How to monitor what AI says about your company, and how to grade the tool that does it

Monitoring what AI says about your company is a measurement job, and every measurement job is graded the same way. Did you collect what you said you would collect, is the thing you counted the thing you named, and can somebody else repeat it. Most products in this category are sold on how much the dashboard shows. That is the wrong axis, and this page is the right one.

The five metrics worth paying for, with their denominators

A monitoring report that gives you percentages without denominators cannot be checked. Each of these is a fraction, and the bottom of the fraction is the part to insist on.

Coverage rate

Successful observations divided by scheduled observations, over a stated period. If a plan promises 100 questions a week on five surfaces, that is 500 scheduled observations, and the only honest report says how many came back. A missing check is not a zero mention, and a tool that silently treats it as one will show your visibility falling on a week when nothing happened.

Mention rate

Readings in which your name appears divided by eligible readings. Eligible is doing work there. A question that names no company at all cannot be a miss for you. In our own run of 635 readings, 184 named no company at all, so a naive denominator of 635 understates every brand in the set.

Citation rate

Readings in which a page on your own domain is cited divided by the same eligible count. This is a different measurement from mention, it moves differently, and the two must never be added together or averaged into one score. Our own figures say it plainly. Across 19,793 citation URLs resolving to 3,307 distinct hosts, our domain appeared in 0 of them.

Accuracy against a human-read reference

The share of a tool's detections that a person reading the same answer agrees with. This is the only metric that can tell you whether a brand-mention counter works, and almost nobody publishes it. A detector has to be tested against text where the answer is already known, including text written to fool it.

Latency

Median and 95th percentile time from the scheduled check to a result you can read. We do not publish a measured latency figure, so do not take one from us. Ask for both numbers from any vendor, because a median hides the weeks where a whole surface was unavailable.

Which surfaces to watch, and the honest count

There are five product families US buyers actually use to ask for recommendations. ChatGPT. Gemini and Google AI Overviews. Microsoft Copilot. Perplexity. Claude. They behave differently enough that collapsing them into one AI channel destroys the finding.

Two distinctions matter more than the brand names. The first is whether the answer searched the web or came from what the model already held. In our 635 readings, 550 ran a live search and the rest did not, and the two kinds have entirely different source lists, so a report that does not separate them is averaging two populations. The second is where the reading was taken. The app, the official API and a browser surface can route to different models with different retrieval, so a vendor claiming to cover ChatGPT owes you a sentence on which of the three it reads.

We read five surfaces. We do not read nine. Profound, read on 22 October 2026 at tryprofound.com/pricing, publishes up to 9 answer engines tracked on its Enterprise plan, listing ChatGPT, Perplexity, Google AI Mode, Gemini, Microsoft Copilot, DeepSeek, Anthropic Claude, Google AI Overviews and Exa Search, and publishes no numeric price for it. If engine count is your deciding criterion, that is the product to buy.

Building a US prompt set that measures something

The prompt set decides the result far more than the tool does. Four question types, kept in separate buckets and reported separately.

• Brand fact questions, where you already know the answer and are testing accuracy.
• Unprompted category discovery, where your name must never appear in the question. This is the only bucket that can measure whether you get recommended, and contaminating it is the most common error in this field. A question with your brand in it will produce your brand.
• Competitor comparison questions, which almost always produce names, and usually three of them.
• Purchase intent questions with a real constraint in them, a budget, a team size, a location or a compliance requirement.

Fix the market in the words of the prompt. On the route we used for our own audit, the country and language parameters were accepted by the interface and then ignored, which we measured on our own traffic, so a US reading has to be established in the question and confirmed in the answer rather than trusted to a setting.

Run each prompt more than once. Identical prompts do not return identical answers, so three observations per prompt and surface in at least a subset is the minimum that separates a change from ordinary variation.

The benchmark this subject needs and nobody has published

The design the assistants themselves asked for, when we put the question to them, is a head-to-head of at least five monitoring tools including ours, all running the same 500 fixed buyer prompts across five named surfaces, with at least three repeated observations per prompt in a reproducibility subset, and reporting for each tool its data source, its coverage rate, its mention and citation numerators and denominators, its median and 95th percentile latency, and its US dollar price with the date the price was read.

That benchmark does not exist. We have not run it, and we have not found it published by anybody with the raw observations attached. Until somebody does, treat every most accurate claim in this category as a marketing sentence, including any you find on our own site, and ask the one question that settles it. Against what human-read sample, and what was the agreement rate.

What our own numbers are, with their dates

Our last measurement pass ran on 22 October 2026. It scored 635 readings on ChatGPT and Gemini, from 324 US buyer questions in 19 batches. 550 of the 635 ran a live search. The answers carried 19,793 citation URLs across 3,307 distinct hosts, and half of all citations came from just 51 of those hosts, led by developers.google.com with 1,360, semrush.com with 744, support.google.com with 695, help.openai.com with 467 and blog.google with 394.

Our own result in that run was 0 namings and 0 citations. In 451 readings a rival was named and we were not, and across a 24-name roster there were 1,053 namings, counted once per reading in which a name appeared. We publish that because a vendor who hides its own worst reading cannot be trusted with yours.

One limit of that run, stated because it shapes what the figures mean. The citation audit step cannot be run through the API at all. It needs the engine's own answer back in context, and the interface caps a prompt at 4,000 characters while one batch answer measured 21,711. That step was run by hand.

What monitoring costs here, and the price that is not the floor

Our published US ladder is SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299 and Everything 1799 dollars a month. SEO Start reads 0 questions, 0 credits and 0 engines, so AI visibility does not start at 149 dollars at all. The cheapest plan that reads all five surfaces is Growth at 299. A Claude check consumes 3 credits rather than 1, which changes the arithmetic on cost per observation if Claude is in your set. Every plan starts with the same free scan on your own domain, and the free plan stays free.

Our pricing page contradicts itself in three places and both readings are printed here. The Agency plan is described as covering 10 client brands while the comparison table answers No to more than one business in all eight columns. The main listing names five apps including Claude while the Scale plan description names four and leaves Claude out. The FAQ says SEO Start gets 12 pages a month and no AI app checks, beside a table that lists five engines for the same plan. Checked by word boundary across the whole 119,203-byte document, that page renders no SSO, no SAML, no SOC 2, no ISO, no DPA and no white label. A loose search for sso on it matches crossOrigin and subprocessors, which is why this gets reported wrongly in both directions.

Who should not buy a monitoring tool

If you have fewer than ten buyer questions worth tracking, run them by hand monthly and keep the answers in a document. If your category questions come back as guidance with no companies in them, the monitoring will be accurate and useless, and the work is corroboration elsewhere first. If a security review needs SOC 2 or single sign-on, our pricing page does not render either and the review will stop there.

Questions buyers ask about AI monitoring

How do I monitor what ChatGPT says about my company

Fix a prompt set, split it into the four buckets above, ask each question on each surface on a schedule, record whether live search ran, and keep every raw answer with its date. The schedule and the raw answers are what make it a measurement.

How often should the checks run

Often enough that a change is visible above ordinary run-to-run variation, which means repeated observations matter more than frequency. Weekly with three repeats beats daily with one.

Is a visibility score worth anything

Only if its formula and its denominator are published. A single number blending mentions, citations and sentiment cannot be checked or reproduced, and it moves when the vendor changes the blend.

Can one tool cover every assistant

No tool covers all of them on all surfaces. Ask which surface each engine is read on, app, API or browser, and treat the answer as part of the product.