Tracking Individuals

Spyglasses can run an AI visibility report on a person, not just a brand. Use it to see how ChatGPT, Claude, Gemini, and Perplexity describe an executive, an industry expert, an author, or any public figure — and whether they surface when people ask about leaders in that person's field.

What You'll Learn

• When to use a person-level report (and when to use a brand report)
• How to add a person, and what the "role or Wikidata reference" field controls
• How the neutral identity snapshot is built — and what the source label means
• How to read the identity accuracy and tracked attributes sections
• How to use key messages to track positive and negative narratives
• What's different about a person report compared to a brand report

When to Use a Person Report

A person report is the right tool when the subject is an individual — for example:

• An executive whose AI reputation you manage on behalf of a client (the ORM use case)
• An industry expert who wants to be discoverable when AI is asked about their field
• A founder or thought leader building authority in a specific space

If the subject is a company, product, or category, run a brand report instead.

Adding a Person

In Properties → Add Property, choose Person as the subject type, then fill in:

• Name — the person's name, e.g. Mellody Hobson. This is required and is the search anchor for everything that follows.
• Role, Wikidata QID, or affiliation (optional) — a single field that drives how the identity snapshot is built:
  • A Wikidata QID (Q42) or a Wikidata URL → the snapshot is built from Wikidata + Wikipedia (the neutral primary source).
  • A role or affiliation in plain text, e.g. Co-CEO, Ariel Investments → the snapshot is built from a Google web search. The role is added to the search query as a disambiguator.
  • A LinkedIn URL → treated as "not a QID." The snapshot still comes from a web search (LinkedIn itself is never scraped), and the URL is not used as a role.
  • Empty → the snapshot is built from a web search by name alone.

You can change this field later from the property's settings and rebuild the snapshot.

The Identity Snapshot

Every person report starts from a neutral baseline — a profile of who the person actually is — that the AI models' answers are measured against. Each snapshot is labeled with where it came from:

When the source is Web sources, the report shows a one-line caveat: "Built from public web sources, not a neutral encyclopedia — treat identity accuracy as directional." This is honest framing — the same kind of pages the AI models read are now the baseline, so accuracy comparisons are directional rather than authoritative.

Editing the snapshot

Click Edit snapshot on the property's overview to override any field by hand. Manual edits persist across rebuilds (they "win" over both Wikidata and web data). You can also rebuild the snapshot at any time from the property's settings — for a Wikidata-sourced person it re-fetches Wikidata/Wikipedia; for a web-sourced person it re-runs the search.

Reading a Person Report

A person report has four sections, in order:

1. Identity snapshot

The neutral facts the AI is being measured against — name, role, employer, field of work, nationality, education, notable work, and (where available) a Wikipedia intro. Source links sit at the bottom of the card.

2. Identity accuracy

This is the person equivalent of brand consistency. We ask every model "Who is [name]?" — both with and without web browsing — and compare the answers to the neutral snapshot.

The section is tabbed per model. Inside each tab, a Without browsing / With browsing toggle picks the answer to show. The card surfaces:

• The full AI response (click "Show full response" to expand), rendered as markdown.
• Differences vs. neutral source — role/employer flags, a list of factual errors, and two specific warnings:
  • Same-name confusion — the model is describing a different person who shares the name.
  • Outdated information — the model is giving facts that were once true but are no longer current (a common failure for executives who recently moved roles).

Two scores sit at the top: average accuracy across all model/mode combinations, and cross-model consistency — how closely the models agree with each other. A high accuracy with low consistency means some models have a sharply different picture than others.

3. AI Visibility Metrics

This is the same AiVisibilitySection used by brand reports, but seeded differently. For a person, we auto-generate five field-level discovery prompts from the snapshot — questions like "Who are the most influential figures in [field]?" or "Name leading voices in [specialty]." Then we run them across every AI platform and measure whether the person surfaces.

The prompts are seeded from two inputs:

• The snapshot's structured facts (field of work, role, employer, nationality).
• The person's positive key messages, if any — these tell the generator what the person wants to be known for, so the prompts target the specific niche.

If your person has a clear specialty, capture it as a positive key message before running the first report — the discovery prompts will then test that exact framing.

The panel reads Discovery Queries · Mentions · Share of Voice · Identity Accuracy for a person (no Citations tile — citations are brand-domain-specific and don't apply).

4. Tracked attributes

Each tracked attribute (see Key Messages, below) appears as a collapsible row, grouped by sentiment. The header shows how often the narrative surfaced (e.g. "Found in 17 of 30 answers"); expanding the row reveals the matched snippets with the model that produced each one.

The denominator is every AI answer in the report — the four identity questions (with and without browsing) plus the five field-level discovery prompts across every platform — so the count is a live total of successful answers, not a fixed number.

Key Messages for Persons

Person reports use the standard Key Messages feature, with one extension: each message has a sentiment — Positive, Negative, or Neutral.

• Positive narratives — what the person wants to be known for. Example: "Leading voice in financial literacy." These also seed the field-level discovery prompts.
• Negative narratives — what you want to watch for. Example: "Embezzlement scandal." Useful for early-warning monitoring of reputational risk.
• Neutral associations — factual links you want to track without a positive/negative framing.

A narrative "appears" in an AI answer when a sentence in the response expresses the same idea semantically — not just a literal substring match. The matcher uses embeddings, so phrasings like "a champion of financial literacy" match "leading voice in financial literacy" even though the wording differs.

You can manage tracked attributes from the Key Messages page in the property sidebar.

Historical Trends

The Historical Metrics dashboard works the same way it does for brands, with two adjustments for persons:

• The Brand Consistency card and chart are relabeled Identity Accuracy and trend the avg-accuracy score over time.
• The Citations card and the Citations Over Time chart are hidden (no brand domain to cite).
• The Tracked Attribute Mentions Over Time chart appears whenever you have key messages defined — positive narratives chart in green, negative in red — so you can see, for example, a negative narrative rising over a quarter.

Person reports also feed the same nightly prompt runs as brand reports — every overnight prompt execution updates the tracked-attribute trend lines, so you don't have to run a fresh report to see movement.

What's Different from a Brand Report

For reference, here's what's hidden or relabeled in the PERSON flow:

• Sidebar: Placements, Publisher Lookup, and the entire Content section (FAQ Generator) are hidden — they're brand-PR tools that don't apply to a person.
• Report: the Robots.txt, Technical Analysis, Grounding Searches, and AI Readiness Audit sections are hidden — they all assume a crawlable website.
• Historical dashboard: Citations are hidden; Performance vs. Competitors is relabeled Performance vs. Peers; "Brand Mentions" becomes simply "Mentions."

Everything else — peer comparison, share of voice, share-of-voice trends, the public report link, the AI report scheduling — works the same way it does for a brand.

Practical Tips

• For non-notable people, lean on the role/affiliation field. "Director of Procurement, Acme Industries" yields much better disambiguation than the name alone.
• For same-name cases, watch the identity-accuracy tabs for the "Same-name confusion" warning — it tells you exactly which models are answering about the wrong person.
• Add positive key messages before the first report run — they're what shape the discovery prompts, so capturing the right niche up front means the report tests what you actually care about.
• Reading the "Web sources" caveat seriously — when the snapshot has no Wikidata anchor, the accuracy score is a directional signal (the AI models and your baseline are reading similar pages). Use trend movement and the differences list, not the absolute score, to drive decisions.