Most AI visibility claims are a screenshot of one good answer. This page is the standard Filtrs measurement follows instead, published so anyone can check our numbers against our rules.
The Metric
Citation frequency, never a one-shot snapshot.
AI answers change by engine, question, model, location, and day. So the unit we report is frequency: how many times a source appeared across how many observations in a stated window. A favorable screenshot proves an answer existed once. Frequency tells you whether it repeats.
The Rubric
Three states. Strict definitions.
Cited
The domain appears in the AI answer's own source list. Nothing else counts as a citation.
Named
The person or business is named in the answer text without a link back. Reported separately, never blended into citation counts.
Not present
Neither of the above. Ranking in ordinary search results near an AI answer is organic placement, not a citation, and is never counted as one.
When we tightened these definitions in August 2026, we restated our own first client's July numbers under the stricter rule and published the restatement. The case study shows it.
The Panel
Fixed questions, frozen before collection.
Measurement runs on a fixed panel of neutral local questions: the things buyers and sellers actually ask AI about a market, spanning agent selection, valuation, neighborhoods, market conditions, and more.
No prompt names any agent, brokerage, or brand. The panel is a neutral census, not a leaderboard built to flatter anyone.
Once collection starts, a panel version is frozen and never edited. New questions form a new version, reported separately.
The same questions repeat on a fixed schedule, so month-over-month numbers compare like with like.
Collection
Repeated runs, engines never blended.
The standard schedules four fixed collection slots per week across different days and dayparts, because answers vary by both.
ChatGPT answers are nondeterministic, so each question is sampled multiple times per run: presence is a rate, never one roll of the dice.
Every engine is reported separately. Results are never pooled, because engines produce different observation counts and pooling would silently overweight one.
Microsoft Copilot data comes from Bing Webmaster Tools and is Microsoft-sampled; it is always labeled that way, with its verification date.
The engine, provider, model, and settings are recorded on every run. A material change starts a new series; incompatible series are never silently joined.
The Honesty Rules
Rules that cost us good-looking numbers.
Every published number carries its denominator and collection period: "cited in 9 of 12 observations, July 1 to 31," never a bare percentage.
Failed collections are excluded from denominators. A timeout is not an observation.
History is append-only. Log entries are never edited or deleted; corrections are published as restatements, on the record.
Results are reported as observed, not as caused. "Visibility increased following publication" is the strongest claim we make without attribution evidence.
Losses, regressions, and flat stretches are published with the same prominence as wins. That is what makes the wins believable.
See the standard applied to a real client.
Every rule on this page is in force on the case study: dated numbers, denominators, restatements, and the losses next to the wins.