The Second Opinion

Your quote is in a buyer’s inbox. It is Tuesday night, the kids are down, and before they reply they pick up their phone and ask an assistant a version of the oldest question in business: is this mob any good?

They used to ask a neighbour. Now they ask a machine, and it answers in full paragraphs. It has an opinion on your pricing and knows, or thinks it knows, what you specialise in. It volunteers two other names in the same suburb and gives reasons for each. That conversation happens about you, without you, before you hear a word. Then you get a call or you get silence, and you never find out which sentence caused which.

This is not fringe behaviour any more. In BrightLocal’s 2026 consumer survey of 1,002 US adults, 45% said they had used an AI tool to get a local business recommendation in the past twelve months, against 6% the year before. Hold the size of that jump loosely; it is self-reported. Hold the direction firmly. And it is not only households. G2 surveyed 1,076 B2B software buyers in March 2026 and found 51% now open their research in an AI chatbot rather than a search engine, with 69% ending up on a different vendor than the one they walked in planning to buy. A different market to yours, but not a consumer novelty either.

None of this is Australian data. I looked and could not find an Australian read worth quoting, which tells you something on its own: nobody has measured this here yet. Take the direction, not the decimal places, then go and get your own number. That is what the next twenty minutes is for.

You cannot edit the model. You can edit what it reads. Which makes the first move obvious, and almost nobody makes it: go and hear the answer yourself.

The interrogation

Do this cold. Cold needs defining, because opening a new chat clears nothing. Log out entirely and use a private browsing window. If the assistant has a memory or personalisation setting, switch it off first. If that sounds like more fiddling than you will actually do, borrow your partner’s phone: same test, none of the setup. What you are avoiding is an assistant serving you a version of yourself it has spent two years learning to be nice about.

Ask five questions and ask them the way a buyer would, dropping in your own trade, suburb and business name. Who should I call for this job around here. Is that business, named outright, any good. How does it compare to the competitor everyone in your market already knows. What does it charge for your most common job. What do people complain about when they deal with it.

Run all five through at least two assistants. Screenshot every answer and put the date on it. Do not argue with the screen, do not follow up with a correction, do not tell it who you are. You are collecting evidence, not having a conversation. The urge to type “actually, we do offer that” is strong. Resist it. The buyer never types that sentence either.

Ten screenshots, twenty minutes, no cost. That is the whole instrument.

Four ways it goes wrong

Now read what came back, and give each answer exactly one word.

Absent is when you are not in it at all. The assistant answered happily, named three businesses, and none of them was you. Expect Absent on the open questions — who should I call, how does it compare — far more often than on the ones that name you. Ask a machine directly about a named business and it will usually find something to say; whether that something is accurate is the next question. Absent on a named query is rarer, and worth taking seriously when it happens. (On open queries the vendor numbers are bleak: SOCi, which sells AI-visibility monitoring, reported ChatGPT recommending 1.2% of the 350,000-plus locations it analysed.) Absent is also the mode operators most often misread as a technical fault. It is not. It means there is not enough evidence on the open web that you do this work, in this place, for the machine to have anything to assemble.

Wrong is when you are in it and the facts are false. A price you have never charged. An address you left in 2022. A service you do not offer, or worse, one you dropped because it was a nightmare. Wrong is the mode people assume is rare, and what testing exists suggests it is not. Searchable put 13,365 questions about 165 London businesses to ChatGPT, Gemini and Perplexity and found 93% had at least one basic fact wrong or missing; a larger run across UK high-street retailers returned at least one false fact for 64% of them, wrong postcode the most common error. Searchable sells AI-visibility monitoring and has not published its method, so treat those numbers as a prompt rather than a finding. Which is the argument for measuring your own.

Stale is when everything is accurate and three years old. The old positioning, the old service list, the review themes from before you fixed the thing everyone used to complain about. Stale is the quietest of the four because nothing in the answer is a lie. It just describes a business you have already stopped being.

Outshone is when you appear, the facts are right, and the competitor’s paragraph is better than yours. Theirs names a specific reason to pick them. Yours says “also well regarded in the area.” This is the mode that stings, because you are visible and losing anyway. It is also the most fixable: not a data problem but a positioning problem that has finally become machine-readable.

One word per answer. Do not write “a bit of all four.” Pick the worst.

Where the answer came from

Most of the time the machine is not inventing anything. It is reading. Every sentence it gave you was assembled from something you can go and find.

Google is explicit about this for its own AI features. Its guidance says there are no additional requirements or special optimisations to appear in AI Overviews or AI Mode, and no new machine-readable file to create. What it asks for instead: keep your Business Profile current, allow crawling, keep your content available as text. It describes the mechanism as a fan-out: a spread of related searches across subtopics and sources, composed into one answer. OpenAI publishes the same plumbing from the other side: several separate crawlers, each with its own robots.txt control. One of them, OAI-SearchBot, is what surfaces your site in ChatGPT search. Opt that one out and OpenAI says your site will not be shown in ChatGPT search answers. Worth checking whether whoever set up your robots.txt knew the difference between the crawler that trains a model and the crawler that recommends you.

Which means the source chain is boringly concrete. Your verified profile. Your reviews and, more importantly, the words inside them, because the machine reads sentences, not star ratings. The text on your own site. The directories holding a version of your details. Mentions elsewhere. Structured data where you have it. Nothing exotic, nothing hidden.

So for each failing answer, go and find the sentence’s parents. Search the exact phrase the assistant used. Usually it lands you on a page you forgot existed, a directory listing you never claimed, or a run of reviews all making the same observation.

Fix the source, not the answer

Each failure mode has a different repair. They are not interchangeable.

Absent is an evidence problem, which makes it a build rather than a repair. Start with the smallest possible version, before you go and read anything longer. Take your single best job from the last three months, the one you would want ten more of, and write four hundred words about it on your own site: what the problem was, what suburb, what you did, what it cost, how long it took. That is the first piece of evidence a machine can read, and it is one evening. The rest of the build is longer than this guide. The Recommendation makes the case and names what an agent is actually reading; Agent-Ready is the install order, scoped to one service so it finishes inside a fortnight. The only thing worth adding here: re-asking will not change the answer, because there is nothing new for it to read. You cannot be recommended for work you have published no proof of doing.

Wrong is a consistency problem. One name, one number, one address, one service list, everywhere it appears, including the listings you did not create. Google’s documented route for place data is Suggest an edit in Search, or editing the profile directly from the account that owns it, and it tells you an edit usually takes up to ten minutes to review but can take thirty days. Fix the worst source first, then hunt the copies, because a corrected profile sitting alongside four stale directory entries has not settled the argument. It has only entered it. Hunting every copy is a sit-down job with a list, not an afternoon’s tidy-up, and it is install five in Agent-Ready. What this guide hands you is the reason to bother: a false sentence you have now read with your own eyes.

Stale is a recency problem. Reviews arriving this month rather than a wall of them from 2023. Proof with dates on it. A site that says what you do now instead of what you did when you built it. Recency is the one signal you cannot write your way to. You earn it by having done something recently.

Outshone is the one that is not fixed by data at all. The repair is not more words. It is one provable reason a buyer picks you, set down plainly on a page a machine can quote, backed by something a sceptic could go and check. Most operators already have that reason. They have simply never written it down, because it has always lived in the conversation, and the conversation is now the part being skipped.

The habit

Once is a diagnostic. Monthly is a metric.

Put the same five prompts in your calendar, same wording, first week of the month, and log what comes back. Twenty minutes. Do not improve the prompts: reword them and you have lost the comparison, which was the entire point.

What you are tracking is not the sentences. It is the mode. Absent in March and outshone in June is progress you would have had no other way of seeing. Outshone in March and absent in June is drift, and it reaches your enquiry count well after it reaches the answer, by which time you will be blaming the season.

Logging the mode matters because the number you used to watch has stopped reporting. Pew tracked 68,879 real Google searches by 900 US adults. March 2025 data, worth naming as such, and still the strongest independent behavioural read published. On results pages carrying an AI summary, users clicked a traditional result on 8% of visits, against 15% where there was no summary, and clicked a link inside the summary itself on 1%. Watching your traffic to know how you are doing is watching a door most people no longer walk through.

The BrightLocal numbers hold one comfort. Among the consumers who had used AI for local recommendations, 63% said they trusted what it told them, and 88% said they fact-checked it anyway. The answer is not the verdict. It is the shortlist, and the fact-check lands on your profile, your reviews, and your site. Which is exactly the material this exercise makes you go and repair.

Where this stops being enough

Everything above is manual, free, and yours to run this week. For most operators it finds the first broken thing.

There is a second interrogation worth running, and this is not it. This one says your name out loud and reads what comes back about you. Agent-Ready’s machine test never mentions your name at all: it asks an assistant to find, compare and book somebody who does what you do, and reads where the assistant stalls. One tells you what your reputation looks like from outside. The other tells you whether you are reachable in the first place. Twenty minutes for this one. Agent-Ready’s is three short runs across a few days, because a single run of a non-deterministic assistant is an anecdote. Neither is a substitute for the other.

What it will not do is rank the damage. Twenty minutes of screenshots tells you three of five answers came back stale. It does not tell you whether stale is costing you more than the competitor’s better paragraph, or which of the two deserves the month. The Visibility read scores the same territory out of 75 and names the first thing to fix. Run the interrogation yourself now. Get it scored when the modes start competing for your attention.

Either way, the conversation is already happening. The only question is whether you have read it.

How to run the second opinion on yourself

Questions