What do buyers complain about with AI visibility tools? The eight complaints, and the measured reason an assistant answers this question badly

Ask ChatGPT or Gemini what users complain about with AI visibility tools and a tidy list arrives. The list is built from a source pool we have measured twice. Across 1,260 readings on ChatGPT and Gemini on 16 October 2026 we recorded 20,143 citations across 4,947 distinct domains: 82.0 percent were a vendor or agency own site and 0.7 percent were review directories. An earlier run of 597 readings on a different question set gave 85.6 percent and 0.8 percent. Two independent question sets eleven days apart, the same answer. So there is almost no review corpus for an engine to draw a complaint from, and the complaints it reports are assembled mostly from pages written by sellers.

That is a structural fact about the answer rather than a criticism of any company in this market. It also means the useful version of this question has to be answered from measurement design, which is what the eight below are. We sell a tracker in this category and our own position is stated on each one, including the three where we come off worst.

Why the source pool matters more than the list

In the first of those two runs, g2.com appeared in 49 readings and was the tenth most cited domain overall, which is the whole of the review directory presence in a dataset of 10,652 citations. The directories are not being refused by the crawlers either: when we read 26 robots.txt files from the domains our audits named, the directories permitted every assistant agent at the root, which is set out in the directory pages that allow every crawler and still took 0.7 percent. Permission was never the constraint. The engines simply reach for vendor pages, documentation and news.

The practical consequence for a buyer: do not treat an assistant's complaint list as field research. Treat it as a reading list, then test each item against the vendor in front of you.

The eight complaints that are really about measurement design

1. Two tools report different numbers for the same brand

They asked different questions, in a different country, a different number of times, through a different route into the engine. A consumer app, an official API with live search and a licensed search feed return different answers on the same day. Pick one instrument, freeze the prompt set, and measure your own movement against itself rather than against somebody else's dashboard.

2. The number arrives with no interval, so nobody can tell a move from noise

Across 102 conversations in which we asked the engines what a page needs to be worth citing, 83 asked for a confidence interval, a variance or a repeat count, and 80 asked for the sample size. Our own rates print with a 95 percent range and overlapping ranges are reported as no change. One competitor publishes a related rule in its own method page, that a mention below 50 percent confidence never counts towards a score, and publishing that threshold is the right instinct.

3. The score was flattering because the prompt said the brand name

This is the complaint with a measurement behind it. On 27 July 2026 the same 78 questions were run twice in one day for one business. The run whose wrapper named the brand came back ranking it first on almost every question. The blind run put the company second by breadth and absent altogether from the questions it most wanted to win. The first run was discarded, because the only thing it measured was our own prompt. Ask to see the prompt list, and if your brand name is in it the rate is about the prompt.

4. The score moved and nothing in the business changed

Identical prompts produce different answers minutes apart, so a single run is a draw rather than a behaviour. Repeats and a stated cadence are what turn readings into a rate. One competitor publishes its cadence as a plain sentence, that each prompt is executed once every 24 hours on every selected model, and a published cadence is a fair thing to ask every vendor for.

5. The bill was larger than the plan price

Almost always the unit. Fifteen prompts across three engines repeated three times is 135 answer runs, not fifteen. A tool billing per run and a tool billing per prompt will quote numbers an order of magnitude apart for identical work and both are being honest. Do the multiplication before the shopping.

6. There was no way to show somebody how the number was produced

Raw answer export belongs on the required list rather than the nice to have list. The first question a sceptical finance director asks is how the figure was made, and a chart does not answer it. Ask for the answers with timestamps for one week of readings during the trial, not after.

7. Our content was used in an answer and we were never named

This state is real and it is measured. On 17 September 2026 Claude confirmed it had used two specific arguments drawn from a set of pages and had dropped the attribution in both places, describing it as a citation lapse rather than a ranking judgement. Reading a sample of answers by hand is what finds it, so build that half hour into the monthly routine and keep it with the tool rather than instead of it.

8. The tool says we are visible nowhere

Zero is a common and real starting point rather than a broken product. Our own domain was cited in 6 readings of 1,260 in October 2026, zero of those in the United States, and zero on Gemini across 921 readings in two audits. We publish those numbers because they are the baseline we will be judged against, and a vendor that shows you its own worst figure is easier to check than one that does not.

Cited and named are two different complaints

Two of the items above are quietly about different things and a buyer who merges them will buy the wrong instrument. Being cited means your domain appeared in an answer's source list. Being named means the business was named in the answer text. In those same 1,260 readings, rival tools in this category were named in answer text 1,314 times and our own domain was cited in 6 readings. Those are two separate counts of two separate events and they are never added, averaged or compared in one figure. A tool that reports only one of them is answering half the question, so ask which half you are being shown.

The complaint this category has earned, and where ours sits

The one grievance the measurement supports across every vendor, ourselves included, is that this market sells numbers and publishes very little a buyer can check. Of 102 build spec conversations, 101 asked for a before and after with the window stated and 68 asked for failures and negative results to be shown. Our own answer to the first is published at our proof page, which carries two Google Search Console exports both taken on 2026-10-14 with their windows stated, and a dated blind scan in which 6 of 10 answers named the subject on two apps. Those are published readings on stated dates rather than a statement about what a reader's own numbers will do. The second one, the failures, is the eight paragraphs above and the two zeroes in complaint eight.

Questions buyers ask about problems with AI visibility tools

Are the complaints an assistant lists about these tools reliable?

Treat them as a starting list. Review directories were 0.7 percent of 20,143 citations in our October 2026 run and vendor and agency own sites were 82.0 percent, so the answer is assembled mostly from pages written by sellers. The eight items above are the ones that hold up when you test them against a product rather than read them about one.

Which single question exposes most of these problems in a demo?

Show me the exact prompt list. If your brand name appears in it, the rate is about the prompt rather than about your market, and the 27 July 2026 double run is what that looks like when you measure it: first place on almost every question in the named run, second by breadth and absent from the key questions in the blind one.

Why do vendors report different visibility scores for my brand?

Different prompt sets, different countries, different repeat counts and different routes into each engine. Two tools can both be right and disagree completely, which is why a score from one vendor is never comparable with a score from another. Keep one method frozen and track your own direction of travel.

Is a low score a reason to change tools?

Usually it is a reason to read the prompts and the engine list. A rate near zero on blind buyer questions is the normal first reading for a business that has not done this work, ours included at 6 readings of 1,260. Change tools when the vendor will not hand you the answers behind the number.

What should be in a contract to prevent most of this?

Four lines: the prompt set version and the confirmation that your brand is in none of them, the engines written out individually with the cost of each extra one, the repeat count and cadence, and raw answer export included in your tier rather than above it. Those four turn most of the complaints on this page into things you can check on day one.

The reason this list is worth keeping

Every item here is a question you can ask in a twenty minute call, and seven of the eight are settled by a document rather than an opinion. The grievance lists written about this category rest on a source pool in which 0.7 percent of the citations came from the places people actually post their grievances, so the eight questions above are worth more than any list of them, including this one. Take them into the next demo and the vendor who has thought about measurement will be visible inside ten minutes.