How ChatGPT decides who to recommend for “electrician in Phoenix, Arizona”

0Verdict

1What the customer actually saw

The delivered answer, verbatim. Citation markers removed because they are transport plumbing; not one other character is changed.

2Why these businesses and not this one

The businesses the answer recommended, set against this one, on the two structured fields the model held while it decided.

Every business it recommended carried a rating of at least 4.8 and at least 106 reviews.

A rating of 4.6 sits below every business it recommended.

493 reviews match or beat 4 of the 6 businesses it recommended.

Review volume alone did not decide this answer: the business with the most reviews in the whole set, Mister Sparky® of Phoenix with 2710, was not recommended.

Read off one recording. These are what this retrieval did, not thresholds ChatGPT publishes.

3Where the decision was actually made

Two surfaces can produce a recommendation: the local business records, and the pages it opened on the web. They are bought in completely different ways, so which one carried this answer decides where the work goes.

ChatGPT opened 20 results across 11 domains before it answered.

angi.com, apmarizonaelectric.com, bestnearme.pro, bestprosintown.com, consumeraffairs.com, dispatchcore.io, reviews.birdeye.com, thumbtack.com, electriciansinphoenix.com, expertise.com, homeadvisor.com

None of them was the website of any business it returned. This answer was decided on directory, review and aggregator pages. Work on a company website does not touch those pages; getting listed and rated on them does.

The site for this business, turnitonelectric.com, was not among the pages it read.

Nearly everything here came through the local business record. That record is the product: the categories on it, the rating, the review count and whether it is claimed. Web placement does not reach this route.

4What this business has to change

Each item carries where it was measured: this recording, the reference sweep, or both. Guidance is written once per signal rather than generated per report, so the numbers move and the meaning does not.

5Fan-out searches, verbatim

Reproduced byte-for-byte in the order emitted. This is the primary evidence: everything above is an interpretation of these lines.

search("electrician in Phoenix Arizona")
business|Phoenix, AZ, USA|electricians;licensed electricians;emergency electricians
fast|electrician Phoenix Arizona licensed electricians reviews

5bDecoded grammar — every line, every field

The payload is pipe-delimited: call_type|query|freshness_days|domain. An em dash means the slot was present in the grammar and left empty, which is itself a choice the model made.

Call-type mix: search ×1, business ×1, fast ×1.

Domain-scoped searches: 0 of 3. These are the searches that read one site in particular — either the business's own domain or a directory.

6Reasoning summaries, verbatim

Model-authored summaries, not raw reasoning tokens. They describe method in general terms and rarely name a source.

7Brand × answer matrix

Every business the retrieval surfaced, whether it survived into the delivered answer, and the structured scores the model held on it while deciding.

7bThe places channel — every field on every record

A second retrieval surface with no web URL. Businesses here are visible to the model whether or not they rank in search.

7cField census

8Pages it actually read

Every result it opened, with the page title verbatim.

8bFormat of the pages it read

9The brief the model wrote itself

Words that appear in its own searches but not in the customer's question. This is what it went looking for and could not get from the prompt.

10Cross-channel contradictions

Figures stated in the answer checked against the structured record for the same business in the same turn.

No cross-channel contradiction found in this scan: every figure stated in the answer matched the structured record it came from, or could not be checked against one.

11Citation channels

Which retrieval channel each citation marker points back to.

12The reference sweep behind the guidance

The scan above is this business, recorded once. The figures here come from the reference sweep — 27 captures and 82 fan-out searches across 12 categories on gpt-5-6. A single capture cannot carry a percentage; the sweep can. The two are never averaged together.

AIntegrity gate and capture record

Every recording taken for this business, whether it passed, and why not.