01Tracking
AI brand monitoring: one reading is noise, a trend is information
Tracking from $900 a monthA frozen prompt set sampled on cadence. Three cycles before anything is called a trend.
Quick answer
Tracking samples the same frozen prompt set on a fixed cadence and reports how citation presence, share of voice, factual accuracy and source usage move over time, per platform.
The reason it exists: these systems are non-deterministic. One 2026 analysis found roughly 30% of brands appearing in one answer stayed visible in the next for the same query. A single check cannot tell you whether you have a problem.
Key takeaways
- Monthly suits most brands. Weekly variance usually exceeds the change you are trying to detect.
- The prompt set is frozen, and changes only with a version bump and a dated note.
- Three cycles before anything is called a trend. Two points make a line, not a direction.
- Alerts fire on sustained movement, never a single reading. Alerting on noise teaches people to ignore alerts.
- Accuracy is tracked, not just presence. A drop there is more urgent than a drop in citations.
- It will not prove causation, and we label claims Estimated rather than Verified when it cannot.
Why a single check tells you almost nothing
Ask an assistant the same question twice and you can get two different answers, with different sources. That is not a bug in the system or an error in your measurement. It is how these systems work.
One 2026 analysis put numbers on it: roughly 30% of brands appearing in one answer remained visible in the next for the same query. Which means a single check has something close to a coin-flip relationship with your actual position.
The practical consequence is uncomfortable for anyone selling a one-off report. If you check once and appear, you cannot conclude you are visible. If you check once and do not, you cannot conclude you have a problem. Either reading is as likely to mislead you as inform you.
What produces a reliable signal is repetition: the same prompts, sampled multiple times, on a fixed cadence, with the distribution reported rather than a single result. That is the whole idea, and it is why tracking is a subscription rather than a purchase.
What gets reported
Five measures, each per platform, never blended into a single score.
| Measure | What it is | What a change means |
|---|---|---|
| Citation presence | Share of runs where you are cited with a link | The headline number. Slow to move, and the one clients watch |
| Named without citation | Runs where you are mentioned but not linked | Tracked separately because it is invisible to analytics entirely |
| Factual accuracy | Whether what is said about you is correct | A drop here is more urgent than a drop in citations |
| Share of voice | You against the competitor set agreed before sampling | Rising while your citations fall means the category grew, not you |
| Source usage | Which domains are cited alongside or instead of you | The most actionable line. It is your off-page target list |
The competitor set is agreed before the first cycle and frozen with the prompt set. Choosing it afterwards is how measurement quietly becomes marketing.
Choosing a cadence
Monthly for most brands. Not because it is a comfortable billing period, but because weekly variance in a non-deterministic system frequently exceeds the change you are trying to detect. Weekly sampling costs four times as much and mostly measures noise.
Two exceptions. During an active fix cycle we sample Perplexity more often, because it is retrieval-led and reflects corrections within weeks rather than months. And after a migration or a major content change, a short burst of denser sampling catches damage early, when it is still cheap to reverse.
Quarterly works for slow categories with long buying cycles. It is a genuine option and we will say so rather than selling the higher cadence by default.
Alerting that does not cry wolf
An alert fires on sustained movement across readings, never on a single fluctuation.
This matters more than it sounds. In a system where 30% of brands hold position between consecutive answers, single-reading alerts fire constantly and almost all of them are noise. What happens next is predictable: within two months everyone mutes the channel, and the alert that mattered arrives into a room where nobody is listening.
A quiet alerting system that fires four times a year and is right each time is worth more than one that fires weekly. That is a deliberate trade of sensitivity for trust.
Reading a trend against a moving background
Your numbers move for reasons that have nothing to do with you, and a report that ignores that is worse than no report.
- Platform retrieval changes without notice. A citation rate that shifts across every client in the same week is a platform event, not a client one.
- Competitors publish and earn placements. Your share of voice can fall while your citations hold.
- Query demand is seasonal. Trigger rates move with it.
- The answer format changes. More or fewer sources cited per answer shifts everyone’s numbers at once.
We annotate these on your timeline. When a number improves and we think a platform change caused it, the report says so rather than accepting the credit, which is the only way the good months stay believable.
What this means for you
- Do not act on a single reading. In this system it is close to a coin flip.
- Freeze the competitor set before cycle one. Choosing it later is how a report becomes a pitch.
- Watch accuracy, not just presence. Being described wrongly costs more than being absent.
