The observations, counted once.
12 answers
One answer for each combination of three cities and four trades. Each was allowed to list up to five businesses. All twelve completed; none was repeated.
54 entries
Business entries across the answers. Some businesses appeared more than once, so this is not a count of 54 different firms.
55 cited URL occurrences
Each URL was counted once within its answer. One business entry cited two pages. This is not a count of 55 different websites.
A citation did not always have a separate page return.
For 36 of the 55 cited URL occurrences, the visible tool record contained a successful page return with metadata. For the other 19, it contained search-result text without a successful separate page return. Every final URL could be matched to a returned search or page record.
That distinction describes the record we could inspect. It does not establish everything the provider fetched or read internally. The record also contained 22 page-retrieval error results, including three for pages that were not cited. Those errors do not establish that 22 websites were broken.
Two details worth checking before choosing.
One answer flagged conflicting branch-location information on a removals website. Another distinguished an air-conditioning supplier advertising commercial work from suppliers offering household services.
These were observations in the assistant’s answers, not our independent audits of the businesses. They show why the cited service and location details deserve a closer look before you use a shortlist. A citation alone does not verify quality or current availability.
What ran, and when.
We observed the round on 28 October 2026, from 19:14:51 to 19:17:01 UTC. It used native Codex CLI 0.162.1 with the configured model gpt-6-astra through an existing ChatGPT subscription. This was a Codex observation, not a test of the consumer ChatGPT product.
The twelve questions crossed Johannesburg, Cape Town and Durban with household removals; air-conditioning installation, servicing and repairs; electrical installation and repair; and plumbing installation and repair. Each began: “Which businesses offer [service] in [city], South Africa?” A shared instruction requested public sources, up to five unranked entries, URLs and uncertainty. It also asked the answer to distinguish advertised services from verified performance; all twelve did so.
We froze the method and exact prompts before starting. Each question ran once in a fresh session without our project instructions, skills, apps, plugins or memory. Live web search was enabled and search activity appeared in every session. OpenAI’s configuration guide, linked below, documents the distinction between live and cached search.
The prompts excluded four preselected business domains and their subdomains. That restriction could change which businesses appeared. The names and exact prompts are retained privately, so this public description alone cannot fully reproduce the run. We contacted no business.
What this round cannot tell us.
One answer per question cannot establish how often a business will be recommended, compare cities’ visibility, or explain why a provider was chosen. We did not measure customer outcomes or test a website change. These observations do not establish results for other assistants.
The visible record did not expose every retrieved page body, search location, personalisation or the provider’s backend model revision. Naming a South African city in a question does not prove the search ran from that city.
A useful check for your own business.
Save the exact question, answer and date. Read the cited pages and check the service, customer type and area they describe. If your own public information conflicts, correct the specific mismatch you can demonstrate.
Keep a later answer as another observation. A changed answer would not, by itself, prove that your edit caused it or brought new work.
A few useful answers.
Does an answer with fewer businesses mean fewer local options exist?
No. The prompt allowed up to five entries and each question ran once. A shorter answer does not establish the number of suitable businesses in that city.
Does a retrieval error mean a website is broken?
No. It records a failed page return in that session. It does not show how the site works for other visitors, crawlers or later requests.
Was this a consumer ChatGPT or multi-assistant study?
No. It was one native Codex condition: CLI 0.162.1, configured model gpt-6-astra and live web search, observed on 28 October 2026. Using a ChatGPT subscription to access Codex does not make it a consumer ChatGPT test.