanently.