Disbelief is a reasonable first reaction to an AI visibility audit. A business with 4.8 stars, 200 reviews, twelve years of trading and a full order book asks the obvious question: if we are this good, why does the assistant not know we exist?
Because rating is not what it is reading. Here are six reasons this happens. They are not ranked by frequency, because nobody has measured that; each one is worth checking on its own.
1. You are not on the pages it reads
An assistant answering "best plumber in {city}" fetches a handful of pages and writes a paragraph from them. Directories, a local "best of" post, a review aggregator and forum threads are all fair game, and the assistant will tell you which ones it used if you ask it for sources. How to show up in ChatGPT recommendations goes through that sequence step by step, including the crawler setting that can keep a website out of ChatGPT's search answers.
If your business is not on any of those pages, your rating is invisible in the most literal sense. It exists on a profile the model did not open.
This is the least interesting cause on the list and the most tractable, because the set of pages is finite: ask the question, ask for sources, and you have your list.
2. Your category is wrong
Google Business Profile categories are how a lot of downstream data gets filed. A business whose primary category is "Contractor" or "Business service" instead of "Plumber" propagates as a contractor everywhere that data is syndicated.
The effect is quiet. You still show up when someone searches your name. You still get map traffic for nearby generic searches. But when a model is composing a paragraph about plumbers, you are not in the set it drew from.
Check the primary category, not just the additional ones. It carries most of the weight. The Google Business Profile signals post covers the category and the other fields in the order to fix them.
3. Your reviews say nothing quotable
This is the counterintuitive one. Two hundred reviews saying "great service, highly recommend" give a model nothing to write with.
Assistants write recommendation sentences with a reason attached: "known for emergency callouts", "frequently mentioned for boiler repairs", "reviewers highlight fixed pricing". Those clauses come from review text and page text that says something specific.
A business with 40 reviews that repeatedly mention "same day" and "emergency" is easier to name than one with 300 generically positive reviews, because the first one comes with a clause and the second one does not.
You can influence this without gaming anything. When you ask for a review, ask a specific question: "if you have a minute, it helps most if you mention what the job was and how quickly we got there". People answer the question they are asked.
4. Your name is inconsistent
Models match names loosely. They do not match magically.
If you are "Smith & Sons Plumbing Ltd" on Companies House, "Smith and Sons Plumbing" on Google, "Smith Plumbing" on two directories and "S&S Plumbing" on a listicle, you have split your evidence across four entities. Each one individually looks like a minor local business. Combined they would look like the obvious answer.
Worse, a fuzzy matcher, ours included, has to make a judgement call about whether those are the same company. Our matching uses a token overlap score with a confident threshold and a review band below it, and the methodology page sets out the exact numbers. A business with four name variants is the kind of case that lands in that band, which means it gets counted conservatively until someone confirms the match.
Pick one form. Use it everywhere. Update the odd ones out.
5. You are competing for the wrong question
"Best plumber in {city}" is the least commercially interesting question in your category. The person asking it is comparing, not buying.
The questions that convert are narrower: "emergency plumber", "who replaces water heaters", "commercial plumbing contractor", "who fixes a burst pipe on a Sunday". Those answers draw on smaller candidate sets.
The reason to expect them to be more winnable is competition: fewer businesses have written the plain page that answers a narrow question. That is worth testing rather than believing. Ask the broad question and a narrow one, and compare how many businesses each answer names.
6. Your website says nothing a model can compress
Open your homepage and read it as if you had to summarise the business in one clause.
If the page says you are professional, reliable, trusted, and committed to quality, there is nothing to compress. Every competitor says the same words, which means the words carry no information.
If it says you are gas certified, family run since 2009, cover four named towns, answer calls 24 hours, and quote fixed prices, you have handed a model five possible clauses. It will use one.
This is not a copywriting flourish, it is an information problem. Specific facts survive compression. Adjectives do not.
What a 4.8 rating actually buys you
It buys you the conversion, once you are in the answer. It is not what gets you into the answer.
That distinction is worth being precise about, because the reflex is to assume the system is broken or unfair. It is neither. The assistant is picking a few names from the sources it read. Your rating never entered the calculation because your business never entered the source set.
Which is good news, mostly. A rating takes years to build. By our own estimates, a directory listing takes ten minutes, a category fix takes two, and a plain page answering a narrow question takes an afternoon.
The order to fix them in
- Find the citation set. Ask your top questions, ask for sources, list the URLs. Half an hour.
- Fix the Google Business Profile. Primary category, service list, hours, description with the city named in text. One hour.
- Join the listings in bucket one. The directories and association listings that appeared in your citation set and take a form. An afternoon.
- Standardise the name. Everywhere. One afternoon, mostly waiting on support tickets.
- Write one narrow page. The service you most want to sell, answered plainly, city named, price range stated. One afternoon.
- Change how you ask for reviews. Ask for the specific, not the general. Ongoing, costs nothing.
- Re-audit in a month. Same questions, same conditions.
The times are our own estimates, not measured averages. Items 1 to 5 are roughly two working days spread over two weeks. Nothing on the list is clever, and nothing on it is a guarantee: each item changes what an assistant can find and quote about you, and item 7 is how you find out whether that changed the answer.
Checking whether it moved
Do not trust a single answer, in either direction. These models are not deterministic: the same question can produce different names on consecutive asks. That is why a real measurement asks the same question several ways and reports a rate, and why "we appeared once yesterday" is not evidence of anything.
Ask eight phrasings, logged out, fresh conversation each time, and count. Named in three of eight is a real improvement over zero of eight. Named once, once, is noise. The 15-minute manual audit sets out the questions, the recording sheet and the scoring.
If you would rather not run that by hand every month, that is the job AIgnored does, with one difference: a scan asks its own fixed set of 8 broad recommendation questions per model, listed in every scan's why drawer, not the service, price and personal questions in the manual audit, so those stay a manual check. It matches the names back to real businesses with the borderline cases shown rather than guessed, and gives a score per business per scan. It runs on your own API keys, so a scan costs what the provider charges. The pricing page has a free plan that covers one city.
The one line to take away
Your rating tells people to hire you once they have heard of you. It does not tell an assistant that you exist. Those are two different problems, and only one of them takes years to solve.