The question an assistant asked Google on a buyer's behalf
Between June and September 2026, more than 600 questions of twelve words or longer showed searchintel.tech in Google results without a single click. We pulled them from our own Search Console. Read in full, most of them are not what a person types into a search box. They are prompts: a persona block, a task, an instruction to search the web and shortlist providers. One of them, in full:
"We've been tracking our AI visibility with a prompt monitoring tool for a few months, but honestly nobody looks at it anymore and no actions come out of it. We're looking for an agency that reviews the setup and turns the data into an actual strategy. Search online for specialized providers and suggest specific providers."
Somebody typed that into an assistant. The assistant ran it through Google, and our site was in the results it read. Nearby sit "is an AI search monitoring platform worth the investment", "what's the difference between entry-level and enterprise AI visibility tools?", and "our competitors keep coming up in AI answers and we dont, what tools can help us figure out why and fix it". Search Console does not label which of these came from a person and which from an agent, and an impression count says nothing about how many people sit behind it, so treat the set as a shape rather than a market size. The shape says one thing plainly: people have bought monitoring, are asking what it was for, and are asking a machine to find them the answer.
Why the dashboard stalls
Five causes come up again and again. Each has a fix you can apply to the tool you already pay for.
1. The score has no decision attached
Most dashboards report a visibility percentage and stop. Gartner's September 2026 guidance on answer engine journeys recommends adding metrics such as Share of Answer and Citation Presence, and says in the same breath that "these metrics do not replace traditional marketing measures". The risk Gartner names is "misdiagnosing performance declines, misallocating investment and losing visibility into brand influence". A score on its own does none of that work.
Ask one question of every chart: if this number moved five points next month, what would we do differently? If the answer is nothing, retire the chart. It was reassuring you, not measuring anything.
2. The prompts are not your buyer's questions
Tools ship with generic prompt sets, and generic prompts produce generic answers. The questions that decide revenue are the ones a buyer asks before they know your name: "which UK tour operators are best for solo travellers", "who should I use for a SIPP transfer", "best platform for enterprise drone fleet management". In our experience twenty of those, written for your category and held steady for months, beat two hundred generic ones. Our guide to checking competitors in AI answers shows how to build the set.
3. A cut-off answer looks like an absence
This one is invisible from the dashboard. Engines return answers with a length cap, and when the cap hits, the answer stops. Of 138 ChatGPT answers we collected on our UK travel question panel between 10 August and 21 September 2026, 46 came back cut off before the end. That is one answer in three. A brand that would have been named in the second half of a cut answer is scored absent, and the dashboard reports a decline that never happened.
Ask your vendor one question: does the tool read the stop reason on every answer, and does it exclude cut answers from the count? If they do not know what you mean, that is the answer.
4. Brand matching is a text search, and the engines use shorthand
A monitoring tool finds your brand by matching text. Engines shorten names. One UK operator on our panel was counted in 9 of 21 answers when we matched its full name, and 16 of 21 once we read the shorthand the engines actually use for it. The dashboard had halved the brand's presence, and nobody inside the company would have known.
Read the raw answers for your own brand once a quarter. Note every spelling and abbreviation the engines use. Give the list to the tool.
5. It reports presence, and presence is not the lever
Whether an engine recommends you is decided upstream, in the sources it reads. Third-party reviews, trade press, forums and comparison sites carry most of the weight, as our analysis of which pages AI pulls from sets out. A dashboard that shows your presence going down without showing which sources changed hands leaves you with a symptom and no cause.
Google's own Search Console report for generative AI, rolled out to every property in August 2026, has the same limit. It shows impressions, pages, countries and devices, and it does not show the prompt, the competitors named beside you, or whether you were chosen. Knowing that you appeared is not knowing why.
The loop never closes for a second reason. Profound's US panel study of two million AI conversations, published in August 2026 and limited to American users, found that more than 97 percent of site visits following an AI mention carried no AI referral tag. Analytics will not show you the payoff, so the dashboard never gets credit and never gets blamed.
What a monthly read should contain
The read that gets opened has four parts.
One number that means selection. Named is not recommended. On one week of our travel panel, 119 completed answers named a tracked brand 445 times; 432 of those were recommendations and 13 were passing mentions, most of them cautions or price comparisons. For most brands the two numbers are the same, so a tool that sells you a "recommendation rate" should be able to show you the thirteen.
Movement on a rolling four-week window, never week to week. Engines vary from run to run, and a weekly line chart is mostly noise.
The sources that changed. Which third-party pages gained or lost the citations behind your questions, and who is now named in your place.
Three actions, each with an owner and a date. If a month's read produces no action, the read is too long or the questions are wrong.
What to do this week with the data you already have
- Export the answers, not the scores. Every serious tool stores the raw text. Read every answer where you are absent and write down who is named instead and which source the engine cites for them.
- Check the stop reasons. Ask the vendor for the completion field, or read the answers and count the ones that end mid-list.
- Pick one question you should win and are not winning. Trace the sources behind it. Fix or earn one of them.
- Retire any chart that has not changed a decision in a quarter.
How to check where you stand
If you want an outside read before touching the tool, our free AI visibility check runs your brand across ChatGPT, Claude, Gemini, Perplexity and Google AI Mode and shows who is named instead of you. For a full AI visibility monitoring programme built on buyer questions, sources and a monthly decision read, book a call.
Frequently asked questions
Is an AI search monitoring platform worth the investment?
Only if the read changes a decision each month. A platform that reports a visibility score with no prompt set built for your buyers, no completion check on the answers, and no view of the sources behind the answers tends to be ignored within a quarter. Judge a platform on whether it shows you who is named instead of you and why, not on how many engines it covers.
What is the difference between entry-level and enterprise AI visibility tools?
Entry-level tools track a fixed prompt set and report presence. Enterprise tools add custom prompts, more markets and API access. Neither level guarantees the two things that matter: completed answers counted correctly, and a source layer that explains movement. Ask both questions before looking at the price.
Why does nobody look at our AI visibility dashboard?
Because it reports a number without a decision. The common causes are generic prompts, cut-off answers scored as absences, brand names matched by exact text while engines use shorthand, and no view of the sources. Any one of these makes the number unreliable, and people stop looking at numbers they cannot act on.
How often should AI visibility be reviewed?
Monthly, on a rolling four-week window. Weekly readings swing with engine variation and produce reactions to noise. A monthly read with three owned actions is the shape that survives.
Can Google Analytics show the effect of AI mentions?
Mostly no. Profound's August 2026 US panel study found that over 97 percent of visits following an AI mention arrived with no AI referral tag, so they appear as direct or organic traffic. Treat the analytics view as a floor, and measure the mention itself.