If your ranking monitor sends forty alerts a week, you stop reading it. If it sends none, you find out about a drop from a sales call. Both are the same failure: nobody defined what a change worth acting on looks like.
This is not a monitoring-loop article. We already published how to run the weekly Hermes loop and we stand by it. This piece is narrower and more annoying to get right: the alerting design. Get the bands wrong and the loop is just a ritual.
The likely problem
Your monitor is probably not broken. It is reporting on the wrong thing at the wrong volume.
Check these four symptoms before you touch the skill file.
Symptom | What it usually means |
|---|---|
You get more than ten alerts a week and act on none | No threshold bands; the monitor is a change log, not an alert system |
Positions look stable while clicks fall | You are monitoring position only, and the signal that moved is click or brand related |
Everything alerts on Monday morning | Single-day comparison; weekend data is noisy and will always trip a narrow band |
Alerts arrive after the fix has already shipped | The cadence is weekly but the change cycle is daily; the monitor needs a different trigger |
If two of those rows describe your setup, the fix is the same list of four changes below. None of them require a new tool.
Fix 1: Replace "alert on any change" with three bands
A useful monitor has exactly three states, and the middle one is what makes it bearable.
Band | Definition (example) | What the agent does |
|---|---|---|
Report | Position moves 3–5 spots, or impressions move 10–30% | Writes it to the weekly report. No notification. |
Watch | Position moves 6–10 spots, or a page leaves the top 20 | Adds it to a watch list with the date it entered |
Act | Position moves more than 10 spots, or a money query leaves the first page | Sends one notification, with the page and query attached |
Three things make these bands work. They are written in numbers, they are stored in the skill file rather than in your head, and the Act band is deliberately small. If your Act band fires more than twice a week, it is too wide.
Three bands, one notification channel. The Act band should stay quiet enough that you still read it.
Quality check: run the bands against last month's data before you switch them on. Count how many rows would have landed in Act. If the number is above eight per month, move the threshold.
If it fails: on low-traffic sites, position bands alone will never fire. Use impression bands as the primary trigger and positions as the secondary one.
Fix 2: Freeze the query set and group it
Monitoring drifts when the query list changes every week. Freeze it for a quarter, then review.
Group the queries the way the business already talks, not by search volume:
- Money queries — the ones that map to revenue. These get the tightest Act band.
- Comparison queries — where buyers evaluate. These move when competitors move, so they need annotation more than alerting.
- Brand queries — your own name and product terms. A drop here usually means something other than SEO, which is exactly why it belongs in the report.
- Support queries — where existing customers look for answers. Low commercial value, high signal about content decay.
Twenty to fifty queries is the working range. Below twenty you miss clusters; above fifty the Act band fires constantly because you are watching too much surface area.
Fix 3: Annotate before you read
The single highest-value habit in this workflow costs two minutes. Before the weekly report is generated, add a short note about what shipped: releases, template changes, redirects, pricing edits, PR pushes.
Without annotations, every movement gets an SEO explanation by default. That is how teams end up chasing algorithm updates for a week when the actual cause was a template change on Tuesday.
The Hermes skill should require this note as an input, not treat it as optional. If the note is missing, the skill writes the report with an explicit "no change log provided" line in the header rather than guessing.
The annotation enters before the analysis, not after. It is the cheapest accuracy improvement in the whole workflow.
Fix 4: Monitor the signals that actually moved in 2026
This is where most monitoring setups are a year behind. The Zyppy expert survey published on September 9, 2026 collected 13,665 data points from 131 practitioners. Behavior and click signals reached 29.4%, brand signals 27.0%, and technical SEO health 17.5%. Our breakdown of the page-versus-brand gap covers the blind spot in that survey, but the practical point for a monitor is blunt: if you only track positions, you are tracking the cheapest signal to replace.
Two additions are worth the effort:
Clicks per impression on flagged queries. A query holding position 4 while impressions rise and CTR falls is not stable. It is decaying, and positions will not tell you.
Branded search as its own line. Brand demand moves for reasons SEO cannot see, and it changes how every other number should be read. Track it separately so a brand-driven traffic swing does not get mistaken for a content win.
