How We Measured Home-Services Visibility in AI Search

Measure mentions, citations and recommendations separately

AI-search visibility has at least three observable states, and they should not be treated as synonyms.

Caldrin's panel stored mentions and owned-domain citations as separate fields. Recommendation ordering was not reliable enough for headline use, so it stays NOT_MEASURED in this page.

Eligible observations are different from failed and quarantined attempts

An eligible observation was a successful, grounded output under the panel's inherited rules. A failed attempt never became a negative brand observation. A quarantined attempt stayed in the operational record but did not enter the eligible denominator.

All 40 prompt-and-surface cells reached the target of 30 eligible observations. Individual cells required 30 to 35 attempts under a maximum of 40.

Accounting note: Caldrin internal panel, collected from 2026-08-31T14:33:27.406Z to 2026-08-31T17:51:13.325Z, across 20 frozen prompts, ChatGPT consumer ground truth and rendered Google AI Overviews, United States at country grain. The table separates 1,200 eligible observations from 57 failed and 6 quarantined attempts. Failures and quarantines are retained run states, not evidence that a brand was absent.

Caldrin did not appear inside this panel

Caldrin was not named and caldrin.co was not cited in any of the 1,200 eligible observations. No prompt in the frozen 20-prompt cohort produced an eligible Caldrin appearance.

Bounded result: 0/1,200 eligible observations and 0/20 frozen prompts with an eligible Caldrin appearance. The 1,200 eligible observations came from 1,263 attempts, with 57 failed and 6 quarantined, across ChatGPT consumer ground truth and rendered Google AI Overviews at United States country grain from 2026-08-31T14:33:27.406Z to 2026-08-31T17:51:13.325Z. This result does not prove that Caldrin is absent from other prompts, surfaces, dates, countries or local markets.

Caldrin's citation rate is NOT_MEASURED, not 0%. The rate is conditional on an eligible Caldrin appearance, and that denominator was empty.

This is the uncomfortable value of a baseline: it records the result that occurred, without polishing it into a better story.

Repetition measures variation inside a window

One answer per prompt cannot show whether the answer or source set changes on the next attempt. This panel repeated each prompt-and-surface cell until it reached 30 eligible observations, then stored per-cell answer and source-overlap measures.

Those repetitions describe variation inside one short collection window. They do not establish stability or persistence across days, weeks, later reruns, changed prompts, changed locations or provider changes. A later baseline should preserve the cohort and measurement rules if the goal is a like-for-like comparison.

Variance boundary: The variance artifact covers all 40 completed prompt-and-surface cells in Caldrin's frozen 20-prompt US-country panel across ChatGPT consumer ground truth and rendered Google AI Overviews. Collection ran from 2026-08-31T14:33:27.406Z to 2026-08-31T17:51:13.325Z, retaining 1,200 eligible observations from 1,263 attempts, with 57 failed and 6 quarantined. The measures describe same-window variation only. Machine extraction does not establish recommendation order, sentiment or market share.

Organic and AI-source overlap is a co-occurrence check

Ten domains appeared in both the panel's 783-domain AI-source set and a supplied 16-domain organic comparison set. The union contained 789 domains.

The ten co-occurring domains were birdeye.com, firstpagesage.com, getcourtyard.ai, hookagency.com, intleacht.ai, plumberseo.net, sequoiageo.com, shiftflow.app, silverbackstrategies.com and webfx.com.

Overlap boundary: The AI-source set comes from Caldrin's frozen 20-prompt US-country panel across ChatGPT consumer ground truth and rendered Google AI Overviews, collected from 2026-08-31T14:33:27.406Z to 2026-08-31T17:51:13.325Z, with 1,200 eligible observations from 1,263 attempts, 57 failed and 6 quarantined. The supplied organic set is a separate dataset. 10/789 is domain-level co-occurrence, not same-URL overlap, a link intersection, causation or proof that organic ranking produced an AI citation.

Intervene, then remeasure the same population

An intervention log should name the changed page or asset, the change date, the intended measurement, the owner and the rollback state. The next run should reuse the frozen prompts, surfaces, geography, eligibility rule, repetition target and attempt cap.

A post-change difference is an observation, not automatic proof that the intervention caused it. Provider behaviour, answer variance, changed source availability and time can all move between runs. If the population changes, version it and stop calling the comparison like for like.

The rerun records what moved. It still does not tell you why. Causal credit needs a design that can rule out competing explanations.

What this panel cannot tell a home-services team

The panel does not support claims about:

• universal AI-search visibility;
• city, service-area, near-me or local-pack performance;
• Gemini, model APIs or every Google AI experience;
• which recurring source is most trusted, accurate or authoritative;
• recommendation order, sentiment or competitor market share;
• leads, pipeline, revenue, conversion or return on investment;
• whether reviews, links, schema, content or domain consolidation caused a citation; or
• persistence across days, weeks, changed prompts, changed locations or provider changes.

The page format is also an editorial direction. No counted live-SERP structural study was supplied, so dominant page type, competitor word-count range, entity consensus and NLP salience remain NOT_MEASURED.