A customer types "best commercial electrician in Geelong" into ChatGPT or Gemini, gets three names back, and yours isn't one of them. Your competitor's is. Nobody sent you a report about it. You found out because a mate mentioned it, or because enquiries went quiet without your rankings moving.
That is the actual problem, and it is not the same problem as ranking on a search results page. Frase, writing in August 2026, defines visibility in generated answers as how often a brand appears in the answer itself across ChatGPT, Perplexity, Google AI Overviews and similar — and notes it is measured differently from traditional ranking position. Superlines reported in March 2026 that around 93% of these sessions end without a click through to a website; treat that figure as indicative rather than precise, but the direction matters. Being named in the answer is the outcome. Sitting at position four underneath it may not be.
The surface has split, so stop optimising for one
Most advice on this still assumes ChatGPT is the whole game. It isn't any more. Similarweb's traffic panel, reported around 29 July 2026 with data to May 2026, has ChatGPT's share of generative-AI website visits falling from roughly 76% in June 2025 to about 53% by May 2026, with Gemini and Claude holding meaningful and growing shares. Similarweb's own reading of this is that a strategy built around a single assistant now reaches a shrinking fraction of the audience.
Practical consequence: if you check your visibility, check it on at least three assistants, not one. Any work you do that only pays off inside one product is worth less than it was a year ago.
Run the ten-prompt check yourself this week
Before paying for a tracking dashboard, do this manually. It takes about an hour and it tells you most of what a paid tool would.
- Write down the ten questions a real customer would actually ask. Not keywords — sentences. "Who should I use for emergency plumbing in Wollongong on a Sunday?" "Best bookkeeper for a small cafe in Fremantle." "I need a structural engineer in Ballarat for a home extension, who's good?" Include price questions, urgency questions and "who's near me" questions.
- Run all ten through ChatGPT, Gemini and one other assistant your customers plausibly use. Use a logged-out or fresh session where you can, so you aren't seeing answers shaped by your own history.
- Log the result in a spreadsheet: prompt, assistant, date, businesses named in order, and any sources the assistant cites.
- Repeat monthly, same prompts, same day of the month.
What you're looking for is not your own rank. It's the pattern in who does get named, and where the assistant appears to be pulling from. If the same three directories, review sites or association pages keep surfacing across assistants, those are your targets. That list is worth more than any generic checklist, because it is specific to your trade and your town.
One caution: assistants give varying answers to the same prompt, and none of the operators publish how local recommendations are assembled. Don't over-read a single result. Look for what repeats across ten prompts and three assistants over three months.
Fix the record before you write anything
In practice, the sources assistants cite in local answers tend to be surfaces the business doesn't own — directory entries, review text, maps data, industry association listings, supplier "find an installer" pages, local news. That is an observation you can make yourself by logging the citations in the check above; no operator publishes how these answers are assembled, so treat it as a working assumption rather than a rule. Your own ten-prompt log is better evidence for your trade than anything generic, including this article.
Blog output is the thing vendors sell you because it's the thing they can bill for. It is rarely the constraint. The constraint is usually that the record is a mess. Work through it in this order:
- Name, address, phone — identical everywhere. Pick one exact form of your business name, one address format, one phone number. Then go and make every listing match it. Old suite numbers, a trading name on one site and the Pty Ltd name on another, a mobile on one and a landline on another — each inconsistency is a reason for a system to treat you as two half-businesses instead of one solid one.
- Claim and complete the obvious listings. Your maps profile, the major national directories, and whatever the dominant directory is in your trade. Complete beats clever: categories, service area, hours, services list, photos.
- Kill the duplicates and the dead entries. Old office addresses, a listing from a previous business name, a franchise page that still lists you. These actively confuse the record.
- Get onto the third-party surfaces that already rank in your category. Industry association member directories, licensing body registers, supplier and manufacturer partner pages, chamber of commerce listings, awards finalist pages. These are often thin, well-structured, trusted pages that say plainly "this business, this trade, this town".
Reviews: volume and recency, not just stars
Review text is where the descriptive detail lives — the suburb, the specific job, the problem solved. A 4.9 average from eleven reviews, the most recent of them two years old, does less work than a 4.6 from a hundred with a dozen from the last quarter.
So build a boring, repeatable ask into the job. After completion, one message, one link, no incentive. Reply to every review, including the bad ones, in plain language that names what the job actually was. That reply text becomes part of the corpus too.
Don't buy reviews. Beyond the platform risk, fake reviews read as generic, and generic text gives an assistant nothing specific to repeat.
What your own site still needs to do
Your site is not the main input, but it has one job: be the unambiguous source of truth that everything else can be checked against.
- A service area stated in words, with the suburbs and towns named.
- Services listed explicitly, in the words customers use, not internal jargon.
- Contact details matching the canonical version exactly.
- Structured data marking up business type, location and hours. Review markup is a separate question — search platforms place restrictions on self-serving review snippets on a business's own site, and getting it wrong can see your markup ignored or worse. Check the current guidance for the platform before you add it, and get your developer to do the same.
- A short page per major service-plus-location combination, where those combinations are real. Two honest pages beat forty spun ones.
Whether any of this directly influences a given assistant is not something the operators disclose. It's basic hygiene that pays off across surfaces regardless.
What to do next
Block out an hour this week for the ten-prompt check and save the spreadsheet. Spend a second hour auditing name, address and phone across every listing you can find, and write down the inconsistencies. Those two hours will tell whether the problem is that the record is fragmented, that the reviews have gone stale, or that competitors are simply present on trade surfaces where the business isn't.
Then fix one category at a time and re-run the same ten prompts in a month. Movement here is slow and lumpy — expect quarters, not weeks. But an owner with a clean record, current reviews and a presence on the trusted third-party surfaces in their trade is giving these systems something to name. An owner with three conflicting phone numbers isn't.