02Honest limits
What tracking will not do
It will not prove your work caused a change. Platforms change retrieval, competitors act, demand shifts. We can correlate and annotate. We cannot isolate, and anyone claiming attribution certainty in this environment is overstating what the method supports.
It will not see unattributed mentions. A meaningful share of assistant answers name no source at all, and no tool including ours can measure what is not attributed.
It will not cover every conversation your buyers have. It covers the prompts in your set. A wider set costs more and still is not exhaustive, and we would rather say that than imply total coverage.
It will not give you one score. The two most similar platforms shared only 24.1% of cited pages in 2026 research. A blended number would average away the finding it exists to surface, which is precisely why we do not produce one.
Tracking or audit?
Different jobs, and buying them in the wrong order wastes money.
The $99 audit establishes where you stand today across seven workstreams. It is a baseline, and it tells you what to fix. Tracking measures movement against that baseline over time.
Buying tracking without an audit means watching a number with nothing to compare it to and no diagnosis of why it sits where it does. Almost every engagement starts with the audit for that reason.
Tracking is included in every retainer, with cadence and platform coverage scaling by tier. Programmes start at $900 a month.
03Common questions
What buyers ask about ongoing measurement
How often should we sample?
Monthly suits most brands. Weekly variance in a non-deterministic system frequently exceeds the change you are trying to detect, so weekly sampling mostly measures noise at four times the cost.
Why does the prompt set have to be frozen?
Because a set that changes between cycles makes the trend meaningless. If you add three prompts in March and citation rate moves, you cannot tell whether visibility changed or the measurement did.
What if our business changes and the prompts no longer fit?
Then the set changes, with a version bump and a dated note explaining what was added or removed. The trend restarts at that point, and the report says so rather than pretending continuity.
How many readings before a trend is real?
Three cycles. Two points make a line, not a direction. We will show you the second reading and we will not call it a trend.
Can tracking prove our work caused the change?
No, and nothing can. Platforms change retrieval without notice, competitors publish, seasonality moves demand. We annotate what we know changed and label causal claims Estimated rather than Verified.
What triggers an alert?
Sustained movement across readings, not a single fluctuation. Alerting on one bad reading in a non-deterministic system trains people to ignore alerts, which is worse than having none.
Do you track sentiment?
We track factual accuracy, which is more actionable. Whether an answer describes you correctly is checkable against a fact set. Whether it sounds positive is a judgement that varies between coders.
Should we buy tracking or an audit first?
The audit, almost always. Tracking measures change against a baseline. Without the baseline you are watching a number with nothing to compare it to.
Sources
Every figure on this page comes from one of the following. Where a claim is ours rather than a source’s, it carries an evidence label instead.
- Rankability — State of AI Search, 1,645 citation observations across 31 topics, 2026. Source for the 24.1% cross-platform overlap. Independent study
- xSeek — 2026 analysis finding that roughly 30% of brands appearing in one AI answer remained visible in the next for the same query. Independent study
- Google Search Central — Optimizing your website for generative AI features (15 May 2026, updated 10 July 2026), and Google I/O announcements, 19 May 2026. Primary
- Searlixa sampling across client engagements, labelled Observed with sample and date. Our own
Ask an assistant
Hand this page to an assistant and ask it to summarise or challenge the argument. If it gets something wrong about us, that is itself a useful finding.
→What you are buying
AI visibility tracking services: what is included
Tracking is a cadence, not a dashboard. This is what arrives, and how often.
- A frozen prompt set, version-locked so the series stays comparable
- Monthly sampling, three runs per prompt, per platform
- Named or absent, and whether what is said about you is accurate
- Competitor presence recorded beside yours on every prompt
- Platform changes annotated, so a shift in their system is not read as a shift in yours
- Raw sampling records and coding rules handed over, yours to keep
Who this is for
Anyone who has done the work and needs to know whether it landed, and anyone who needs to report AI visibility to someone who will ask how it was measured.
Who it is not for
Anyone needing a reading this week. Three cycles pass before anything here is called a trend, because one reading on a non-deterministic system is noise.
From $900 a month. Monthly sampling at every tier. What changes is the prompt count, the platforms and the number of markets. SEO and AI visibility pricing is published in full.
The audit sets the baseline the series is measured against. Without it, cycle one has nothing to compare to. The $99 SEO and AI visibility audit comes back in three working days, and the fee comes off your first invoice if you continue.
→Next step
Get the baseline before you start measuring movement.
The measurement method · Pricing · How each platform behaves