If you want to go further and add AI citation visibility, keep it in a separate section with its own thresholds. Mixing answer-engine visibility into a position monitor makes both harder to read, and the sources are not comparable.
Build it as a Hermes Agent skill
Hermes Agent is a reasonable fit for this because it runs a bundled skill repeatedly with memory across sessions, which is exactly what a weekly monitor needs. If you have not set it up yet, our Hermes SEO/GEO operator guide covers the project folder and the approval rules.
The skill file needs five blocks:
- Inputs — last week's data export, the change-log note, and the frozen query list.
- Bands — the three thresholds from Fix 1, in numbers.
- Groups — the four query groups from Fix 2, with which band applies to each.
- Output — a report in the fixed three sections plus a watch list carried forward from last week.
- Boundaries — read only, no site changes, no notifications outside the Act band.
A copy-ready starting point:
## Ranking monitor skill
Inputs: data/latest.csv, notes/change-log.md, queries/frozen-list.csv
Cadence: weekly
Bands:
- Report: position move 3-5, or impressions move 10-30%
- Watch: position move 6-10, or a page leaves the top 20
- Act: position move 10+, or a money query leaves the first page
Groups and band overrides:
- Money: Act band fires at 6+ positions
- Comparison: Watch band only, never notify
- Brand: Report band; flag separately, do not merge with content movement
- Support: Report band
Output: reports/monitor/YYYY-MM-DD.md
1. Moved rows that cleared a band
2. Likely explanation, limited to the inputs. Write "not explained by this data"
when the files do not cover it.
3. Watch list with the date each row entered it
If notes/change-log.md is missing, write "no change log provided" in the header.
Never recommend site changes. Never edit files outside reports/monitor/.Run it weekly. Read the report once. Act only on the Act band, and only on rows that carried the same explanation for two weeks.
Done looks like this: you can say how many alerts fired last month, how many became actions, and what your average time-to-detect was. If you cannot answer those three, the monitor is still a log.
The band design here came out of a mistake I made twice. The first monitor I set up alerted on every position change, which trained me to ignore it within a month. The second alerted on nothing, because I had set the Act band for a site with ten times my traffic. The version above is the third attempt, and the only one I still read.
Measure whether the monitor improved
Three numbers tell you whether any of this worked, and none of them are ranking positions.
Metric | Before | Target after 4 weeks |
|---|---|---|
Alerts per month | 20–40 | 4–8 |
Alerts that became a decision | 0–2 | Half of them |
Time from drop to detection | 2–6 weeks | 1 week |
If alerts fall but decisions do not rise, your Act band is too narrow rather than your monitor being quiet. Widen it by one step and run another month.
When not to use this
Skip the whole thing if you have no owner for the output. A weekly report that nobody reads is more expensive than no report, because it creates the impression that ranking work is covered.
Skip it too if your site gets fewer than a few hundred impressions a day. At that volume, use a monthly year-over-year comparison instead of a weekly monitor — the signal is not there yet, and you will spend the time tuning thresholds that cannot fire.
FAQ
Does this replace a paid rank tracker? No. Search Console covers your own property, which is what this monitor needs. Paid trackers add competitor positions, daily cadence, and locations you cannot see in your own data. Start free, add a tracker when you need a comparison rather than a baseline.
How many alerts per week is normal? Four to eight per month is the working range for a site with a few thousand daily impressions. Two per week is a reasonable ceiling. Above that, the bands are too narrow.
Should the monitor run daily or weekly? Weekly for the report, daily only for the Act band. Daily reports create daily reading, which is how monitoring becomes a job instead of a safeguard.
Can Codex or Claude Code run the same design? Yes. The bands, groups, annotations, and output shape are agent-neutral. Only the file conventions differ, which is why we keep the walkthroughs separate for Codex and Claude Code.
What do I do when a query alerts in the Act band? Check the page first, then the SERP. If the page is fine and the SERP changed shape, the fix is usually different content or a different format, not a rewrite. Our notes on how AI SEO platforms track ranking performance cover the review sequence.
Author: Camille Rhodes, Architect of 300+ AI Content Workflows at Auspia. Camille writes about workflow design, automation boundaries, and the review steps that keep AI-assisted SEO work useful.
