Module 07 · Lesson 2 of 4

Tracking AI Mentions and Citations

About 12 minutesPrerequisite: Lessons 3.2 and 7.1

After this lesson you canupgrade the Module 3 log into a durable tracking system, add context and sentiment labels without over-engineering, and decide honestly whether a paid tool earns its place.

You already own the core system: a fixed query set, a repeatable run, a log. What turns a log into a tracking system is structure that survives months, labels that make trends answerable, and a decision about automation. This lesson is deliberately unglamorous; it is also the difference between practitioners and tourists.

The upgraded schema

Keep every field from lesson 3.2 (run, date, query, surface, mentioned, cited domains, competitors, classic position, notes) and add three:

Mention type: cited link, named with no link, or quoted text. The three mean different things and convert differently.
Context label: recommended, listed-alongside-competitors, contrasted-negatively, neutral. One label per mention, chosen from a fixed list so the column stays countable.
Answer changed: yes or no versus last run. This single column is your variability early-warning: a wave of "yes" across queries usually means a surface or corpus shifted, not that you did something.

DefinitionA tracking system is a log plus fixed vocabularies plus a cadence. Fixed labels are what let you ask "how many negative-context mentions this quarter?" six months from now and get an answer.

Cadence and retention

Weekly runs, monthly roll-ups, quarterly reviews. The roll-up is one hour: count mentions by type, count context labels, list queries where the answer changed three weeks running, and write three sentences of interpretation. Keep the raw rows forever; storage is free and the historical record is the asset. When someone asks "did we have visibility for X in March?", the system answers instead of someone's recollection.

Where tools fit, honestly

Approach	Strengths	Costs and limits
Manual runs in a spreadsheet	Total control of conditions; you see the answers yourself; free	Time-bounded: fine to about 40 queries, a few surfaces
Browser automation you script	Cheaper repetition of the same conditions	Surfaces change; scripts break; bot detection complicates runs
Paid AI-visibility tools	Breadth, scheduled runs, dashboards	Costs; their sampling differs from yours, so numbers will not match your log

If you adopt a tool, run it alongside your manual log for a month and compare. Where they disagree, trust your controlled conditions and note the divergence. The tool can carry breadth; the method carries truth. And whatever carries the runs, the interpretation stays yours: no dashboard knows that the competitor surge in week six coincided with their data story launch. You do, because you logged it in lesson 6.2.

Worked example: the monthly roll-up (illustrative)

Mentions by type: 9 cited, 14 named, 3 quoted. Citations up 3 from last month.
Context: 11 recommended, 6 alongside-competitors, 2 negative (both on pricing queries).
Answer changed: 6 of 30 queries, three weeks running. All six on the pricing cluster.
Interpretation: citation growth is real and broad; the pricing cluster is unstable and slightly negative. Next month's priority: the pricing comparison slot from the opportunity queue.

Four counts and three sentences. That is a report any client or board can use.

Workbench 7.2

Add the three new fields to your log and fix their vocabularies in a sheet note.
Run this week's run with the full schema, including the answer-changed column.
Book the monthly one-hour roll-up, and write last month's three interpretation sentences if you have the history.
If you use a paid tool, start the month-long side-by-side comparison and note the first divergences.

Self-check

Why fixed vocabularies for context labels?

So the column stays countable over months. Free-text notes cannot be aggregated; a fixed list turns sentiment and context into numbers you can trend and report.

Your tool reports share numbers different from your log. Which is wrong?

Neither, necessarily: they sample differently. Your log reflects your controlled conditions; the tool reflects its own. Trust your conditions for decisions, keep the tool for breadth, and note divergences instead of averaging them away.

What does a wave of "answer changed = yes" usually mean?

A surface or corpus shift on the platform side, not something you did. It is an early-warning column precisely so you do not misattribute platform churn to your own work.

Key principleA log becomes a system when its labels are fixed and its cadence is booked: structure is what makes months of runs answerable.

Sources used in this lesson
Google Search Central: AI features and your website

Previous 7.1 Metrics That Still Matter Next lesson 7.3 Source Share and Competitor Benchmarks